From d998bb65fc143e2855e9e0f0d88f4c5712632e5d Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Mon, 18 Aug 2025 19:31:13 +0300
Subject: [PATCH 01/86] scripts : update sync scripts

---
 scripts/sync-ggml-am.sh | 50 +++--------------------------------------
 scripts/sync-ggml.sh    | 17 +-------------
 2 files changed, 4 insertions(+), 63 deletions(-)

diff --git a/scripts/sync-ggml-am.sh b/scripts/sync-ggml-am.sh
index 924f67ee5b5..1f87e23122b 100755
--- a/scripts/sync-ggml-am.sh
+++ b/scripts/sync-ggml-am.sh
@@ -61,21 +61,7 @@ while read c; do
         cmake/ggml-config.cmake.in \
         src/ggml-cpu/cmake/FindSIMD.cmake \
         src/ggml*.h \
-        src/ggml*.c \
-        src/ggml*.cpp \
-        src/gguf*.cpp \
-        src/ggml-blas/* \
-        src/ggml-cann/* \
-        src/ggml-cpu/* \
-        src/ggml-cuda/* \
-        src/ggml-hip/* \
-        src/ggml-kompute/* \
-        src/ggml-metal/* \
-        src/ggml-musa/* \
-        src/ggml-opencl/* \
-        src/ggml-rpc/* \
-        src/ggml-sycl/* \
-        src/ggml-vulkan/* \
+        src/ggml* \
         include/ggml*.h \
         include/gguf*.h \
         examples/common.h \
@@ -118,22 +104,7 @@ if [ -f $SRC_WHISPER/ggml-src.patch ]; then
     # cmake/ggml-config.cmake.in        -> ggml/cmake/ggml-config.cmake.in
     # src/ggml-cpu/cmake/FindSIMD.cmake -> ggml/src/ggml-cpu/cmake/FindSIMD.cmake
     #
-    # src/ggml*.c          -> ggml/src/ggml*.c
-    # src/ggml*.cpp        -> ggml/src/ggml*.cpp
-    # src/ggml*.h          -> ggml/src/ggml*.h
-    # src/gguf*.cpp        -> ggml/src/gguf*.cpp
-    # src/ggml-blas/*      -> ggml/src/ggml-blas/*
-    # src/ggml-cann/*      -> ggml/src/ggml-cann/*
-    # src/ggml-cpu/*       -> ggml/src/ggml-cpu/*
-    # src/ggml-cuda/*      -> ggml/src/ggml-cuda/*
-    # src/ggml-hip/*       -> ggml/src/ggml-hip/*
-    # src/ggml-kompute/*   -> ggml/src/ggml-kompute/*
-    # src/ggml-metal/*     -> ggml/src/ggml-metal/*
-    # src/ggml-musa/*      -> ggml/src/ggml-musa/*
-    # src/ggml-opencl/*     > ggml/src/ggml-opencl/*
-    # src/ggml-rpc/*       -> ggml/src/ggml-rpc/*
-    # src/ggml-sycl/*      -> ggml/src/ggml-sycl/*
-    # src/ggml-vulkan/*    -> ggml/src/ggml-vulkan/*
+    # src/ggml* -> ggml/src/ggml*.c
     #
     # include/ggml*.h -> ggml/include/ggml*.h
     # include/gguf*.h -> ggml/include/gguf*.h
@@ -154,22 +125,7 @@ if [ -f $SRC_WHISPER/ggml-src.patch ]; then
         -e 's/(^[[:space:]]| [ab]\/)cmake\/common.cmake/\1ggml\/cmake\/common.cmake/g' \
         -e 's/(^[[:space:]]| [ab]\/)cmake\/ggml-config.cmake.in/\1ggml\/cmake\/ggml-config.cmake.in/g' \
         -e 's/(^[[:space:]]| [ab]\/)src\/ggml-cpu\/cmake\/FindSIMD.cmake/\1ggml\/src\/ggml-cpu\/cmake\/FindSIMD.cmake/g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml(.*)\.c/\1ggml\/src\/ggml\2.c/g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml(.*)\.cpp/\1ggml\/src\/ggml\2.cpp/g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml(.*)\.h/\1ggml\/src\/ggml\2.h/g' \
-        -e 's/([[:space:]]| [ab]\/)src\/gguf(.*)\.cpp/\1ggml\/src\/gguf\2.cpp/g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml-blas\//\1ggml\/src\/ggml-blas\//g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml-cann\//\1ggml\/src\/ggml-cann\//g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml-cpu\//\1ggml\/src\/ggml-cpu\//g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml-cuda\//\1ggml\/src\/ggml-cuda\//g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml-hip\//\1ggml\/src\/ggml-hip\//g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml-kompute\//\1ggml\/src\/ggml-kompute\//g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml-metal\//\1ggml\/src\/ggml-metal\//g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml-musa\//\1ggml\/src\/ggml-musa\//g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml-opencl\//\1ggml\/src\/ggml-opencl\//g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml-rpc\//\1ggml\/src\/ggml-rpc\//g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml-sycl\//\1ggml\/src\/ggml-sycl\//g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml-vulkan\//\1ggml\/src\/ggml-vulkan\//g' \
+        -e 's/([[:space:]]| [ab]\/)src\/ggml(.*)/\1ggml\/src\/ggml\2/g' \
         -e 's/(^[[:space:]]| [ab]\/)include\/ggml(.*)\.h/\1ggml\/include\/ggml\2.h/g' \
         -e 's/(^[[:space:]]| [ab]\/)include\/gguf(.*)\.h/\1ggml\/include\/gguf\2.h/g' \
         -e 's/(^[[:space:]]| [ab]\/)examples\/common\.h/\1examples\/common.h/g' \
diff --git a/scripts/sync-ggml.sh b/scripts/sync-ggml.sh
index 00162daa05b..4296ddf5f50 100755
--- a/scripts/sync-ggml.sh
+++ b/scripts/sync-ggml.sh
@@ -6,22 +6,7 @@ cp -rpv ../ggml/src/CMakeLists.txt   ./ggml/src/CMakeLists.txt
 cp -rpv ../ggml/cmake/*              ./ggml/cmake/
 cp -rpv ../ggml/src/ggml-cpu/cmake/* ./ggml/src/ggml-cpu/cmake/
 
-cp -rpv ../ggml/src/ggml*.c        ./ggml/src/
-cp -rpv ../ggml/src/ggml*.cpp      ./ggml/src/
-cp -rpv ../ggml/src/ggml*.h        ./ggml/src/
-cp -rpv ../ggml/src/gguf*.cpp      ./ggml/src/
-cp -rpv ../ggml/src/ggml-blas/*    ./ggml/src/ggml-blas/
-cp -rpv ../ggml/src/ggml-cann/*    ./ggml/src/ggml-cann/
-cp -rpv ../ggml/src/ggml-cpu/*     ./ggml/src/ggml-cpu/
-cp -rpv ../ggml/src/ggml-cuda/*    ./ggml/src/ggml-cuda/
-cp -rpv ../ggml/src/ggml-hip/*     ./ggml/src/ggml-hip/
-cp -rpv ../ggml/src/ggml-kompute/* ./ggml/src/ggml-kompute/
-cp -rpv ../ggml/src/ggml-metal/*   ./ggml/src/ggml-metal/
-cp -rpv ../ggml/src/ggml-musa/*    ./ggml/src/ggml-musa/
-cp -rpv ../ggml/src/ggml-opencl/*  ./ggml/src/ggml-opencl/
-cp -rpv ../ggml/src/ggml-rpc/*     ./ggml/src/ggml-rpc/
-cp -rpv ../ggml/src/ggml-sycl/*    ./ggml/src/ggml-sycl/
-cp -rpv ../ggml/src/ggml-vulkan/*  ./ggml/src/ggml-vulkan/
+cp -rpv ../ggml/src/ggml* ./ggml/src/
 
 cp -rpv ../ggml/include/ggml*.h ./ggml/include/
 cp -rpv ../ggml/include/gguf*.h ./ggml/include/

From 74ed8d825d259bee9b3394cf512172206b85ca22 Mon Sep 17 00:00:00 2001
From: Kai Pastor <dg0yt@darc.de>
Date: Wed, 30 Jul 2025 14:52:26 +0200
Subject: [PATCH 02/86] vulkan : fix 32-bit builds (ggml/1313)

The pipeline member can be cast to VkPipeline.
This is a VkPipeline_T* on 64 bit but a uint64_t on 32 bit.
Cf. VK_DEFINE_NON_DISPATCHABLE_HANDLE documentation.
---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index a99b1c73130..b1a1cff904b 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -1341,7 +1341,7 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
         vk::DebugUtilsObjectNameInfoEXT duoni;
         duoni.objectType = vk::ObjectType::ePipeline;
         duoni.pObjectName = pipeline->name.c_str();
-        duoni.objectHandle = reinterpret_cast<uint64_t>(static_cast<VkPipeline_T*>(pipeline->pipeline));
+        duoni.objectHandle = /*reinterpret_cast*/(uint64_t)(static_cast<VkPipeline>(pipeline->pipeline));
         vk_instance.pfn_vkSetDebugUtilsObjectNameEXT(device->device, &static_cast<VkDebugUtilsObjectNameInfoEXT &>(duoni));
     }
 

From c6d170bec525debe22ea8cc9346d12e3743ddea5 Mon Sep 17 00:00:00 2001
From: Kai Pastor <dg0yt@darc.de>
Date: Wed, 30 Jul 2025 14:53:16 +0200
Subject: [PATCH 03/86] cmake : Fix BLAS link interface (ggml/1316)

---
 ggml/cmake/ggml-config.cmake.in | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/ggml/cmake/ggml-config.cmake.in b/ggml/cmake/ggml-config.cmake.in
index fe34cda4e01..2322c6cd9d0 100644
--- a/ggml/cmake/ggml-config.cmake.in
+++ b/ggml/cmake/ggml-config.cmake.in
@@ -34,8 +34,8 @@ if (NOT GGML_SHARED_LIB)
 
     if (GGML_BLAS)
         find_dependency(BLAS)
-        list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES ${BLAS_LIBRARIES})
-        list(APPEND GGML_CPU_INTERFACE_LINK_OPTIONS   ${BLAS_LINKER_FLAGS})
+        list(APPEND GGML_BLAS_INTERFACE_LINK_LIBRARIES ${BLAS_LIBRARIES})
+        list(APPEND GGML_BLAS_INTERFACE_LINK_OPTIONS   ${BLAS_LINKER_FLAGS})
     endif()
 
     if (GGML_CUDA)

From aa662f2431ad80d3f446eb80b699d96cb502b290 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Alberto=20Cabrera=20P=C3=A9rez?=
 <alberto.cabrera@codeplay.com>
Date: Mon, 28 Jul 2025 11:05:53 +0100
Subject: [PATCH 04/86] sycl: refactor quantization to q8_1 (llama/14815)

* sycl: quantization to q8_1 refactor

* Refactored src1 copy logic in op_mul_mat
---
 ggml/src/ggml-sycl/backend.hpp   |   1 +
 ggml/src/ggml-sycl/ggml-sycl.cpp | 256 ++++++-------------------------
 ggml/src/ggml-sycl/quantize.hpp  | 133 ++++++++++++++++
 3 files changed, 184 insertions(+), 206 deletions(-)
 create mode 100644 ggml/src/ggml-sycl/quantize.hpp

diff --git a/ggml/src/ggml-sycl/backend.hpp b/ggml/src/ggml-sycl/backend.hpp
index f839a42bc90..410a67b0195 100644
--- a/ggml/src/ggml-sycl/backend.hpp
+++ b/ggml/src/ggml-sycl/backend.hpp
@@ -28,6 +28,7 @@
 #include "mmvq.hpp"
 #include "norm.hpp"
 #include "outprod.hpp"
+#include "quantize.hpp"
 #include "quants.hpp"
 #include "rope.hpp"
 #include "set_rows.hpp"
diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
index a023d6fb452..b08941c328b 100644
--- a/ggml/src/ggml-sycl/ggml-sycl.cpp
+++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
@@ -44,6 +44,7 @@
 #include "ggml-sycl/set_rows.hpp"
 #include "ggml-sycl/sycl_hw.hpp"
 #include "ggml-sycl/getrows.hpp"
+#include "ggml-sycl/quantize.hpp"
 #include "ggml.h"
 
 static bool g_sycl_loaded = false;
@@ -1373,120 +1374,6 @@ typedef void (*ggml_sycl_op_mul_mat_t)(
 
 
 
-template<int QUANT_BLOCK_TILE>
-static void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int kx, const int kx_padded,
-                          const sycl::nd_item<3> &item_ct1) {
-    const int ix = (item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                    item_ct1.get_local_id(2)) * QUANT_BLOCK_TILE;
-
-    if (ix >= kx_padded) {
-        return;
-    }
-
-    const int iy = item_ct1.get_local_range(1) * item_ct1.get_group(1) +
-                   item_ct1.get_local_id(1);
-
-    const int i_padded = iy*kx_padded + ix;
-
-    block_q8_1 * y = (block_q8_1 *) vy;
-
-    const int ib = i_padded / QK8_1; // block index
-    const int iqs = i_padded % QK8_1; // quant index
-    typedef  sycl::vec<float, QUANT_BLOCK_TILE> TC;
-    typedef  sycl::vec<int8_t, QUANT_BLOCK_TILE> TQ;
-    TC zeros;
-    TQ qzeros;
-#pragma unroll
-    for (int i = 0; i < QUANT_BLOCK_TILE; i++)
-    {
-        zeros[i] = 0.f;
-        qzeros[i] = 0;
-    }
-    const TC xi = ix < kx ? *(const TC *)&x[iy * kx + ix] : zeros;
-    float sum = xi[0];
-    float amax = sycl::fabs(xi[0]);
-#pragma unroll
-    for (int i = 1; i < QUANT_BLOCK_TILE; i++)
-    {
-        sum += xi[i];
-        amax = sycl::fmax(sycl::fabs(xi[i]), amax);
-    }
-    sum = warp_reduce_sum(sum, item_ct1);
-    amax = warp_reduce_max(amax, item_ct1);
-
-    const float d = amax / 127;
-    TQ q = qzeros;
-    if (amax != 0.0f)
-    {
-#pragma unroll
-        for (int i = 0; i < QUANT_BLOCK_TILE; i++) {
-            q[i] = sycl::round(xi[i] / d);
-        }
-    }
-
-    *(TQ *)&y[ib].qs[iqs] = q;
-
-    if (iqs > 0) {
-        return;
-    }
-
-    reinterpret_cast<sycl::half &>(y[ib].ds.x()) = d;
-    reinterpret_cast<sycl::half &>(y[ib].ds.y()) = sum;
-}
-
-template <int ElementsPerWI>
-static __dpct_inline__ void quantize_and_reorder_q8_1(const float * __restrict__ x, void * reordered_q8_tensor,
-                                                      const int kx, const int kx_padded, const sycl::nd_item<1> & it) {
-    /*
-        Quantizes and reorders the resultant q8 tensor in a per row fashion
-        Each sub-group calculates one quant block. i.e. QK8_1 quant values and the d and sum values
-    */
-
-    auto subgroup_id = it.get_group(0);
-    auto wi_id       = it.get_local_id(0);
-
-    const int num_blocks_per_row = kx / QK8_1;
-    auto      row                = subgroup_id / num_blocks_per_row;
-    auto      col                = subgroup_id % num_blocks_per_row;
-
-    auto row_offset = row * (kx_padded / QK8_1) * sizeof(block_q8_1);
-    auto col_offset = QK8_1 * col + wi_id * ElementsPerWI;
-
-    auto quant_ptr = (int8_t *) ((char *) reordered_q8_tensor + row_offset + col_offset);
-    auto ds_ptr    = (sycl::half2 *) ((char *) reordered_q8_tensor + row_offset + kx + col * sizeof(sycl::half2));
-
-    sycl::vec<float, ElementsPerWI>  wi_f32_vals;
-    sycl::vec<int8_t, ElementsPerWI> quantized_values;
-
-    auto float_ptr_offset = subgroup_id * QK8_1 + ElementsPerWI * wi_id;
-    wi_f32_vals           = *reinterpret_cast<const sycl::vec<float, ElementsPerWI> *>(x + float_ptr_offset);
-
-    float sum  = 0.0f;
-    float amax = 0.0f;
-
-#pragma unroll(ElementsPerWI)
-    for (int i = 0; i < ElementsPerWI; i++) {
-        sum += wi_f32_vals[i];
-        amax                = sycl::fmax(amax, sycl::fabs(wi_f32_vals[i]));
-        quantized_values[i] = 0;
-    }
-    sum     = sycl::reduce_over_group(it.get_group(), sum, sycl::plus<float>());
-    amax    = sycl::reduce_over_group(it.get_group(), amax, sycl::maximum<float>());
-    float d = amax == 0 ? 1 : amax / 127;
-
-#pragma unroll(ElementsPerWI)
-    for (int i = 0; i < ElementsPerWI; i++) {
-        quantized_values[i] = sycl::round(wi_f32_vals[i] / d);
-    }
-
-    d = amax == 0 ? 0 : d;
-
-    *reinterpret_cast<sycl::vec<int8_t, ElementsPerWI> *>(quant_ptr) = quantized_values;
-    if (wi_id == 0) {
-        *ds_ptr = sycl::half2(sycl::half(d), sycl::half(sum));
-    }
-}
-
 static void mul_mat_p021_f16_f32(
     const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst,
     const int ncols_x, const int nrows_x, const int nchannels_x, const int nchannels_y,
@@ -1770,32 +1657,6 @@ static  void pool2d_nchw_kernel(
         o_ptr[cur_oh * ow + cur_ow] = res;
 }
 
-static void quantize_row_q8_1_sycl(const float * x, void * vy, const int kx, const int ky, const int kx_padded,
-                                   bool reorder_q8_tensor, queue_ptr stream) {
-    if (reorder_q8_tensor) {
-        auto local_range      = std::size_t(WARP_SIZE);
-        auto num_quant_blocks = ky * (kx / QK8_1);
-        auto global_range     = num_quant_blocks * local_range;
-        stream->parallel_for(sycl::nd_range<1>({ global_range }, { local_range }),
-                             [=](sycl::nd_item<1> it) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                 quantize_and_reorder_q8_1<QK8_1 / WARP_SIZE>(x, vy, kx, kx_padded, it);
-                             });
-    } else {
-        const int            block_num_x = (kx_padded + SYCL_QUANTIZE_BLOCK_SIZE - 1) / SYCL_QUANTIZE_BLOCK_SIZE;
-        const sycl::range<3> num_blocks(1, ky, block_num_x);
-        int constexpr QUANT_BLOCK_TILE = QK8_1 / WARP_SIZE;
-        static_assert(QK8_1 % WARP_SIZE == 0);
-        const sycl::range<3> block_size(1, 1, SYCL_QUANTIZE_BLOCK_SIZE / QUANT_BLOCK_TILE);
-        {
-            dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
-
-            stream->parallel_for(sycl::nd_range<3>(num_blocks * block_size, block_size),
-                                 [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                     quantize_q8_1<QUANT_BLOCK_TILE>(x, vy, kx, kx_padded, item_ct1);
-                                 });
-        }
-    }
-}
 
 static void ggml_mul_mat_p021_f16_f32_sycl(const void *vx, const float *y,
                                            float *dst, const int ncols_x,
@@ -2372,10 +2233,10 @@ static void ggml_sycl_set_peer_access(const int n_tokens, int main_device) {
     peer_access_enabled = enable_peer_access;
 }
 
+template <template <int> typename quantize_f>
 static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
                                  const ggml_tensor *src1, ggml_tensor *dst,
-                                 ggml_sycl_op_mul_mat_t op,
-                                 const bool convert_src1_to_q8_1) try {
+                                 ggml_sycl_op_mul_mat_t op) try {
 
     GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne);
 
@@ -2470,6 +2331,8 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
         }
     }
 
+    constexpr bool quantize_enabled = !std::is_same_v<quantize_f<QK8_1 / WARP_SIZE>,
+                                                      no_quantize_q8_1<QK8_1 / WARP_SIZE>>;
     for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
         if ((!split && i != ctx.device) || dev[i].row_low == dev[i].row_high) {
             continue;
@@ -2495,20 +2358,19 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
             dev[i].src1_ddf = dev[i].src1_ddf_alloc.alloc(ctx.pool(i), ggml_nelements(src1));
         }
 
-        if (convert_src1_to_q8_1) {
+        if constexpr(quantize_enabled) {
             dev[i].src1_ddq = dev[i].src1_ddq_alloc.alloc(ctx.pool(i), nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs);
 
             if (src1_on_device && src1_is_contiguous) {
-                bool reorder_q8_tensor = src0->extra && ((ggml_tensor_extra_gpu *)src0->extra)->optimized_feature.reorder;
                 scope_op_debug_print scope_dbg_print(__func__, "/quantize_row_q8_1_sycl", dst,
                                                      /*num_src=*/2, " : converting src1 to Q8_1");
-                quantize_row_q8_1_sycl(dev[i].src1_ddf, dev[i].src1_ddq, ne10, nrows1, src1_padded_col_size, reorder_q8_tensor, stream);
-                /*
-                DPCT1010:90: SYCL uses exceptions to report errors and does not
-                use the error codes. The call was replaced with 0. You need to
-                rewrite this code.
-                */
-                SYCL_CHECK(0);
+                try {
+                    quantize_row_q8_1_sycl<quantize_f>(dev[i].src1_ddf, dev[i].src1_ddq, ne10, nrows1, src1_padded_col_size, stream);
+                } catch (sycl::exception const &exc) {
+                    std::cerr << "Quantize_row_q8_1_sycl error" << exc.what() << "Exception caught at file:" << __FILE__
+                              << ", line:" << __LINE__ << std::endl;
+                    std::exit(1);
+                }
             }
         }
 
@@ -2524,11 +2386,6 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
     // here an event is recorded that signals that the main device has finished calculating the input data
     if (split && used_devices > 1) {
         ggml_sycl_set_device(ctx.device);
-        /*
-        DPCT1024:91: The original code returned the error code that was further
-        consumed by the program logic. This original code was replaced with 0.
-        You may need to rewrite the program logic consuming the error code.
-        */
         SYCL_CHECK(CHECK_TRY_ERROR(
             *src0_extra->events[ctx.device][0] =
                 ctx.stream()->ext_oneapi_submit_barrier()));
@@ -2552,11 +2409,6 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
 
             // wait for main GPU data if necessary
             if (split && (i != ctx.device || is != 0)) {
-                /*
-                DPCT1009:163: SYCL uses exceptions to report errors and does not
-                use the error codes. The original code was commented out and a
-                warning string was inserted. You need to rewrite this code.
-                */
                 SYCL_CHECK(CHECK_TRY_ERROR(stream->ext_oneapi_submit_barrier(
                     {*src0_extra->events[ctx.device][0]})));
             }
@@ -2582,39 +2434,42 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
                 // copy src0, src1 to device if necessary
                 if (src1_is_contiguous) {
                     if (i != ctx.device) {
-                        if (convert_src1_to_q8_1) {
+                        if constexpr (quantize_enabled) {
                             char * src1_ddq_i_source = dev[ctx.device].src1_ddq + src1_ddq_i_offset;
-                          SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy(
-                                src1_ddq_i, src1_ddq_i_source,
-                                src1_ncols * src1_padded_col_size * q8_1_ts /
-                                    q8_1_bs).wait()));
+                            SYCL_CHECK(
+                                CHECK_TRY_ERROR(stream
+                                                    ->memcpy(src1_ddq_i, src1_ddq_i_source,
+                                                             src1_ncols * src1_padded_col_size * q8_1_ts / q8_1_bs)
+                                                    .wait()));
                         } else {
-
                             float * src1_ddf_i_source = (float *) src1_extra->data_device[ctx.device];
-                            src1_ddf_i_source += (i0*ne11 + src1_col_0) * ne10;
+                            src1_ddf_i_source += (i0 * ne11 + src1_col_0) * ne10;
 
-                            SYCL_CHECK(CHECK_TRY_ERROR(dev2dev_memcpy(*stream, *main_stream,
-                                src1_ddf_i, src1_ddf_i_source,
-                                src1_ncols * ne10 * sizeof(float))));
+                            SYCL_CHECK(
+                                CHECK_TRY_ERROR(dev2dev_memcpy(*stream, *main_stream, src1_ddf_i, src1_ddf_i_source,
+                                                               src1_ncols * ne10 * sizeof(float))));
                         }
                     }
-                } else if (src1_on_device && !src1_is_contiguous) {
-                    SYCL_CHECK(ggml_sycl_cpy_tensor_2d(
-                                   src1_ddf_i, src1, i03, i02, src1_col_0, src1_col_0+src1_ncols, stream));
                 } else {
-                    GGML_ABORT("fatal error");
-                }
+                    if (src1_on_device) {
+                        SYCL_CHECK(ggml_sycl_cpy_tensor_2d(src1_ddf_i, src1, i03, i02, src1_col_0,
+                                                           src1_col_0 + src1_ncols, stream));
+                    } else {
+                        GGML_ABORT("src1 is non-contiguous and not on device");
+                    }
 
-                if (convert_src1_to_q8_1 && !src1_is_contiguous) {
-                    scope_op_debug_print scope_dbg_print(__func__, "/quantize_row_q8_1_sycl", dst,
-                                                         /*num_src=*/2, " : converting src1 to Q8_1");
-                    quantize_row_q8_1_sycl(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, src1_padded_col_size, false, stream);
-                    /*
-                    DPCT1010:92: SYCL uses exceptions to report errors and does
-                    not use the error codes. The call was replaced with 0. You
-                    need to rewrite this code.
-                    */
-                    SYCL_CHECK(0);
+                    if constexpr (quantize_enabled) {
+                        scope_op_debug_print scope_dbg_print(__func__, "/quantize_row_q8_1_sycl", dst,
+                                                             /*num_src=*/2, " : converting src1 to Q8_1");
+                        try {
+                            quantize_row_q8_1_sycl<quantize_q8_1>(src1_ddf_i, src1_ddq_i, ne10, src1_ncols,
+                                                                  src1_padded_col_size, stream);
+                        } catch (const sycl::exception & exc) {
+                            std::cerr << "Quantize_row_q8_1_sycl error" << exc.what()
+                                      << "Exception caught at file:" << __FILE__ << ", line:" << __LINE__ << std::endl;
+                            std::exit(1);
+                        }
+                    }
                 }
 
                 if (src1_col_0 == 0 && !src0_is_contiguous && i02 % i02_divisor == 0) {
@@ -2626,12 +2481,6 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
                 // do the computation
                 SYCL_CHECK(CHECK_TRY_ERROR(op(ctx, src0, src1, dst, src0_dd_i, src1_ddf_i, src1_ddq_i, dst_dd_i,
                     dev[i].row_low, dev[i].row_high, src1_ncols, src1_padded_col_size, stream)));
-                /*
-                DPCT1010:93: SYCL uses exceptions to report errors and does not
-                use the error codes. The call was replaced with 0. You need to
-                rewrite this code.
-                */
-                SYCL_CHECK(0);
 
                 // copy dst to host or other device if necessary
                 if (!dst_on_device) {
@@ -2662,12 +2511,6 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
 
                 // add event for the main device to wait on until other device is done
                 if (split && (i != ctx.device || is != 0)) {
-                    /*
-                    DPCT1024:94: The original code returned the error code that
-                    was further consumed by the program logic. This original
-                    code was replaced with 0. You may need to rewrite the
-                    program logic consuming the error code.
-                    */
                     SYCL_CHECK(CHECK_TRY_ERROR(
                         *src0_extra->events[i][is] =
                             stream->ext_oneapi_submit_barrier()));
@@ -3351,19 +3194,20 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
         // KQ + KQV multi-batch
         ggml_sycl_mul_mat_batched_sycl(ctx, src0, src1, dst);
     } else if (use_dequantize_mul_mat_vec) {
-        constexpr bool convert_src1_to_q8_1 = false;
         opt_for_reorder(&ctx, src0, src1, dst, mul_mat_algo::DMMV);
-        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_dequantize_mul_mat_vec, convert_src1_to_q8_1);
+        ggml_sycl_op_mul_mat<no_quantize_q8_1>(ctx, src0, src1, dst, ggml_sycl_op_dequantize_mul_mat_vec);
     } else if (use_mul_mat_vec_q) {
-        constexpr bool convert_src1_to_q8_1 = true;
         opt_for_reorder(&ctx, src0, src1, dst, mul_mat_algo::MMVQ);
-        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_vec_q, convert_src1_to_q8_1);
+        ggml_tensor_extra_gpu * extra = static_cast<ggml_tensor_extra_gpu *>(src0->extra);
+        if (extra && extra->optimized_feature.reorder) {
+            ggml_sycl_op_mul_mat<quantize_and_reorder_q8_1_soa>(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_vec_q);
+        } else {
+            ggml_sycl_op_mul_mat<quantize_q8_1>(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_vec_q);
+        }
     } else if (use_mul_mat_q) {
-        constexpr bool convert_src1_to_q8_1 = true;
-        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_q, convert_src1_to_q8_1);
+        ggml_sycl_op_mul_mat<quantize_q8_1>(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_q);
     } else {
-        constexpr bool convert_src1_to_q8_1 = false;
-        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_sycl, convert_src1_to_q8_1);
+        ggml_sycl_op_mul_mat<no_quantize_q8_1>(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_sycl);
     }
 }
 
diff --git a/ggml/src/ggml-sycl/quantize.hpp b/ggml/src/ggml-sycl/quantize.hpp
new file mode 100644
index 00000000000..b5c7a54b792
--- /dev/null
+++ b/ggml/src/ggml-sycl/quantize.hpp
@@ -0,0 +1,133 @@
+/***************************************************************************
+ *
+ *  Copyright (C) 2025 Codeplay Software Ltd.
+ *  Copyright (C) 2025 Intel Corporation
+ *
+ *  MIT License
+ *
+ *  Unless required by applicable law or agreed to in writing, software
+ *  distributed under the License is distributed on an "AS IS" BASIS,
+ *  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ *  See the License for the specific language governing permissions and
+ *  limitations under the License.
+ *
+ *  quantize.hpp
+ *
+ *  Description:
+ *     Sycl backend specific quantization functions
+ **************************************************************************/
+
+#pragma once
+
+#include <sycl/nd_item.hpp>
+
+#include "ggml-sycl/dpct/helper.hpp"
+
+template <int ElementsPerWI>
+__dpct_inline__ static void quantize_q8_1_impl(const float * __restrict__ x,
+                                               sycl::vec<int8_t, ElementsPerWI> & quantized_values, float & d,
+                                               float & sum, const sycl::nd_item<1> & it) {
+    auto subgroup_id = it.get_group(0);
+    auto wi_id       = it.get_local_id(0);
+
+    sycl::vec<float, ElementsPerWI> wi_f32_vals;
+
+    auto float_ptr_offset = subgroup_id * QK8_1 + ElementsPerWI * wi_id;
+    wi_f32_vals           = *reinterpret_cast<const sycl::vec<float, ElementsPerWI> *>(x + float_ptr_offset);
+
+    float amax = 0.0f;
+
+#pragma unroll(ElementsPerWI)
+    for (int i = 0; i < ElementsPerWI; i++) {
+        sum += wi_f32_vals[i];
+        amax                = sycl::fmax(amax, sycl::fabs(wi_f32_vals[i]));
+        quantized_values[i] = 0;
+    }
+    sum  = sycl::reduce_over_group(it.get_sub_group(), sum, sycl::plus<float>());
+    amax = sycl::reduce_over_group(it.get_sub_group(), amax, sycl::maximum<float>());
+    d    = amax == 0 ? 1 : amax / 127;
+
+#pragma unroll(ElementsPerWI)
+    for (int i = 0; i < ElementsPerWI; i++) {
+        quantized_values[i] = sycl::round(wi_f32_vals[i] / d);
+    }
+
+    d = amax == 0 ? 0 : d;
+}
+
+// No op to control codepath in ggml_sycl_op_mul_mat
+template <int ElementsPerWI> struct no_quantize_q8_1 {
+    void operator()(const float *, void *, int, int, const sycl::nd_item<1> &) const {}
+};
+
+template <int ElementsPerWI> struct quantize_and_reorder_q8_1_soa {
+    __dpct_inline__ void operator()(const float * __restrict__ x, void * reordered_q8_tensor, const int kx,
+                                    const int kx_padded, const sycl::nd_item<1> & it) const {
+        /*
+        Quantizes and reorders the resultant q8 tensor in a per row fashion
+        Each sub-group calculates one quant block. i.e. QK8_1 quant values and the d and sum values
+    */
+        auto subgroup_id = it.get_group(0);
+        auto wi_id       = it.get_local_id(0);
+
+        sycl::vec<int8_t, ElementsPerWI> quantized_values;
+        float                            d   = 0.0f;
+        float                            sum = 0.0f;
+        quantize_q8_1_impl<ElementsPerWI>(x, quantized_values, d, sum, it);
+
+        const int num_blocks_per_row = kx / QK8_1;
+        auto      row                = subgroup_id / num_blocks_per_row;
+        auto      col                = subgroup_id % num_blocks_per_row;
+        auto      row_offset         = row * (kx_padded / QK8_1) * sizeof(block_q8_1);
+        auto      col_offset         = QK8_1 * col + wi_id * ElementsPerWI;
+
+        auto quant_ptr = (int8_t *) ((char *) reordered_q8_tensor + row_offset + col_offset);
+        *reinterpret_cast<sycl::vec<int8_t, ElementsPerWI> *>(quant_ptr) = quantized_values;
+
+        auto ds_ptr = (sycl::half2 *) ((char *) reordered_q8_tensor + row_offset + kx + col * sizeof(sycl::half2));
+        if (wi_id == 0) {
+            *ds_ptr = sycl::half2(sycl::half(d), sycl::half(sum));
+        }
+    }
+};
+
+template <int ElementsPerWI> struct quantize_q8_1 {
+    __dpct_inline__ void operator()(const float * __restrict__ x, void * q8_tensor, const int kx, const int kx_padded,
+                                    const sycl::nd_item<1> & it) const {
+        auto subgroup_id = it.get_group(0);
+        auto wi_id       = it.get_local_id(0);
+
+        const int num_blocks_per_row = kx / QK8_1;
+        auto      row                = subgroup_id / num_blocks_per_row;
+        const int pitch              = kx_padded / QK8_1;
+
+        sycl::vec<int8_t, ElementsPerWI> quantized_values;
+        float                            d   = 0.0f;
+        float                            sum = 0.0f;
+        quantize_q8_1_impl<ElementsPerWI>(x, quantized_values, d, sum, it);
+
+        block_q8_1 * quant_ptr = (block_q8_1 *) q8_tensor;
+        auto         block_id  = subgroup_id % num_blocks_per_row + row * pitch;
+
+        int8_t * qs                                               = &(quant_ptr[block_id].qs[wi_id * ElementsPerWI]);
+        *reinterpret_cast<sycl::vec<int8_t, ElementsPerWI> *>(qs) = quantized_values;
+        if (wi_id == 0) {
+            quant_ptr[block_id].ds = sycl::half2(sycl::half(d), sycl::half(sum));
+        }
+    }
+};
+
+template <template <int> typename quantize_f>
+void quantize_row_q8_1_sycl(const float * x, void * vy, const int kx, const int ky, const int kx_padded,
+                            dpct::queue_ptr stream) {
+    static_assert(QK8_1 % WARP_SIZE == 0);
+    auto local_range      = std::size_t(WARP_SIZE);
+    auto num_quant_blocks = ky * (kx / QK8_1);
+    auto global_range     = num_quant_blocks * local_range;
+    dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
+
+    stream->parallel_for(sycl::nd_range<1>({ global_range }, { local_range }),
+                         [=](sycl::nd_item<1> it) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                             quantize_f<QK8_1 / WARP_SIZE>()(x, vy, kx, kx_padded, it);
+                         });
+}

From 5cc0448cc0fe148d8804297f74c34eb9c95d30e6 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= <johannesg@5d6.de>
Date: Mon, 28 Jul 2025 14:30:22 +0200
Subject: [PATCH 05/86] CUDA: fix pointer incrementation in FA (llama/14916)

---
 ggml/src/ggml-cuda/fattn-vec-f16.cuh | 9 ++++-----
 ggml/src/ggml-cuda/fattn-vec-f32.cuh | 9 ++++-----
 2 files changed, 8 insertions(+), 10 deletions(-)

diff --git a/ggml/src/ggml-cuda/fattn-vec-f16.cuh b/ggml/src/ggml-cuda/fattn-vec-f16.cuh
index e9b5c306365..109253838f2 100644
--- a/ggml/src/ggml-cuda/fattn-vec-f16.cuh
+++ b/ggml/src/ggml-cuda/fattn-vec-f16.cuh
@@ -174,7 +174,10 @@ static __global__ void flash_attn_vec_ext_f16(
     K     += blockIdx.y*D * nb11;
     V     += blockIdx.y*D * nb21;
     maskh += blockIdx.y*D;
-    for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*D) {
+    for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*D,
+             // Increment pointers after each loop:
+             K += gridDim.y*D*nb11, V += gridDim.y*D*nb21, maskh += gridDim.y*D) {
+
         // Calculate KQ tile and keep track of new maximum KQ values:
 
         if (mask) {
@@ -291,10 +294,6 @@ static __global__ void flash_attn_vec_ext_f16(
             }
         }
 
-        K     += gridDim.y*D * nb11;
-        V     += gridDim.y*D * nb21;
-        maskh += gridDim.y*D;
-
         __syncthreads();
     }
 
diff --git a/ggml/src/ggml-cuda/fattn-vec-f32.cuh b/ggml/src/ggml-cuda/fattn-vec-f32.cuh
index 6a4bdc0ff9a..2cf2e408e92 100644
--- a/ggml/src/ggml-cuda/fattn-vec-f32.cuh
+++ b/ggml/src/ggml-cuda/fattn-vec-f32.cuh
@@ -180,7 +180,10 @@ static __global__ void flash_attn_vec_ext_f32(
     K     += blockIdx.y*D * nb11;
     V     += blockIdx.y*D * nb21;
     maskh += blockIdx.y*D;
-    for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*D) {
+    for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*D,
+             // Increment pointers after each loop:
+             K += gridDim.y*D*nb11, V += gridDim.y*D*nb21, maskh += gridDim.y*D) {
+
         // Calculate KQ tile and keep track of new maximum KQ values:
 
         if (mask) {
@@ -286,10 +289,6 @@ static __global__ void flash_attn_vec_ext_f32(
             }
         }
 
-        K     += gridDim.y*D * nb11;
-        V     += gridDim.y*D * nb21;
-        maskh += gridDim.y*D;
-
         __syncthreads();
     }
 

From 8ed95829fba09c84c4bd1def42327f1510256721 Mon Sep 17 00:00:00 2001
From: Akarshan Biswas <akarshan@menlo.ai>
Date: Mon, 28 Jul 2025 20:32:15 +0530
Subject: [PATCH 06/86] SYCL: Add set_rows support for quantized types
 (llama/14883)

* SYCL: Add set_rows support for quantized types

This commit adds support for GGML_OP_SET_ROWS operation for various
quantized tensor types (Q8_0, Q5_1, Q5_0, Q4_1, Q4_0, IQ4_NL) and BF16
type in the SYCL backend.

The quantization/dequantization copy kernels were moved from cpy.cpp
to cpy.hpp to make them available for set_rows.cpp.

This addresses part of the TODOs mentioned in the code.

* Use get_global_linear_id() instead

ggml-ci

* Fix formatting

ggml-ci

* Use const for ne11 and size_t variables in set_rows_sycl_q

ggml-ci

* Increase block size for q kernel to 256

ggml-ci

* Cleanup imports

* Add float.h to cpy.hpp
---
 ggml/src/ggml-sycl/cpy.cpp       | 212 ------------------------------
 ggml/src/ggml-sycl/cpy.hpp       | 214 ++++++++++++++++++++++++++++++-
 ggml/src/ggml-sycl/ggml-sycl.cpp |  11 +-
 ggml/src/ggml-sycl/set_rows.cpp  |  94 ++++++++++++++
 4 files changed, 313 insertions(+), 218 deletions(-)

diff --git a/ggml/src/ggml-sycl/cpy.cpp b/ggml/src/ggml-sycl/cpy.cpp
index 1ffd7f12267..3d321b58ac6 100644
--- a/ggml/src/ggml-sycl/cpy.cpp
+++ b/ggml/src/ggml-sycl/cpy.cpp
@@ -1,31 +1,12 @@
 #include "cpy.hpp"
 
 #include <float.h>
-#include <string>
 
 #include "dequantize.hpp"
 #include "ggml-sycl/common.hpp"
 #include "ggml-sycl/presets.hpp"
 #include "ggml.h"
 
-static __dpct_inline__ int best_index_int8(int n, const int8_t * val, float x) {
-    if (x <= val[0]) {
-        return 0;
-    }
-    if (x >= val[n - 1]) {
-        return n - 1;
-    }
-    int ml = 0, mu = n - 1;
-    while (mu - ml > 1) {
-        int mav = (ml + mu) / 2;
-        if (x < val[mav]) {
-            mu = mav;
-        } else {
-            ml = mav;
-        }
-    }
-    return x - val[mu - 1] < val[mu] - x ? mu - 1 : mu;
-}
 
 static void cpy_1_f32_f32(const char * cxi, char * cdsti) {
     const float * xi   = (const float *) cxi;
@@ -97,28 +78,6 @@ static void cpy_f32_f16(const char * cx, char * cdst, const int ne, const int ne
     cpy_1(cx + x_offset, cdst + dst_offset);
 }
 
-static void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
-    const float * xi   = (const float *) cxi;
-    block_q8_0 *  dsti = (block_q8_0 *) cdsti;
-
-    float amax = 0.0f;  // absolute max
-
-    for (int j = 0; j < QK8_0; j++) {
-        const float v = xi[j];
-        amax          = sycl::fmax(amax, sycl::fabs((float) v));
-    }
-
-    const float d  = amax / ((1 << 7) - 1);
-    const float id = d ? 1.0f / d : 0.0f;
-
-    dsti->d = d;
-
-    for (int j = 0; j < QK8_0; ++j) {
-        const float x0 = xi[j] * id;
-
-        dsti->qs[j] = sycl::round((float) x0);
-    }
-}
 
 /* quantized type same copy */
 template<typename T>
@@ -140,178 +99,7 @@ static void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) {
     }
 }
 
-static void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) {
-    const float * xi   = (const float *) cxi;
-    block_q4_0 *  dsti = (block_q4_0 *) cdsti;
-
-    float amax = 0.0f;
-    float vmax = 0.0f;
-
-    for (int j = 0; j < QK4_0; ++j) {
-        const float v = xi[j];
-        if (amax < sycl::fabs((float) v)) {
-            amax = sycl::fabs((float) v);
-            vmax = v;
-        }
-    }
-
-    const float d  = vmax / -8;
-    const float id = d ? 1.0f / d : 0.0f;
-
-    dsti->d = d;
-
-    for (int j = 0; j < QK4_0 / 2; ++j) {
-        const float x0 = xi[0 + j] * id;
-        const float x1 = xi[QK4_0 / 2 + j] * id;
-
-        const uint8_t xi0 = dpct::min(15, (int8_t) (x0 + 8.5f));
-        const uint8_t xi1 = dpct::min(15, (int8_t) (x1 + 8.5f));
-
-        dsti->qs[j] = xi0;
-        dsti->qs[j] |= xi1 << 4;
-    }
-}
-
-static void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) {
-    const float * xi   = (const float *) cxi;
-    block_q4_1 *  dsti = (block_q4_1 *) cdsti;
-
-    float vmin = FLT_MAX;
-    float vmax = -FLT_MAX;
-
-    for (int j = 0; j < QK4_1; ++j) {
-        const float v = xi[j];
-
-        if (v < vmin) {
-            vmin = v;
-        }
-        if (v > vmax) {
-            vmax = v;
-        }
-    }
-
-    const float d  = (vmax - vmin) / ((1 << 4) - 1);
-    const float id = d ? 1.0f / d : 0.0f;
-
-    dsti->dm.x() = d;
-    dsti->dm.y() = vmin;
-
-    for (int j = 0; j < QK4_1 / 2; ++j) {
-        const float x0 = (xi[0 + j] - vmin) * id;
-        const float x1 = (xi[QK4_1 / 2 + j] - vmin) * id;
-
-        const uint8_t xi0 = dpct::min(15, (int8_t) (x0 + 0.5f));
-        const uint8_t xi1 = dpct::min(15, (int8_t) (x1 + 0.5f));
 
-        dsti->qs[j] = xi0;
-        dsti->qs[j] |= xi1 << 4;
-    }
-}
-
-static void cpy_blck_f32_q5_0(const char * cxi, char * cdsti) {
-    const float * xi   = (const float *) cxi;
-    block_q5_0 *  dsti = (block_q5_0 *) cdsti;
-
-    float amax = 0.0f;
-    float vmax = 0.0f;
-
-    for (int j = 0; j < QK5_0; ++j) {
-        const float v = xi[j];
-        if (amax < sycl::fabs((float) v)) {
-            amax = sycl::fabs((float) v);
-            vmax = v;
-        }
-    }
-
-    const float d  = vmax / -16;
-    const float id = d ? 1.0f / d : 0.0f;
-
-    dsti->d = d;
-
-    uint32_t qh = 0;
-    for (int j = 0; j < QK5_0 / 2; ++j) {
-        const float x0 = xi[0 + j] * id;
-        const float x1 = xi[QK5_0 / 2 + j] * id;
-
-        const uint8_t xi0 = dpct::min(31, (int8_t) (x0 + 16.5f));
-        const uint8_t xi1 = dpct::min(31, (int8_t) (x1 + 16.5f));
-
-        dsti->qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
-        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
-        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0 / 2);
-    }
-    memcpy(dsti->qh, &qh, sizeof(qh));
-}
-
-static void cpy_blck_f32_q5_1(const char * cxi, char * cdsti) {
-    const float * xi   = (const float *) cxi;
-    block_q5_1 *  dsti = (block_q5_1 *) cdsti;
-
-    float min = xi[0];
-    float max = xi[0];
-
-    for (int j = 1; j < QK5_1; ++j) {
-        const float v = xi[j];
-        min           = v < min ? v : min;
-        max           = v > max ? v : max;
-    }
-
-    const float d  = (max - min) / 31;
-    const float id = d ? 1.0f / d : 0.0f;
-
-    dsti->dm.x() = d;
-    dsti->dm.y() = min;
-
-    uint32_t qh = 0;
-    for (int j = 0; j < QK5_1 / 2; ++j) {
-        const float x0 = (xi[0 + j] - min) * id;
-        const float x1 = (xi[QK5_1 / 2 + j] - min) * id;
-
-        const uint8_t xi0 = (uint8_t) (x0 + 0.5f);
-        const uint8_t xi1 = (uint8_t) (x1 + 0.5f);
-
-        dsti->qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
-        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
-        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1 / 2);
-    }
-    memcpy(dsti->qh, &qh, sizeof(qh));
-}
-
-static void cpy_blck_f32_iq4_nl(const char * cxi, char * cdsti) {
-    const float *  xi   = (const float *) cxi;
-    block_iq4_nl * dsti = (block_iq4_nl *) cdsti;
-
-    float amax = 0.0f;
-    float vmax = 0.0f;
-
-    for (int j = 0; j < QK4_NL; ++j) {
-        const float v = xi[j];
-        if (amax < sycl::fabs((float) v)) {
-            amax = sycl::fabs((float) v);
-            vmax = v;
-        }
-    }
-
-    float       d  = vmax / kvalues_iq4nl[0];
-    const float id = d ? 1.0f / d : 0.0f;
-
-    float sumqx = 0, sumq2 = 0;
-    for (int j = 0; j < QK4_NL / 2; ++j) {
-        const float   x0  = xi[0 + j] * id;
-        const float   x1  = xi[QK4_NL / 2 + j] * id;
-        const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl, x0);
-        const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl, x1);
-        dsti->qs[j]       = xi0 | (xi1 << 4);
-        const float v0    = kvalues_iq4nl[xi0];
-        const float v1    = kvalues_iq4nl[xi1];
-        const float w0    = xi[0 + j] * xi[0 + j];
-        const float w1    = xi[QK4_NL / 2 + j] * xi[QK4_NL / 2 + j];
-        sumqx += w0 * v0 * xi[j] + w1 * v1 * xi[QK4_NL / 2 + j];
-        sumq2 += w0 * v0 * v0 + w1 * v1 * v1;
-    }
-
-    dsti->d = sumq2 > 0 ? sumqx / sumq2 : d;
-}
 
 template <dequantize_kernel_t dequant, int qk> static void cpy_blck_q_f32(const char * cxi, char * cdsti) {
     float * cdstf = (float *) (cdsti);
diff --git a/ggml/src/ggml-sycl/cpy.hpp b/ggml/src/ggml-sycl/cpy.hpp
index 0a0f561d230..3c331f1ef27 100644
--- a/ggml/src/ggml-sycl/cpy.hpp
+++ b/ggml/src/ggml-sycl/cpy.hpp
@@ -2,10 +2,222 @@
 #define GGML_SYCL_CPY_HPP
 
 #include "common.hpp"
+#include <float.h>
 
 typedef void (*cpy_kernel_t)(const char * cx, char * cdst);
 
+__dpct_inline__ int best_index_int8(int n, const int8_t * val, float x) {
+    if (x <= val[0]) {
+        return 0;
+    }
+    if (x >= val[n - 1]) {
+        return n - 1;
+    }
+    int ml = 0, mu = n - 1;
+    while (mu - ml > 1) {
+        int mav = (ml + mu) / 2;
+        if (x < val[mav]) {
+            mu = mav;
+        } else {
+            ml = mav;
+        }
+    }
+    return x - val[mu - 1] < val[mu] - x ? mu - 1 : mu;
+}
+
+inline void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
+    const float * xi   = (const float *) cxi;
+    block_q8_0 *  dsti = (block_q8_0 *) cdsti;
+
+    float amax = 0.0f;  // absolute max
+
+    for (int j = 0; j < QK8_0; j++) {
+        const float v = xi[j];
+        amax          = sycl::fmax(amax, sycl::fabs((float) v));
+    }
+
+    const float d  = amax / ((1 << 7) - 1);
+    const float id = d ? 1.0f / d : 0.0f;
+
+    dsti->d = d;
+
+    for (int j = 0; j < QK8_0; ++j) {
+        const float x0 = xi[j] * id;
+
+        dsti->qs[j] = sycl::round((float) x0);
+    }
+}
+
+inline void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) {
+    const float * xi   = (const float *) cxi;
+    block_q4_0 *  dsti = (block_q4_0 *) cdsti;
+
+    float amax = 0.0f;
+    float vmax = 0.0f;
+
+    for (int j = 0; j < QK4_0; ++j) {
+        const float v = xi[j];
+        if (amax < sycl::fabs((float) v)) {
+            amax = sycl::fabs((float) v);
+            vmax = v;
+        }
+    }
+
+    const float d  = vmax / -8;
+    const float id = d ? 1.0f / d : 0.0f;
+
+    dsti->d = d;
+
+    for (int j = 0; j < QK4_0 / 2; ++j) {
+        const float x0 = xi[0 + j] * id;
+        const float x1 = xi[QK4_0 / 2 + j] * id;
+
+        const uint8_t xi0 = dpct::min(15, (int8_t) (x0 + 8.5f));
+        const uint8_t xi1 = dpct::min(15, (int8_t) (x1 + 8.5f));
+
+        dsti->qs[j] = xi0;
+        dsti->qs[j] |= xi1 << 4;
+    }
+}
+
+inline void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) {
+    const float * xi   = (const float *) cxi;
+    block_q4_1 *  dsti = (block_q4_1 *) cdsti;
+
+    float vmin = FLT_MAX;
+    float vmax = -FLT_MAX;
+
+    for (int j = 0; j < QK4_1; ++j) {
+        const float v = xi[j];
+
+        vmin = sycl::min(v, vmin);
+        vmax = sycl::max(v, vmax);
+    }
+
+    const float d  = (vmax - vmin) / ((1 << 4) - 1);
+    const float id = d ? 1.0f / d : 0.0f;
+
+    dsti->dm.x() = d;
+    dsti->dm.y() = vmin;
+
+    for (int j = 0; j < QK4_1 / 2; ++j) {
+        const float x0 = (xi[0 + j] - vmin) * id;
+        const float x1 = (xi[QK4_1 / 2 + j] - vmin) * id;
+
+        const uint8_t xi0 = dpct::min(15, (int8_t) (x0 + 0.5f));
+        const uint8_t xi1 = dpct::min(15, (int8_t) (x1 + 0.5f));
+
+        dsti->qs[j] = xi0;
+        dsti->qs[j] |= xi1 << 4;
+    }
+}
+
+inline void cpy_blck_f32_q5_0(const char * cxi, char * cdsti) {
+    const float * xi   = (const float *) cxi;
+    block_q5_0 *  dsti = (block_q5_0 *) cdsti;
+
+    float amax = 0.0f;
+    float vmax = 0.0f;
+
+    for (int j = 0; j < QK5_0; ++j) {
+        const float v = xi[j];
+        if (amax < sycl::fabs((float) v)) {
+            amax = sycl::fabs((float) v);
+            vmax = v;
+        }
+    }
+
+    const float d  = vmax / -16;
+    const float id = d ? 1.0f / d : 0.0f;
+
+    dsti->d = d;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_0 / 2; ++j) {
+        const float x0 = xi[0 + j] * id;
+        const float x1 = xi[QK5_0 / 2 + j] * id;
+
+        const uint8_t xi0 = dpct::min(31, (int8_t) (x0 + 16.5f));
+        const uint8_t xi1 = dpct::min(31, (int8_t) (x1 + 16.5f));
+
+        dsti->qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0 / 2);
+    }
+    memcpy(dsti->qh, &qh, sizeof(qh));
+}
+
+inline void cpy_blck_f32_q5_1(const char * cxi, char * cdsti) {
+    const float * xi   = (const float *) cxi;
+    block_q5_1 *  dsti = (block_q5_1 *) cdsti;
+
+    float min = xi[0];
+    float max = xi[0];
+
+    for (int j = 1; j < QK5_1; ++j) {
+        const float v = xi[j];
+        min           = v < min ? v : min;
+        max           = v > max ? v : max;
+    }
+
+    const float d  = (max - min) / 31;
+    const float id = d ? 1.0f / d : 0.0f;
+
+    dsti->dm.x() = d;
+    dsti->dm.y() = min;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_1 / 2; ++j) {
+        const float x0 = (xi[0 + j] - min) * id;
+        const float x1 = (xi[QK5_1 / 2 + j] - min) * id;
+
+        const uint8_t xi0 = (uint8_t) (x0 + 0.5f);
+        const uint8_t xi1 = (uint8_t) (x1 + 0.5f);
+
+        dsti->qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1 / 2);
+    }
+    memcpy(dsti->qh, &qh, sizeof(qh));
+}
+
+inline void cpy_blck_f32_iq4_nl(const char * cxi, char * cdsti) {
+    const float *  xi   = (const float *) cxi;
+    block_iq4_nl * dsti = (block_iq4_nl *) cdsti;
+
+    float amax = 0.0f;
+    float vmax = 0.0f;
+
+    for (int j = 0; j < QK4_NL; ++j) {
+        const float v = xi[j];
+        if (amax < sycl::fabs((float) v)) {
+            amax = sycl::fabs((float) v);
+            vmax = v;
+        }
+    }
+
+    float       d  = vmax / kvalues_iq4nl[0];
+    const float id = d ? 1.0f / d : 0.0f;
+
+    float sumqx = 0, sumq2 = 0;
+    for (int j = 0; j < QK4_NL / 2; ++j) {
+        const float   x0  = xi[0 + j] * id;
+        const float   x1  = xi[QK4_NL / 2 + j] * id;
+        const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl, x0);
+        const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl, x1);
+        dsti->qs[j]       = xi0 | (xi1 << 4);
+        const float v0    = kvalues_iq4nl[xi0];
+        const float v1    = kvalues_iq4nl[xi1];
+        const float w0    = xi[0 + j] * xi[0 + j];
+        const float w1    = xi[QK4_NL / 2 + j] * xi[QK4_NL / 2 + j];
+        sumqx += w0 * v0 * xi[j] + w1 * v1 * xi[QK4_NL / 2 + j];
+        sumq2 += w0 * v0 * v0 + w1 * v1 * v1;
+    }
+
+    dsti->d = sumq2 > 0 ? sumqx / sumq2 : d;
+}
+
 void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1);
 void ggml_sycl_dup(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
-#endif // GGML_SYCL_CPY_HPP
+#endif  // GGML_SYCL_CPY_HPP
diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
index b08941c328b..2acdef98a6a 100644
--- a/ggml/src/ggml-sycl/ggml-sycl.cpp
+++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
@@ -4229,11 +4229,12 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
             }
         case GGML_OP_SET_ROWS:
             {
-                // TODO: add support
-                // ref: https://github.com/ggml-org/llama.cpp/pull/14274
-#pragma message("TODO: implement BF16, Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, IQ4_NL support (https://github.com/ggml-org/llama.cpp/pull/14661)")
-                return (op->type == GGML_TYPE_F32 || (op->type == GGML_TYPE_F16 && op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_I64));
-            } break;
+                return ((op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_BF16 ||
+                         op->type == GGML_TYPE_Q8_0 || op->type == GGML_TYPE_Q5_1 || op->type == GGML_TYPE_Q5_0 ||
+                         op->type == GGML_TYPE_Q4_1 || op->type == GGML_TYPE_Q4_0 || op->type == GGML_TYPE_IQ4_NL) &&
+                        (op->src[1]->type == GGML_TYPE_I64));
+            }
+            break;
         case GGML_OP_CPY:
             {
                 ggml_type src0_type = op->src[0]->type;
diff --git a/ggml/src/ggml-sycl/set_rows.cpp b/ggml/src/ggml-sycl/set_rows.cpp
index 3091fab3995..7a8e1410b70 100644
--- a/ggml/src/ggml-sycl/set_rows.cpp
+++ b/ggml/src/ggml-sycl/set_rows.cpp
@@ -1,4 +1,5 @@
 #include "set_rows.hpp"
+#include "cpy.hpp"
 
 namespace utils {
 template<typename T>
@@ -15,6 +16,68 @@ convert (const char* src, char* dst) {
    *reinterpret_cast<TOut*>(dst) = dst_val;
 }
 
+template <typename blockType, int qk, cpy_kernel_t cpyblck>
+static void set_rows_sycl_q(const char * __restrict__ src0_d,
+                            const int64_t * __restrict__ src1_d,
+                            blockType * __restrict__ dst_d,
+                            // tensor dimensions src0 and src1
+                            const int64_t ne00,
+                            const int64_t ne01,
+                            const int64_t ne02,
+                            const int64_t ne03,
+                            const int64_t ne10,
+                            const int64_t ne11,
+                            const int64_t ne12,
+                            const int64_t ne13,
+                            // strides for src0
+                            const size_t  nb00,
+                            const size_t  nb01,
+                            const size_t  nb02,
+                            const size_t  nb03,
+                            // strides for src1
+                            const size_t  nb10,
+                            const size_t  nb11,
+                            const size_t  nb12,
+                            const size_t  nb13,
+                            // strides for dst
+                            const size_t  nb1,
+                            const size_t  nb2,
+                            const size_t  nb3,
+                            queue_ptr     stream) {
+    const int64_t total_blocks = (ne00 * ne01 * ne02 * ne03) / qk;
+    constexpr int block_size   = 256;
+    const int64_t grid_size    = ceil_div(total_blocks, block_size);
+
+    sycl_parallel_for(stream, sycl::nd_range<1>(grid_size * block_size, block_size), [=](sycl::nd_item<1> item_ct1) {
+        const int64_t i = item_ct1.get_global_linear_id();
+        if (i >= total_blocks) {
+            return;
+        }
+        const int64_t i_base      = i * qk;
+        const int64_t i03         = i_base / (ne00 * ne01 * ne02);
+        const int64_t rem1        = i_base - i03 * (ne00 * ne01 * ne02);
+        const int64_t i02         = rem1 / (ne00 * ne01);
+        const int64_t rem2        = rem1 - i02 * ne00 * ne01;
+        const int64_t i01         = rem2 / ne00;
+        const int64_t i00         = rem2 - i01 * ne00;
+        const int64_t i12         = i03 % ne12;
+        const int64_t i11         = i02 % ne11;
+        const int64_t i10         = i01;
+        const size_t  src_offset  = calculate_offset<3>({ nb01, nb02, nb03 }, { i01, i02, i03 });
+        const char *  src_block   = src0_d + src_offset + i00 * sizeof(float);
+        const size_t  src1_offset = calculate_offset<3>({ nb10, nb11, nb12 }, { i10, i11, i12 });
+        const int64_t dst_row     = src1_d[src1_offset / sizeof(int64_t)];
+        const size_t  dst_offset =
+            calculate_offset<3>({ nb1, nb2, nb3 }, { dst_row, i02, i03 }) + (i00 / qk) * sizeof(blockType);
+        char * dst_block = reinterpret_cast<char *>(reinterpret_cast<char *>(dst_d) + dst_offset);
+        cpyblck(src_block, dst_block);
+    });
+    GGML_UNUSED(ne10);
+    GGML_UNUSED(ne13);
+    GGML_UNUSED(nb00);
+    GGML_UNUSED(nb13);
+}
+
 template<typename TIn, typename TOut>
 static void k_set_rows(
         const char * __restrict__ src0, const int64_t * __restrict__ src1, char * __restrict__ dst,
@@ -124,6 +187,37 @@ void ggml_sycl_op_set_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
                 stream
             );
             break;
+        case GGML_TYPE_BF16:
+            set_rows_sycl<float, sycl::ext::oneapi::bfloat16>(
+                (const char *)src0->data, src1_dd, (char *)dst->data,
+                ne00, ne01, ne02, ne03,
+                ne11, ne12,
+                nb01, nb02, nb03,
+                nb10, nb11, nb12,
+                nb1, nb2, nb3,
+                sizeof(float), sizeof(sycl::ext::oneapi::bfloat16),
+                stream
+            );
+            break;
+        case GGML_TYPE_Q8_0:
+            set_rows_sycl_q<block_q8_0, QK8_0, cpy_blck_f32_q8_0>((const char *)src0->data, src1_dd, (block_q8_0 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            set_rows_sycl_q<block_q5_1, QK5_1, cpy_blck_f32_q5_1>((const char *)src0->data, src1_dd, (block_q5_1 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            set_rows_sycl_q<block_q5_0, QK5_0, cpy_blck_f32_q5_0>((const char *)src0->data, src1_dd, (block_q5_0 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            set_rows_sycl_q<block_q4_1, QK4_1, cpy_blck_f32_q4_1>((const char *)src0->data, src1_dd, (block_q4_1 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q4_0:
+            set_rows_sycl_q<block_q4_0, QK4_0, cpy_blck_f32_q4_0>((const char *)src0->data, src1_dd, (block_q4_0 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_IQ4_NL:
+            set_rows_sycl_q<block_iq4_nl, QK4_NL, cpy_blck_f32_iq4_nl>((const char *)src0->data, src1_dd, (block_iq4_nl *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            break;
+
         default:
             GGML_ABORT("Unsupported tensor type!");
             break;

From 689060e181e1a49b31e3bf7228b262c68a9b1137 Mon Sep 17 00:00:00 2001
From: xctan <xc-tan@outlook.com>
Date: Mon, 28 Jul 2025 23:40:24 +0800
Subject: [PATCH 07/86] ggml-cpu : deduplicate scalar implementations
 (llama/14897)

* remove redundant code in riscv

* remove redundant code in arm

* remove redundant code in loongarch

* remove redundant code in ppc

* remove redundant code in s390

* remove redundant code in wasm

* remove redundant code in x86

* remove fallback headers

* fix x86 ggml_vec_dot_q8_0_q8_0
---
 ggml/src/ggml-cpu/arch/arm/quants.c       | 667 ++-----------------
 ggml/src/ggml-cpu/arch/arm/repack.cpp     | 300 +--------
 ggml/src/ggml-cpu/arch/loongarch/quants.c | 659 +++----------------
 ggml/src/ggml-cpu/arch/powerpc/quants.c   | 699 +++-----------------
 ggml/src/ggml-cpu/arch/riscv/quants.c     | 396 ++---------
 ggml/src/ggml-cpu/arch/riscv/repack.cpp   |  61 +-
 ggml/src/ggml-cpu/arch/s390/quants.c      | 367 ++---------
 ggml/src/ggml-cpu/arch/wasm/quants.c      | 368 ++---------
 ggml/src/ggml-cpu/arch/x86/quants.c       | 767 +++-------------------
 ggml/src/ggml-cpu/arch/x86/repack.cpp     | 265 +-------
 10 files changed, 562 insertions(+), 3987 deletions(-)

diff --git a/ggml/src/ggml-cpu/arch/arm/quants.c b/ggml/src/ggml-cpu/arch/arm/quants.c
index 3e2d3d03d67..c6d1d852e9a 100644
--- a/ggml/src/ggml-cpu/arch/arm/quants.c
+++ b/ggml/src/ggml-cpu/arch/arm/quants.c
@@ -1236,44 +1236,10 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = sumf;
 
 #else
-    const uint8_t pow3[6] = {1, 3, 9, 27, 81, 243};
-
-    float sumf = 0.0f;
-
-    for (int i = 0; i < nb; ++i) {
-        int sum = 0;
-
-        for (size_t j = 0; j < sizeof(x->qs) - sizeof(x->qs) % 32; j += 32) {
-            for (size_t l = 0; l < 5; ++l) {
-                for (size_t m = 0; m < 32; ++m) {
-                    uint8_t q = x[i].qs[j + m] * pow3[l];
-                    uint16_t xi = ((uint16_t) q * 3) >> 8;
-                    sum += (xi - 1) * y[i].qs[j*5 + l*32 + m];
-                }
-            }
-        }
-        for (size_t j = sizeof(x->qs) - sizeof(x->qs) % 32; j < sizeof(x->qs); j += 16) {
-            for (size_t l = 0; l < 5; ++l) {
-                for (size_t m = 0; m < 16; ++m) {
-                    uint8_t q = x[i].qs[j + m] * pow3[l];
-                    uint16_t xi = ((uint16_t) q * 3) >> 8;
-                    sum += (xi - 1) * y[i].qs[j*5 + l*16 + m];
-                }
-            }
-        }
-
-        for (size_t l = 0; l < 4; ++l) {
-            for (size_t j = 0; j < sizeof(x->qh); ++j) {
-                uint8_t q = x[i].qh[j] * pow3[l];
-                uint16_t xi = ((uint16_t) q * 3) >> 8;
-                sum += (xi - 1) * y[i].qs[sizeof(x->qs)*5 + l*sizeof(x->qh) + j];
-            }
-        }
-
-        sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d);
-    }
-
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_tq1_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1381,25 +1347,10 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = sumf;
 
 #else
-    float sumf = 0.0f;
-
-    for (int i = 0; i < nb; ++i) {
-        int32_t sumi = 0;
-
-        for (size_t j = 0; j < sizeof(x->qs); j += 32) {
-            for (size_t l = 0; l < 4; ++l) {
-                for (size_t k = 0; k < 32; ++k) {
-                    sumi += y[i].qs[j*4 + l*32 + k] * (((x[i].qs[j + k] >> (l*2)) & 3) - 1);
-                }
-            }
-        }
-
-        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
-
-        sumf += (float) sumi * d;
-    }
-
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_tq2_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1729,45 +1680,10 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sum;
 
 #else
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * q2 = x[i].qs;
-        const  int8_t * q8 = y[i].qs;
-        const uint8_t * sc = x[i].scales;
-
-        int summs = 0;
-        for (int j = 0; j < 16; ++j) {
-            summs += y[i].bsums[j] * (sc[j] >> 4);
-        }
-
-        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
-
-        int isum = 0;
-        int is = 0;
-        int d;
-        for (int k = 0; k < QK_K/128; ++k) {
-            int shift = 0;
-            for (int j = 0; j < 4; ++j) {
-                d = sc[is++] & 0xF;
-                int isuml = 0;
-                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                d = sc[is++] & 0xF;
-                isuml = 0;
-                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                shift += 2;
-                q8 += 32;
-            }
-            q2 += 32;
-        }
-        sumf += dall * isum - dmin * summs;
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2057,68 +1973,12 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sum;
 
 #else
-    // scalar version
-    // This function is written like this so the compiler can manage to vectorize most of it
-    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
-    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
-    // The ideal situation would be if we could just write the code once, and the compiler would
-    // automatically produce the best possible set of machine instructions, instead of us having to manually
-    // write vectorized versions for AVX, ARM_NEON, etc.
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    uint32_t auxs[4];
-    const int8_t * scales = (const int8_t*)auxs;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].hmask;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            q3 += 32;
-        }
-        a = aux8;
-
-        memcpy(auxs, x[i].scales, 12);
-        uint32_t tmp = auxs[2];
-        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
-        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
-        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
-        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
-        for (int j = 0; j < QK_K/16; ++j) {
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 
 }
@@ -2431,61 +2291,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sumf;
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            a += 32;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            a += 32; q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2578,66 +2391,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sumf;
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -3093,47 +2854,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     }
     *s = sum;
 #else
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].ql;
-        const uint8_t * GGML_RESTRICT qh = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) {
-                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
-                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
-                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
-                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
-            }
-            a  += 128;
-            q4 += 64;
-            qh += 32;
-        }
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/16; ++j) {
-            int scale = x[i].scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -3229,34 +2953,10 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
     *s = 0.25f * sumf;
 
 #else
-
-    uint32_t aux32[2];
-    const uint8_t * aux8 = (const uint8_t *)aux32;
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
-        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            memcpy(aux32, q2, 2*sizeof(uint32_t));
-            q2 += 4;
-            const uint32_t ls = 2*(aux32[1] >> 28) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
-                const uint8_t  signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.125f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -3327,42 +3027,10 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     *s = 0.125f * sumf;
 
 #else
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
-        const uint8_t  * GGML_RESTRICT sc = x[i].scales;
-        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1;
-            const uint16_t ls2 = 2*(sc[ib32] >>  4) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 2; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
-                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls1;
-            sumi = 0;
-            for (int l = 2; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
-                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls2;
-            q2 += 4;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.125f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -3455,45 +3123,10 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = 0.125f * sumf;
 
 #else
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const int8_t  * q8 = y[i].qs;
-        const uint8_t * qs = x[i].qs;
-        const uint8_t * qh = x[i].qh;
-        const uint8_t * signs = qs + QK_K/8;
-
-        int bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf);
-            int ls2 = 1 + 2*(x[i].scales[ib32] >>  4);
-            int sumi1 = 0, sumi2 = 0;
-            for (int l = 0; l < 2; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
-                for (int j = 0; j < 8; ++j) {
-                    sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            for (int l = 2; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
-                for (int j = 0; j < 8; ++j) {
-                    sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += ls1 * sumi1 + ls2 * sumi2;
-            qs += 4;
-            signs += 4;
-        }
-
-        sumf += d * bsum;
-    }
-
-    *s = 0.125f * sumf;
-
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 
 }
@@ -3553,36 +3186,10 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
     *s = 0.5f * sumf;
 
 #else
-
-    uint32_t aux32;
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
-        const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
-        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            memcpy(&aux32, gas, sizeof(uint32_t)); gas += sizeof(uint32_t);
-            const uint32_t ls = 2*(aux32 >> 28) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]);
-                const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]);
-                const uint8_t  signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            q3 += 8;
-            bsum += sumi * ls;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.25f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq3_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -3689,48 +3296,10 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = sumf;
 
 #else
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * GGML_RESTRICT qs = x[i].qs;
-        const uint8_t * GGML_RESTRICT qh = x[i].qh;
-        const uint8_t * GGML_RESTRICT signs = x[i].signs;
-        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const uint32_t ls1 = 2*(x[i].scales[ib32/2] & 0xf) + 1;
-            const uint32_t ls2 = 2*(x[i].scales[ib32/2] >>  4) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256)));
-                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256)));
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            qs += 8;
-            signs += 4;
-            bsum += sumi * ls1;
-            sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256)));
-                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256)));
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            qs += 8;
-            signs += 4;
-            bsum += sumi * ls2;
-        }
-        sumf += d * bsum;
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq3_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -3793,36 +3362,10 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = sumf;
 
 #else
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint16_t * qh = x[i].qh;
-
-        int sumi = 0, sumi1 = 0;
-        for (int ib = 0; ib < QK_K/32; ++ib) {
-            const int ls = 2*((qh[ib] >> 12) & 7) + 1;
-            const int delta = qh[ib] & 0x8000 ? -1 : 1;
-            int lsum = 0;
-            for (int l = 0; l < 4; ++l) {
-                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
-                for (int j = 0; j < 8; ++j) {
-                    lsum += q8[j] * grid[j];
-                }
-                q8 += 8;
-            }
-            sumi  += ls * lsum;
-            sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]);
-            qs += 4;
-        }
-
-        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
-    }
-
-    *s = sumf;
-
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq1_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -3912,52 +3455,11 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = sumf;
 
 #else
-
-    int sum1[2], sum2[2], delta[4];
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint8_t  * qh = x[i].qh;
-        const uint16_t * sc = (const uint16_t *)x[i].scales;
-
-        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
-
-        int sumi1 = 0, sumi2 = 0;
-        for (int ib = 0; ib < QK_K/32; ++ib) {
-            delta[0] = qh[0] & 0x08 ? -1 : 1;
-            delta[1] = qh[0] & 0x80 ? -1 : 1;
-            delta[2] = qh[1] & 0x08 ? -1 : 1;
-            delta[3] = qh[1] & 0x80 ? -1 : 1;
-            sum1[0] = sum1[1] = sum2[0] = sum2[1] = 0;
-            for (int l = 0; l < 4; ++l) {
-                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((uint16_t)qh[l/2] << (8 - 4*(l%2))) & 0x700)));
-                int lsum1 = 0, lsum2 = 0;
-                for (int j = 0; j < 8; ++j) {
-                    lsum1 += q8[j] * grid[j];
-                    lsum2 += q8[j];
-                }
-                q8 += 8;
-                sum1[l/2] += lsum1;
-                sum2[l/2] += lsum2*delta[l];
-            }
-
-            const int ls1 = 2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1;
-            const int ls2 = 2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1;
-
-            sumi1 += sum1[0] * ls1 + sum1[1] * ls2;
-            sumi2 += sum2[0] * ls1 + sum2[1] * ls2;
-            qs += 4;
-            qh += 2;
-        }
-
-        sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
-    }
-
-    *s = sumf;
-
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(scale);
+    ggml_vec_dot_iq1_m_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -4078,37 +3580,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     *s = sumf;
 
 #else
-    float sumf = 0;
-    for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
-        uint16_t h = x[ibl].scales_h;
-        const uint8_t * qs = x[ibl].qs;
-        const int8_t  * q8 = y[ibl].qs;
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
-            const uint8_t ls2 = (x[ibl].scales_l[ib/2] >>  4) | ((h << 2) & 0x30);
-            h >>= 4;
-            const float d1 = d4d8*(ls1 - 32);
-            const float d2 = d4d8*(ls2 - 32);
-            int sumi1 = 0, sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d1 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
-            sumi1 = sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d2 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
-        }
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
diff --git a/ggml/src/ggml-cpu/arch/arm/repack.cpp b/ggml/src/ggml-cpu/arch/arm/repack.cpp
index 2f8bc9e2517..fdd0a513b83 100644
--- a/ggml/src/ggml-cpu/arch/arm/repack.cpp
+++ b/ggml/src/ggml-cpu/arch/arm/repack.cpp
@@ -86,35 +86,9 @@ void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTR
         }
     }
 #else
-    // scalar
-    const int blck_size_interleave = 4;
-    float srcv[4][QK8_0];
-    float id[4];
-
-    for (int i = 0; i < nb; i++) {
-        for (int row_iter = 0; row_iter < 4; row_iter++) {
-            float amax = 0.0f; // absolute max
-
-            for (int j = 0; j < QK8_0; j++) {
-                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
-                amax = MAX(amax, fabsf(srcv[row_iter][j]));
-            }
-
-            const float d = amax / ((1 << 7) - 1);
-            id[row_iter] = d ? 1.0f / d : 0.0f;
-
-            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
-        }
-
-        for (int j = 0; j < QK8_0 * 4; j++) {
-            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
-            int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
-            src_offset += (j % blck_size_interleave);
-
-            float x0 = srcv[src_id][src_offset] * id[src_id];
-            y[i].qs[j] = roundf(x0);
-        }
-    }
+    UNUSED(nb);
+    UNUSED(y);
+    ggml_quantize_mat_q8_0_4x4_generic(x, vy, k);
 #endif
 }
 
@@ -205,35 +179,9 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
     }
 
 #else
-    // scalar
-    const int blck_size_interleave = 8;
-    float srcv[4][QK8_0];
-    float id[4];
-
-    for (int i = 0; i < nb; i++) {
-        for (int row_iter = 0; row_iter < 4; row_iter++) {
-            float amax = 0.0f; // absolute max
-
-            for (int j = 0; j < QK8_0; j++) {
-                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
-                amax = MAX(amax, fabsf(srcv[row_iter][j]));
-            }
-
-            const float d = amax / ((1 << 7) - 1);
-            id[row_iter] = d ? 1.0f / d : 0.0f;
-
-            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
-        }
-
-        for (int j = 0; j < QK8_0 * 4; j++) {
-            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
-            int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
-            src_offset += (j % blck_size_interleave);
-
-            float x0 = srcv[src_id][src_offset] * id[src_id];
-            y[i].qs[j] = roundf(x0);
-        }
-    }
+    UNUSED(nb);
+    UNUSED(y);
+    ggml_quantize_mat_q8_0_4x8_generic(x, vy, k);
 #endif
 }
 
@@ -295,29 +243,7 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     }
     return;
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
-    float sumf[4];
-    int sumi;
-
-    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-    for (int x = 0; x < nc / ncols_interleaved; x++) {
-        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
-
-        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-        for (int l = 0; l < nb; l++) {
-            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    sumi = 0;
-                    for (int i = 0; i < blocklen; ++i) {
-                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
-                    }
-                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
-                }
-            }
-        }
-        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
-    }
+    ggml_gemv_q4_0_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -383,29 +309,7 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     }
     return;
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
-    float sumf[4];
-    int sumi;
-
-    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-    for (int x = 0; x < nc / ncols_interleaved; x++) {
-        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
-
-        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-        for (int l = 0; l < nb; l++) {
-            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    sumi = 0;
-                    for (int i = 0; i < blocklen; ++i) {
-                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
-                    }
-                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
-                }
-            }
-        }
-        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
-    }
+    ggml_gemv_q4_0_4x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -497,31 +401,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
 #endif // #if defined(__ARM_FEATURE_SVE)
 
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
-    {
-        float sumf[8];
-        int sumi;
-
-        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
-
-            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int j = 0; j < ncols_interleaved; j++) {
-                        sumi = 0;
-                        for (int i = 0; i < blocklen; ++i) {
-                            const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
-                        }
-                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
-                    }
-                }
-            }
-            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
-        }
-    }
+    ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -591,31 +471,7 @@ void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
     }
     return;
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
-    {
-        float sumf[4];
-        int sumi;
-
-        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
-
-            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int j = 0; j < ncols_interleaved; j++) {
-                        sumi = 0;
-                        for (int i = 0; i < blocklen; ++i) {
-                            const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
-                            const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
-                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
-                        }
-                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
-                    }
-                }
-            }
-            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
-        }
-    }
+    ggml_gemv_iq4_nl_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -1096,40 +952,7 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     );
     return;
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
-    {
-        float sumf[4][4];
-        int sumi;
-
-        for (int y = 0; y < nr / 4; y++) {
-            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-            for (int x = 0; x < nc / ncols_interleaved; x++) {
-                const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
-                for (int m = 0; m < 4; m++) {
-                    for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
-                }
-                for (int l = 0; l < nb; l++) {
-                    for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                        for (int m = 0; m < 4; m++) {
-                            for (int j = 0; j < ncols_interleaved; j++) {
-                                sumi = 0;
-                                for (int i = 0; i < blocklen; ++i) {
-                                    const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                                    const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                                    sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
-                                }
-                                sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
-                            }
-                        }
-                    }
-                }
-                for (int m = 0; m < 4; m++) {
-                    for (int j = 0; j < ncols_interleaved; j++)
-                        s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
-                }
-            }
-        }
-    }
+    ggml_gemm_q4_0_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -1550,38 +1373,7 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     );
     return;
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
-    float sumf[4][4];
-    int sumi;
-
-    for (int y = 0; y < nr / 4; y++) {
-        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
-            }
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int m = 0; m < 4; m++) {
-                        for (int j = 0; j < ncols_interleaved; j++) {
-                            sumi = 0;
-                            for (int i = 0; i < blocklen; ++i) {
-                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                        (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
-                            }
-                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
-                        }
-                    }
-                }
-            }
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++)
-                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
-            }
-        }
-    }
+    ggml_gemm_q4_0_4x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -2019,38 +1811,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
 #endif // #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
 
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
-    float sumf[4][8];
-    int sumi;
-
-    for (int y = 0; y < nr / 4; y++) {
-        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
-            }
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int m = 0; m < 4; m++) {
-                        for (int j = 0; j < ncols_interleaved; j++) {
-                            sumi = 0;
-                            for (int i = 0; i < blocklen; ++i) {
-                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
-                            }
-                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
-                        }
-                    }
-                }
-            }
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++)
-                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
-            }
-        }
-    }
+    ggml_gemm_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -2126,38 +1887,5 @@ void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
     }
     return;
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
-    {
-        float sumf[4][4];
-        int sumi;
-
-        for (int y = 0; y < nr / 4; y++) {
-            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-            for (int x = 0; x < nc / ncols_interleaved; x++) {
-                const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
-                for (int m = 0; m < 4; m++) {
-                    for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
-                }
-                for (int l = 0; l < nb; l++) {
-                    for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                        for (int m = 0; m < 4; m++) {
-                            for (int j = 0; j < ncols_interleaved; j++) {
-                                sumi = 0;
-                                for (int i = 0; i < blocklen; ++i) {
-                                    const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
-                                    const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
-                                    sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
-                                }
-                                sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
-                            }
-                        }
-                    }
-                }
-                for (int m = 0; m < 4; m++) {
-                    for (int j = 0; j < ncols_interleaved; j++)
-                        s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
-                }
-            }
-        }
-    }
+    ggml_gemm_iq4_nl_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
diff --git a/ggml/src/ggml-cpu/arch/loongarch/quants.c b/ggml/src/ggml-cpu/arch/loongarch/quants.c
index 7908da4d16b..0f9af7bf520 100644
--- a/ggml/src/ggml-cpu/arch/loongarch/quants.c
+++ b/ggml/src/ggml-cpu/arch/loongarch/quants.c
@@ -821,24 +821,15 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = hsum_float_8(acc) + summs;
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[ib].qs[j] & 0x0F);
-            const int v1 = (x[ib].qs[j] >>   4);
-
-            sumi0 += (v0 * y[ib].qs[j]);
-            sumi1 += (v1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q4_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -883,30 +874,15 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = hsum_float_8(acc);
 
-#endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
-            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
-
-            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
-            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(ib);
+    UNUSED(sumf);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -954,30 +930,15 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = hsum_float_8(acc) + summs;
 
-#endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
-            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
-
-            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
-            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(ib);
+    UNUSED(sumf);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -1016,18 +977,15 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = hsum_float_8(acc);
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi = 0;
-
-        for (int j = 0; j < qk; j++) {
-            sumi += x[ib].qs[j]*y[ib].qs[j];
-        }
-
-        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(ib);
+    UNUSED(sumf);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q8_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -1103,45 +1061,10 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = hsum_float_8(acc);
 
 #else
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * q2 = x[i].qs;
-        const  int8_t * q8 = y[i].qs;
-        const uint8_t * sc = x[i].scales;
-
-        int summs = 0;
-        for (int j = 0; j < 16; ++j) {
-            summs += y[i].bsums[j] * (sc[j] >> 4);
-        }
-
-        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
-
-        int isum = 0;
-        int is = 0;
-        int d;
-        for (int k = 0; k < QK_K/128; ++k) {
-            int shift = 0;
-            for (int j = 0; j < 4; ++j) {
-                d = sc[is++] & 0xF;
-                int isuml = 0;
-                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                d = sc[is++] & 0xF;
-                isuml = 0;
-                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                shift += 2;
-                q8 += 32;
-            }
-            q2 += 32;
-        }
-        sumf += dall * isum - dmin * summs;
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1239,70 +1162,13 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = hsum_float_8(acc);
 
 #else
-    // scalar version
-    // This function is written like this so the compiler can manage to vectorize most of it
-    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
-    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
-    // The ideal situation would be if we could just write the code once, and the compiler would
-    // automatically produce the best possible set of machine instructions, instead of us having to manually
-    // write vectorized versions for AVX, ARM_NEON, etc.
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    uint32_t auxs[4];
-    const int8_t * scales = (const int8_t*)auxs;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].hmask;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            q3 += 32;
-        }
-        a = aux8;
-
-        memcpy(auxs, x[i].scales, 12);
-        uint32_t tmp = auxs[2];
-        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
-        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
-        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
-        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
-        for (int j = 0; j < QK_K/16; ++j) {
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-
 }
 
 void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -1391,61 +1257,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = hsum_float_8(acc) + ((v4f32)acc_m)[0];
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            a += 32;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            a += 32; q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1541,66 +1360,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = hsum_float_8(acc) + ((v4f32)acc_m)[0];
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1678,47 +1445,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = hsum_float_8(acc);
 
 #else
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].ql;
-        const uint8_t * GGML_RESTRICT qh = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) {
-                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
-                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
-                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
-                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
-            }
-            a  += 128;
-            q4 += 64;
-            qh += 32;
-        }
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/16; ++j) {
-            int scale = x[i].scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1815,34 +1545,10 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
     *s = 0.125f * hsum_float_8(accumf);
 
 #else
-
-    uint32_t aux32[2];
-    const uint8_t * aux8 = (const uint8_t *)aux32;
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
-        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            memcpy(aux32, q2, 2*sizeof(uint32_t));
-            q2 += 4;
-            const uint32_t ls = 2*(aux32[1] >> 28) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
-                const uint8_t  signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.125f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1978,42 +1684,10 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     *s = 0.125f * hsum_float_8(accumf);
 
 #else
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
-        const uint8_t  * GGML_RESTRICT sc = x[i].scales;
-        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1;
-            const uint16_t ls2 = 2*(sc[ib32] >>  4) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 2; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
-                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls1;
-            sumi = 0;
-            for (int l = 2; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
-                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls2;
-            q2 += 4;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.125f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2105,47 +1779,11 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = 0.125f * hsum_float_8(accumf);
 
 #else
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const int8_t  * q8 = y[i].qs;
-        const uint8_t * qs = x[i].qs;
-        const uint8_t * qh = x[i].qh;
-        const uint8_t * signs = qs + QK_K/8;
-
-        int bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf);
-            int ls2 = 1 + 2*(x[i].scales[ib32] >>  4);
-            int sumi1 = 0, sumi2 = 0;
-            for (int l = 0; l < 2; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
-                for (int j = 0; j < 8; ++j) {
-                    sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            for (int l = 2; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
-                for (int j = 0; j < 8; ++j) {
-                    sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += ls1 * sumi1 + ls2 * sumi2;
-            qs += 4;
-            signs += 4;
-        }
-
-        sumf += d * bsum;
-    }
-
-    *s = 0.125f * sumf;
-
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-
 }
 
 void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -2209,36 +1847,10 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
     *s = 0.25f * hsum_float_8(accumf);
 
 #else
-
-    uint32_t aux32;
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
-        const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
-        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            memcpy(&aux32, gas, sizeof(uint32_t)); gas += sizeof(uint32_t);
-            const uint32_t ls = 2*(aux32 >> 28) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]);
-                const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]);
-                const uint8_t  signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            q3 += 8;
-            bsum += sumi * ls;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.25f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq3_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2338,48 +1950,10 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = hsum_float_8(accumf);
 
 #else
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * GGML_RESTRICT qs = x[i].qs;
-        const uint8_t * GGML_RESTRICT qh = x[i].qh;
-        const uint8_t * GGML_RESTRICT signs = x[i].signs;
-        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const uint32_t ls1 = 2*(x[i].scales[ib32/2] & 0xf) + 1;
-            const uint32_t ls2 = 2*(x[i].scales[ib32/2] >>  4) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256)));
-                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256)));
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            qs += 8;
-            signs += 4;
-            bsum += sumi * ls1;
-            sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256)));
-                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256)));
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            qs += 8;
-            signs += 4;
-            bsum += sumi * ls2;
-        }
-        sumf += d * bsum;
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq3_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2460,36 +2034,10 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
 
 #else
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint16_t * qh = x[i].qh;
-
-        int sumi = 0, sumi1 = 0;
-        for (int ib = 0; ib < QK_K/32; ++ib) {
-            const int ls = 2*((qh[ib] >> 12) & 7) + 1;
-            const int delta = qh[ib] & 0x8000 ? -1 : 1;
-            int lsum = 0;
-            for (int l = 0; l < 4; ++l) {
-                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
-                for (int j = 0; j < 8; ++j) {
-                    lsum += q8[j] * grid[j];
-                }
-                q8 += 8;
-            }
-            sumi  += ls * lsum;
-            sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]);
-            qs += 4;
-        }
-
-        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
-    }
-
-    *s = sumf;
-
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq1_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2603,37 +2151,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     *s = hsum_float_8(accum);
 
 #else
-    float sumf = 0;
-    for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
-        uint16_t h = x[ibl].scales_h;
-        const uint8_t * qs = x[ibl].qs;
-        const int8_t  * q8 = y[ibl].qs;
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
-            const uint8_t ls2 = (x[ibl].scales_l[ib/2] >>  4) | ((h << 2) & 0x30);
-            h >>= 4;
-            const float d1 = d4d8*(ls1 - 32);
-            const float d2 = d4d8*(ls2 - 32);
-            int sumi1 = 0, sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d1 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
-            sumi1 = sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d2 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
-        }
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
diff --git a/ggml/src/ggml-cpu/arch/powerpc/quants.c b/ggml/src/ggml-cpu/arch/powerpc/quants.c
index 053d5cbdc7b..49aae7a23bb 100644
--- a/ggml/src/ggml-cpu/arch/powerpc/quants.c
+++ b/ggml/src/ggml-cpu/arch/powerpc/quants.c
@@ -201,24 +201,14 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = vec_extract(vsumf0, 0);
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[ib].qs[j] & 0x0F) - 8;
-            const int v1 = (x[ib].qs[j] >>   4) - 8;
-
-            sumi0 += (v0 * y[ib].qs[j]);
-            sumi1 += (v1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
-    }
-
     *s = sumf;
+#else
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q4_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -278,24 +268,14 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = vec_extract(vsumf0, 0);
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[ib].qs[j] & 0x0F);
-            const int v1 = (x[ib].qs[j] >>   4);
-
-            sumi0 += (v0 * y[ib].qs[j]);
-            sumi1 += (v1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
-    }
-
     *s = sumf;
+#else
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q4_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -360,30 +340,14 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = vec_extract(vsumf0, 0);
 
-#endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
-            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
-
-            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
-            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
-    }
-
     *s = sumf;
+#else
+    UNUSED(ib);
+    UNUSED(sumf);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -451,30 +415,15 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = vec_extract(vsumf0, 0);
 
-#endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
-            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
-
-            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
-            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(ib);
+    UNUSED(sumf);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -535,18 +484,15 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = vec_extract(vsumf0, 0);
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi = 0;
-
-        for (int j = 0; j < qk; j++) {
-            sumi += x[ib].qs[j]*y[ib].qs[j];
-        }
-
-        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q8_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -695,45 +641,10 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = vec_extract(vsumf0, 0);
 
 #else
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * q2 = x[i].qs;
-        const  int8_t * q8 = y[i].qs;
-        const uint8_t * sc = x[i].scales;
-
-        int summs = 0;
-        for (int j = 0; j < 16; ++j) {
-            summs += y[i].bsums[j] * (sc[j] >> 4);
-        }
-
-        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
-
-        int isum = 0;
-        int is = 0;
-        int d;
-        for (int k = 0; k < QK_K/128; ++k) {
-            int shift = 0;
-            for (int j = 0; j < 4; ++j) {
-                d = sc[is++] & 0xF;
-                int isuml = 0;
-                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                d = sc[is++] & 0xF;
-                isuml = 0;
-                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                shift += 2;
-                q8 += 32;
-            }
-            q2 += 32;
-        }
-        sumf += dall * isum - dmin * summs;
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -907,70 +818,13 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = vec_extract(vsumf0, 0);
 
 #else
-    // scalar version
-    // This function is written like this so the compiler can manage to vectorize most of it
-    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
-    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
-    // The ideal situation would be if we could just write the code once, and the compiler would
-    // automatically produce the best possible set of machine instructions, instead of us having to manually
-    // write vectorized versions for AVX, ARM_NEON, etc.
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    uint32_t auxs[4];
-    const int8_t * scales = (const int8_t*)auxs;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].hmask;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            q3 += 32;
-        }
-        a = aux8;
-
-        memcpy(auxs, x[i].scales, 12);
-        uint32_t tmp = auxs[2];
-        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
-        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
-        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
-        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
-        for (int j = 0; j < QK_K/16; ++j) {
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-
 }
 
 void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -1130,61 +984,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = vec_extract(vsumf0, 0);
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            a += 32;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            a += 32; q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1342,66 +1149,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = vec_extract(vsumf0, 0);
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1556,47 +1311,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = vec_extract(vsumf0, 0);
 
 #else
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].ql;
-        const uint8_t * GGML_RESTRICT qh = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) {
-                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
-                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
-                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
-                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
-            }
-            a  += 128;
-            q4 += 64;
-            qh += 32;
-        }
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/16; ++j) {
-            int scale = x[i].scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1737,34 +1455,10 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
     *s = 0.125f * vec_extract(vsumf0, 0);
 
 #else
-
-    uint32_t aux32[2];
-    const uint8_t * aux8 = (const uint8_t *)aux32;
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
-        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            memcpy(aux32, q2, 2*sizeof(uint32_t));
-            q2 += 4;
-            const uint32_t ls = 2*(aux32[1] >> 28) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
-                const uint8_t  signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.125f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1869,42 +1563,10 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     *s = 0.125f * vec_extract(vsumf0, 0);
 
 #else
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
-        const uint8_t  * GGML_RESTRICT sc = x[i].scales;
-        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1;
-            const uint16_t ls2 = 2*(sc[ib32] >>  4) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 2; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
-                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls1;
-            sumi = 0;
-            for (int l = 2; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
-                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls2;
-            q2 += 4;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.125f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2030,47 +1692,11 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = 0.125f * vec_extract(vsumf0, 0);
 
 #else
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const int8_t  * q8 = y[i].qs;
-        const uint8_t * qs = x[i].qs;
-        const uint8_t * qh = x[i].qh;
-        const uint8_t * signs = qs + QK_K/8;
-
-        int bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf);
-            int ls2 = 1 + 2*(x[i].scales[ib32] >>  4);
-            int sumi1 = 0, sumi2 = 0;
-            for (int l = 0; l < 2; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
-                for (int j = 0; j < 8; ++j) {
-                    sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            for (int l = 2; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
-                for (int j = 0; j < 8; ++j) {
-                    sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += ls1 * sumi1 + ls2 * sumi2;
-            qs += 4;
-            signs += 4;
-        }
-
-        sumf += d * bsum;
-    }
-
-    *s = 0.125f * sumf;
-
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-
 }
 
 void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -2172,36 +1798,10 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
     *s = 0.25f * vec_extract(vsumf0, 0);
 
 #else
-
-    uint32_t aux32;
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
-        const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
-        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            memcpy(&aux32, gas, sizeof(uint32_t)); gas += sizeof(uint32_t);
-            const uint32_t ls = 2*(aux32 >> 28) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]);
-                const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]);
-                const uint8_t  signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            q3 += 8;
-            bsum += sumi * ls;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.25f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq3_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2327,48 +1927,10 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = vec_extract(vsumf0, 0);
 
 #else
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * GGML_RESTRICT qs = x[i].qs;
-        const uint8_t * GGML_RESTRICT qh = x[i].qh;
-        const uint8_t * GGML_RESTRICT signs = x[i].signs;
-        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const uint32_t ls1 = 2*(x[i].scales[ib32/2] & 0xf) + 1;
-            const uint32_t ls2 = 2*(x[i].scales[ib32/2] >>  4) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256)));
-                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256)));
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            qs += 8;
-            signs += 4;
-            bsum += sumi * ls1;
-            sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256)));
-                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256)));
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            qs += 8;
-            signs += 4;
-            bsum += sumi * ls2;
-        }
-        sumf += d * bsum;
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq3_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2481,36 +2043,10 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = vec_extract(vsumf0, 0);
 
 #else
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint16_t * qh = x[i].qh;
-
-        int sumi = 0, sumi1 = 0;
-        for (int ib = 0; ib < QK_K/32; ++ib) {
-            const int ls = 2*((qh[ib] >> 12) & 7) + 1;
-            const int delta = qh[ib] & 0x8000 ? -1 : 1;
-            int lsum = 0;
-            for (int l = 0; l < 4; ++l) {
-                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
-                for (int j = 0; j < 8; ++j) {
-                    lsum += q8[j] * grid[j];
-                }
-                q8 += 8;
-            }
-            sumi  += ls * lsum;
-            sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]);
-            qs += 4;
-        }
-
-        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
-    }
-
-    *s = sumf;
-
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq1_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2581,17 +2117,15 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
 
     sumf = vec_extract(vsumf0, 0);
 
-#endif
-    for (; ib < nb; ++ib) {
-        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
-        int sumi1 = 0, sumi2 = 0;
-        for (int j = 0; j < QK4_NL/2; ++j) {
-            sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
-            sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >>  4];
-        }
-        sumf += d * (sumi1 + sumi2);
-    }
     *s = sumf;
+#else
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_iq4_nl_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -2696,37 +2230,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     *s = vec_extract(vsumf0, 0);
 
 #else
-    float sumf = 0;
-    for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
-        uint16_t h = x[ibl].scales_h;
-        const uint8_t * qs = x[ibl].qs;
-        const int8_t  * q8 = y[ibl].qs;
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
-            const uint8_t ls2 = (x[ibl].scales_l[ib/2] >>  4) | ((h << 2) & 0x30);
-            h >>= 4;
-            const float d1 = d4d8*(ls1 - 32);
-            const float d2 = d4d8*(ls2 - 32);
-            int sumi1 = 0, sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d1 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
-            sumi1 = sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d2 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
-        }
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
diff --git a/ggml/src/ggml-cpu/arch/riscv/quants.c b/ggml/src/ggml-cpu/arch/riscv/quants.c
index 8b64d8adc48..6c74417c90c 100644
--- a/ggml/src/ggml-cpu/arch/riscv/quants.c
+++ b/ggml/src/ggml-cpu/arch/riscv/quants.c
@@ -116,6 +116,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
 //===================================== Dot products =================================
 
 void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+#if defined(__riscv_v)
     const int qk = QK8_0;
     const int nb = n / qk;
 
@@ -132,7 +133,6 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     int ib = 0;
     float sumf = 0;
 
-#if defined(__riscv_v)
     size_t vl = qk / 2;
 
     for (; ib < nb; ++ib) {
@@ -164,27 +164,14 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
     }
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[ib].qs[j] & 0x0F) - 8;
-            const int v1 = (x[ib].qs[j] >>   4) - 8;
-
-            sumi0 += (v0 * y[ib].qs[j]);
-            sumi1 += (v1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
-    }
-
     *s = sumf;
+#else
+    ggml_vec_dot_q4_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+#if defined(__riscv_v)
     const int qk = QK8_1;
     const int nb = n / qk;
 
@@ -201,7 +188,6 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
     int ib = 0;
     float sumf = 0;
 
-#if defined(__riscv_v)
     size_t vl = qk / 2;
 
     for (; ib < nb; ++ib) {
@@ -229,27 +215,14 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[ib].qs[j] & 0x0F);
-            const int v1 = (x[ib].qs[j] >>   4);
-
-            sumi0 += (v0 * y[ib].qs[j]);
-            sumi1 += (v1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
-    }
-
     *s = sumf;
+#else
+    ggml_vec_dot_q4_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+#if defined(__riscv_v)
     const int qk = QK8_0;
     const int nb = n / qk;
 
@@ -267,7 +240,6 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     const block_q5_0 * GGML_RESTRICT x = vx;
     const block_q8_0 * GGML_RESTRICT y = vy;
 
-#if defined(__riscv_v)
     size_t vl;
     size_t vlenb = __riscv_vlenb();
 
@@ -297,33 +269,14 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
     }
 
-#endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
-            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
-
-            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
-            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
-    }
-
     *s = sumf;
+#else
+    ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+#if defined(__riscv_v)
     const int qk = QK8_1;
     const int nb = n / qk;
 
@@ -341,7 +294,6 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
     const block_q5_1 * GGML_RESTRICT x = vx;
     const block_q8_1 * GGML_RESTRICT y = vy;
 
-#if defined(__riscv_v)
     size_t vl;
     size_t vlenb = __riscv_vlenb();
 
@@ -370,30 +322,10 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
-#endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
-            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
-
-            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
-            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
-    }
-
     *s = sumf;
+#else
+    ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -431,18 +363,17 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
     }
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi = 0;
+    *s = sumf;
+#else
 
-        for (int j = 0; j < qk; j++) {
-            sumi += x[ib].qs[j]*y[ib].qs[j];
-        }
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
 
-        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
-    }
-
-    *s = sumf;
+    ggml_vec_dot_q8_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -738,44 +669,11 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
 #else
 
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * q2 = x[i].qs;
-        const  int8_t * q8 = y[i].qs;
-        const uint8_t * sc = x[i].scales;
-
-        int summs = 0;
-        for (int j = 0; j < 16; ++j) {
-            summs += y[i].bsums[j] * (sc[j] >> 4);
-        }
-
-        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
 
-        int isum = 0;
-        int is = 0;
-        int d;
-        for (int k = 0; k < QK_K/128; ++k) {
-            int shift = 0;
-            for (int j = 0; j < 4; ++j) {
-                d = sc[is++] & 0xF;
-                int isuml = 0;
-                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                d = sc[is++] & 0xF;
-                isuml = 0;
-                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                shift += 2;
-                q8 += 32;
-            }
-            q2 += 32;
-        }
-        sumf += dall * isum - dmin * summs;
-    }
-    *s = sumf;
+    ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1147,68 +1045,14 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sumf;
 
 #else
-    // scalar version
-    // This function is written like this so the compiler can manage to vectorize most of it
-    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
-    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
-    // The ideal situation would be if we could just write the code once, and the compiler would
-    // automatically produce the best possible set of machine instructions, instead of us having to manually
-    // write vectorized versions for AVX, ARM_NEON, etc.
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    uint32_t auxs[4];
-    const int8_t * scales = (const int8_t*)auxs;
 
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].hmask;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            q3 += 32;
-        }
-        a = aux8;
-
-        memcpy(auxs, x[i].scales, 12);
-        uint32_t tmp = auxs[2];
-        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
-        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
-        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
-        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
-        for (int j = 0; j < QK_K/16; ++j) {
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
 
+    ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 
 }
@@ -1534,60 +1378,15 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
 #else
 
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(nb);
+    UNUSED(utmp);
 
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            a += 32;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            a += 32; q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1698,65 +1497,15 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
 #else
 
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(nb);
+    UNUSED(utmp);
 
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2024,46 +1773,11 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
 #else
 
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
 
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].ql;
-        const uint8_t * GGML_RESTRICT qh = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) {
-                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
-                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
-                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
-                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
-            }
-            a  += 128;
-            q4 += 64;
-            qh += 32;
-        }
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/16; ++j) {
-            int scale = x[i].scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
diff --git a/ggml/src/ggml-cpu/arch/riscv/repack.cpp b/ggml/src/ggml-cpu/arch/riscv/repack.cpp
index 45c91a69482..2a35ff9ad87 100644
--- a/ggml/src/ggml-cpu/arch/riscv/repack.cpp
+++ b/ggml/src/ggml-cpu/arch/riscv/repack.cpp
@@ -112,31 +112,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     }
 
 #endif
-    {
-        float sumf[8];
-        int sumi;
-
-        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
-
-            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int j = 0; j < ncols_interleaved; j++) {
-                        sumi = 0;
-                        for (int i = 0; i < blocklen; ++i) {
-                            const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
-                        }
-                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
-                    }
-                }
-            }
-            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
-        }
-    }
+    ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -361,37 +337,6 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
         return;
     }
 
-#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
-    float sumf[4][8];
-    int sumi;
-
-    for (int y = 0; y < nr / 4; y++) {
-        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
-            }
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int m = 0; m < 4; m++) {
-                        for (int j = 0; j < ncols_interleaved; j++) {
-                            sumi = 0;
-                            for (int i = 0; i < blocklen; ++i) {
-                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
-                            }
-                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
-                        }
-                    }
-                }
-            }
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++)
-                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
-            }
-        }
-    }
+#endif
+    ggml_gemm_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
diff --git a/ggml/src/ggml-cpu/arch/s390/quants.c b/ggml/src/ggml-cpu/arch/s390/quants.c
index a840219a4fc..7e4229d0e46 100644
--- a/ggml/src/ggml-cpu/arch/s390/quants.c
+++ b/ggml/src/ggml-cpu/arch/s390/quants.c
@@ -172,24 +172,15 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = acc[0] + acc[1] + acc[2] + acc[3];
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[ib].qs[j] & 0x0F) - 8;
-            const int v1 = (x[ib].qs[j] >>   4) - 8;
-
-            sumi0 += (v0 * y[ib].qs[j]);
-            sumi1 += (v1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q4_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -239,24 +230,15 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = acc[0] + acc[1] + acc[2] + acc[3] + summs;
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[ib].qs[j] & 0x0F);
-            const int v1 = (x[ib].qs[j] >>   4);
-
-            sumi0 += (v0 * y[ib].qs[j]);
-            sumi1 += (v1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q4_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -298,18 +280,15 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     sumf = acc[0] + acc[1] + acc[2] + acc[3];
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi = 0;
-
-        for (int j = 0; j < qk; j++) {
-            sumi += x[ib].qs[j]*y[ib].qs[j];
-        }
-
-        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q8_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -442,70 +421,13 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sum;
 
 #else
-    // scalar version
-    // This function is written like this so the compiler can manage to vectorize most of it
-    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
-    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
-    // The ideal situation would be if we could just write the code once, and the compiler would
-    // automatically produce the best possible set of machine instructions, instead of us having to manually
-    // write vectorized versions for AVX, ARM_NEON, etc.
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    uint32_t auxs[4];
-    const int8_t * scales = (const int8_t*)auxs;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].hmask;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            q3 += 32;
-        }
-        a = aux8;
-
-        memcpy(auxs, x[i].scales, 12);
-        uint32_t tmp = auxs[2];
-        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
-        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
-        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
-        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
-        for (int j = 0; j < QK_K/16; ++j) {
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-
 }
 
 void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -600,61 +522,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sumf;
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            a += 32;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            a += 32; q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -767,66 +642,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sumf;
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -969,47 +792,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sum;
 
 #else
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].ql;
-        const uint8_t * GGML_RESTRICT qh = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) {
-                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
-                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
-                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
-                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
-            }
-            a  += 128;
-            q4 += 64;
-            qh += 32;
-        }
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/16; ++j) {
-            int scale = x[i].scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1186,17 +972,15 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
         sumf += GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d) * (v_xy[0] + v_xy[1] + v_xy[2] + v_xy[3]);
     }
 
-#endif
-    for (; ib < nb; ++ib) {
-        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
-        int sumi1 = 0, sumi2 = 0;
-        for (int j = 0; j < QK4_NL/2; ++j) {
-            sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
-            sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >>  4];
-        }
-        sumf += d * (sumi1 + sumi2);
-    }
     *s = sumf;
+#else
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_iq4_nl_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -1264,37 +1048,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     *s = sumf;
 
 #else
-    float sumf = 0;
-    for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
-        uint16_t h = x[ibl].scales_h;
-        const uint8_t * qs = x[ibl].qs;
-        const int8_t  * q8 = y[ibl].qs;
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
-            const uint8_t ls2 = (x[ibl].scales_l[ib/2] >>  4) | ((h << 2) & 0x30);
-            h >>= 4;
-            const float d1 = d4d8*(ls1 - 32);
-            const float d2 = d4d8*(ls2 - 32);
-            int sumi1 = 0, sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d1 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
-            sumi1 = sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d2 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
-        }
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
diff --git a/ggml/src/ggml-cpu/arch/wasm/quants.c b/ggml/src/ggml-cpu/arch/wasm/quants.c
index b0904d8a3ab..74a359e6d12 100644
--- a/ggml/src/ggml-cpu/arch/wasm/quants.c
+++ b/ggml/src/ggml-cpu/arch/wasm/quants.c
@@ -435,30 +435,15 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
            wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
 
-#endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
-            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
-
-            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
-            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(ib);
+    UNUSED(sumf);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -545,30 +530,15 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
     sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
            wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3) + summs;
 
-#endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
-            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
-
-            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
-            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(ib);
+    UNUSED(sumf);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -628,18 +598,15 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
            wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
 
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi = 0;
-
-        for (int j = 0; j < qk; j++) {
-            sumi += x[ib].qs[j]*y[ib].qs[j];
-        }
-
-        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
-    }
-
     *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q8_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
 }
 
 void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -755,45 +722,10 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sumf;
 
 #else
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * q2 = x[i].qs;
-        const  int8_t * q8 = y[i].qs;
-        const uint8_t * sc = x[i].scales;
-
-        int summs = 0;
-        for (int j = 0; j < 16; ++j) {
-            summs += y[i].bsums[j] * (sc[j] >> 4);
-        }
-
-        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
-
-        int isum = 0;
-        int is = 0;
-        int d;
-        for (int k = 0; k < QK_K/128; ++k) {
-            int shift = 0;
-            for (int j = 0; j < 4; ++j) {
-                d = sc[is++] & 0xF;
-                int isuml = 0;
-                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                d = sc[is++] & 0xF;
-                isuml = 0;
-                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                shift += 2;
-                q8 += 32;
-            }
-            q2 += 32;
-        }
-        sumf += dall * isum - dmin * summs;
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -902,68 +834,12 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sumf;
 
 #else
-    // scalar version
-    // This function is written like this so the compiler can manage to vectorize most of it
-    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
-    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
-    // The ideal situation would be if we could just write the code once, and the compiler would
-    // automatically produce the best possible set of machine instructions, instead of us having to manually
-    // write vectorized versions for AVX, ARM_NEON, etc.
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    uint32_t auxs[4];
-    const int8_t * scales = (const int8_t*)auxs;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].hmask;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            q3 += 32;
-        }
-        a = aux8;
-
-        memcpy(auxs, x[i].scales, 12);
-        uint32_t tmp = auxs[2];
-        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
-        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
-        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
-        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
-        for (int j = 0; j < QK_K/16; ++j) {
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 
 }
@@ -1089,61 +965,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sumf;
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            a += 32;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            a += 32; q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1279,66 +1108,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sumf;
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1435,47 +1212,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sumf;
 
 #else
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].ql;
-        const uint8_t * GGML_RESTRICT qh = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) {
-                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
-                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
-                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
-                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
-            }
-            a  += 128;
-            q4 += 64;
-            qh += 32;
-        }
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/16; ++j) {
-            int scale = x[i].scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
diff --git a/ggml/src/ggml-cpu/arch/x86/quants.c b/ggml/src/ggml-cpu/arch/x86/quants.c
index e7527c00a8f..30fd59f7028 100644
--- a/ggml/src/ggml-cpu/arch/x86/quants.c
+++ b/ggml/src/ggml-cpu/arch/x86/quants.c
@@ -702,7 +702,6 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
     const block_q8_1 * GGML_RESTRICT y = vy;
 
     int ib = 0;
-    float sumf = 0;
 
 #if defined(__AVX2__) || defined(__AVX__)
     // Initialize accumulator with zeros
@@ -737,26 +736,14 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 #endif
     }
 
-    sumf = hsum_float_8(acc) + summs;
-
+    *s = hsum_float_8(acc) + summs;
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    ggml_vec_dot_q4_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-    for (; ib < nb; ++ib) {
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[ib].qs[j] & 0x0F);
-            const int v1 = (x[ib].qs[j] >>   4);
-
-            sumi0 += (v0 * y[ib].qs[j]);
-            sumi1 += (v1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
-    }
-
-    *s = sumf;
 }
 
 void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -764,7 +751,6 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     const int nb = n / qk;
 
     int ib = 0;
-    float sumf = 0;
 
     assert(n % qk == 0);
     assert(qk == QK5_0);
@@ -799,7 +785,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         acc = _mm256_fmadd_ps(d, q, acc);
     }
 
-    sumf = hsum_float_8(acc);
+    *s = hsum_float_8(acc);
 #elif defined(__AVX__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
@@ -830,32 +816,14 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         acc = _mm256_add_ps(_mm256_mul_ps(d, q), acc);
     }
 
-    sumf = hsum_float_8(acc);
-
+    *s = hsum_float_8(acc);
+#else
+    UNUSED(nb);
+    UNUSED(ib);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
-            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
-
-            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
-            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
-    }
-
-    *s = sumf;
 }
 
 void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -863,7 +831,6 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
     const int nb = n / qk;
 
     int ib = 0;
-    float sumf = 0;
 
     assert(n % qk == 0);
     assert(qk == QK5_1);
@@ -901,7 +868,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         acc = _mm256_fmadd_ps(q, _mm256_mul_ps(dx, dy), acc);
     }
 
-    sumf = hsum_float_8(acc) + summs;
+    *s = hsum_float_8(acc) + summs;
 #elif defined(__AVX__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
@@ -935,32 +902,14 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         acc = _mm256_add_ps(_mm256_mul_ps(q, _mm256_mul_ps(dx, dy)), acc);
     }
 
-    sumf = hsum_float_8(acc) + summs;
-
+    *s = hsum_float_8(acc) + summs;
+#else
+    UNUSED(nb);
+    UNUSED(ib);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
-            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
-
-            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
-            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
-    }
-
-    *s = sumf;
 }
 
 void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -1017,7 +966,6 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     }
 
     sumf = hsum_float_8(accum);
-
 #endif
     for (; ib < nb; ++ib) {
         int sumi = 0;
@@ -1157,44 +1105,10 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = hsum_float_8(sumf);
 
 #else
-    const uint8_t pow3[6] = {1, 3, 9, 27, 81, 243};
-
-    float sumf = 0.0f;
-
-    for (int i = 0; i < nb; ++i) {
-        int sum = 0;
-
-        for (size_t j = 0; j < sizeof(x->qs) - sizeof(x->qs) % 32; j += 32) {
-            for (size_t l = 0; l < 5; ++l) {
-                for (size_t m = 0; m < 32; ++m) {
-                    uint8_t q = x[i].qs[j + m] * pow3[l];
-                    uint16_t xi = ((uint16_t) q * 3) >> 8;
-                    sum += (xi - 1) * y[i].qs[j*5 + l*32 + m];
-                }
-            }
-        }
-        for (size_t j = sizeof(x->qs) - sizeof(x->qs) % 32; j < sizeof(x->qs); j += 16) {
-            for (size_t l = 0; l < 5; ++l) {
-                for (size_t m = 0; m < 16; ++m) {
-                    uint8_t q = x[i].qs[j + m] * pow3[l];
-                    uint16_t xi = ((uint16_t) q * 3) >> 8;
-                    sum += (xi - 1) * y[i].qs[j*5 + l*16 + m];
-                }
-            }
-        }
-
-        for (size_t l = 0; l < 4; ++l) {
-            for (size_t j = 0; j < sizeof(x->qh); ++j) {
-                uint8_t q = x[i].qh[j] * pow3[l];
-                uint16_t xi = ((uint16_t) q * 3) >> 8;
-                sum += (xi - 1) * y[i].qs[sizeof(x->qs)*5 + l*sizeof(x->qh) + j];
-            }
-        }
-
-        sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d);
-    }
-
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_tq1_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1257,25 +1171,10 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = hsum_float_8(sumf);
 
 #else
-    float sumf = 0.0f;
-
-    for (int i = 0; i < nb; ++i) {
-        int32_t sumi = 0;
-
-        for (size_t j = 0; j < sizeof(x->qs); j += 32) {
-            for (size_t l = 0; l < 4; ++l) {
-                for (size_t k = 0; k < 32; ++k) {
-                    sumi += y[i].qs[j*4 + l*32 + k] * (((x[i].qs[j + k] >> (l*2)) & 3) - 1);
-                }
-            }
-        }
-
-        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
-
-        sumf += (float) sumi * d;
-    }
-
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_tq2_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1464,45 +1363,10 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = hsum_float_8(acc);
 
 #else
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * q2 = x[i].qs;
-        const  int8_t * q8 = y[i].qs;
-        const uint8_t * sc = x[i].scales;
-
-        int summs = 0;
-        for (int j = 0; j < 16; ++j) {
-            summs += y[i].bsums[j] * (sc[j] >> 4);
-        }
-
-        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
-
-        int isum = 0;
-        int is = 0;
-        int d;
-        for (int k = 0; k < QK_K/128; ++k) {
-            int shift = 0;
-            for (int j = 0; j < 4; ++j) {
-                d = sc[is++] & 0xF;
-                int isuml = 0;
-                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                d = sc[is++] & 0xF;
-                isuml = 0;
-                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                shift += 2;
-                q8 += 32;
-            }
-            q2 += 32;
-        }
-        sumf += dall * isum - dmin * summs;
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -1769,70 +1633,13 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = hsum_float_8(acc);
 
 #else
-    // scalar version
-    // This function is written like this so the compiler can manage to vectorize most of it
-    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
-    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
-    // The ideal situation would be if we could just write the code once, and the compiler would
-    // automatically produce the best possible set of machine instructions, instead of us having to manually
-    // write vectorized versions for AVX, ARM_NEON, etc.
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    uint32_t auxs[4];
-    const int8_t * scales = (const int8_t*)auxs;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].hmask;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            q3 += 32;
-        }
-        a = aux8;
-
-        memcpy(auxs, x[i].scales, 12);
-        uint32_t tmp = auxs[2];
-        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
-        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
-        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
-        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
-        for (int j = 0; j < QK_K/16; ++j) {
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-
 }
 
 void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -2002,61 +1809,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m);
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            a += 32;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            a += 32; q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2259,66 +2019,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = hsum_float_8(acc) + summs;
 
 #else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const uint8_t * GGML_RESTRICT hm = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    UNUSED(utmp);
+    ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2520,47 +2228,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = hsum_float_8(acc);
 
 #else
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * GGML_RESTRICT q4 = x[i].ql;
-        const uint8_t * GGML_RESTRICT qh = x[i].qh;
-        const  int8_t * GGML_RESTRICT q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * GGML_RESTRICT a = aux8;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) {
-                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
-                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
-                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
-                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
-            }
-            a  += 128;
-            q4 += 64;
-            qh += 32;
-        }
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/16; ++j) {
-            int scale = x[i].scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -2712,34 +2383,10 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
     *s = 0.125f * hsum_float_8(accumf);
 
 #else
-
-    uint32_t aux32[2];
-    const uint8_t * aux8 = (const uint8_t *)aux32;
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
-        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            memcpy(aux32, q2, 2*sizeof(uint32_t));
-            q2 += 4;
-            const uint32_t ls = 2*(aux32[1] >> 28) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
-                const uint8_t  signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.125f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -3033,42 +2680,10 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     *s = 0.125f * hsum_float_8(accumf);
 
 #else
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
-        const uint8_t  * GGML_RESTRICT sc = x[i].scales;
-        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1;
-            const uint16_t ls2 = 2*(sc[ib32] >>  4) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 2; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
-                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls1;
-            sumi = 0;
-            for (int l = 2; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
-                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls2;
-            q2 += 4;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.125f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -3250,47 +2865,11 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = 0.125f * hsum_float_8(accumf);
 
 #else
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const int8_t  * q8 = y[i].qs;
-        const uint8_t * qs = x[i].qs;
-        const uint8_t * qh = x[i].qh;
-        const uint8_t * signs = qs + QK_K/8;
-
-        int bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf);
-            int ls2 = 1 + 2*(x[i].scales[ib32] >>  4);
-            int sumi1 = 0, sumi2 = 0;
-            for (int l = 0; l < 2; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
-                for (int j = 0; j < 8; ++j) {
-                    sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            for (int l = 2; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
-                for (int j = 0; j < 8; ++j) {
-                    sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += ls1 * sumi1 + ls2 * sumi2;
-            qs += 4;
-            signs += 4;
-        }
-
-        sumf += d * bsum;
-    }
-
-    *s = 0.125f * sumf;
-
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq2_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-
 }
 
 void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -3410,36 +2989,10 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
     *s = 0.25f * hsum_float_8(accumf);
 
 #else
-
-    uint32_t aux32;
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
-        const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
-        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            memcpy(&aux32, gas, sizeof(uint32_t)); gas += sizeof(uint32_t);
-            const uint32_t ls = 2*(aux32 >> 28) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]);
-                const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]);
-                const uint8_t  signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            q3 += 8;
-            bsum += sumi * ls;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.25f * sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq3_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -3646,48 +3199,10 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = hsum_float_8(accumf);
 
 #else
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * GGML_RESTRICT qs = x[i].qs;
-        const uint8_t * GGML_RESTRICT qh = x[i].qh;
-        const uint8_t * GGML_RESTRICT signs = x[i].signs;
-        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const uint32_t ls1 = 2*(x[i].scales[ib32/2] & 0xf) + 1;
-            const uint32_t ls2 = 2*(x[i].scales[ib32/2] >>  4) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256)));
-                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256)));
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            qs += 8;
-            signs += 4;
-            bsum += sumi * ls1;
-            sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256)));
-                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256)));
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            qs += 8;
-            signs += 4;
-            bsum += sumi * ls2;
-        }
-        sumf += d * bsum;
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq3_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -3811,36 +3326,10 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
 
 #else
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint16_t * qh = x[i].qh;
-
-        int sumi = 0, sumi1 = 0;
-        for (int ib = 0; ib < QK_K/32; ++ib) {
-            const int ls = 2*((qh[ib] >> 12) & 7) + 1;
-            const int delta = qh[ib] & 0x8000 ? -1 : 1;
-            int lsum = 0;
-            for (int l = 0; l < 4; ++l) {
-                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
-                for (int j = 0; j < 8; ++j) {
-                    lsum += q8[j] * grid[j];
-                }
-                q8 += 8;
-            }
-            sumi  += ls * lsum;
-            sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]);
-            qs += 4;
-        }
-
-        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
-    }
-
-    *s = sumf;
-
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq1_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -4043,52 +3532,11 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = hsum_float_8(accum1) + IQ1M_DELTA * hsum_float_8(accum2);
 
 #else
-
-    int sum1[2], sum2[2], delta[4];
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint8_t  * qh = x[i].qh;
-        const uint16_t * sc = (const uint16_t *)x[i].scales;
-
-        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
-
-        int sumi1 = 0, sumi2 = 0;
-        for (int ib = 0; ib < QK_K/32; ++ib) {
-            delta[0] = qh[0] & 0x08 ? -1 : 1;
-            delta[1] = qh[0] & 0x80 ? -1 : 1;
-            delta[2] = qh[1] & 0x08 ? -1 : 1;
-            delta[3] = qh[1] & 0x80 ? -1 : 1;
-            sum1[0] = sum1[1] = sum2[0] = sum2[1] = 0;
-            for (int l = 0; l < 4; ++l) {
-                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((uint16_t)qh[l/2] << (8 - 4*(l%2))) & 0x700)));
-                int lsum1 = 0, lsum2 = 0;
-                for (int j = 0; j < 8; ++j) {
-                    lsum1 += q8[j] * grid[j];
-                    lsum2 += q8[j];
-                }
-                q8 += 8;
-                sum1[l/2] += lsum1;
-                sum2[l/2] += lsum2*delta[l];
-            }
-
-            const int ls1 = 2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1;
-            const int ls2 = 2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1;
-
-            sumi1 += sum1[0] * ls1 + sum1[1] * ls2;
-            sumi2 += sum2[0] * ls1 + sum2[1] * ls2;
-            qs += 4;
-            qh += 2;
-        }
-
-        sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
-    }
-
-    *s = sumf;
-
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    UNUSED(scale);
+    ggml_vec_dot_iq1_m_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
@@ -4275,37 +3723,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     *s = hsum_float_8(accum);
 
 #else
-    float sumf = 0;
-    for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
-        uint16_t h = x[ibl].scales_h;
-        const uint8_t * qs = x[ibl].qs;
-        const int8_t  * q8 = y[ibl].qs;
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
-            const uint8_t ls2 = (x[ibl].scales_l[ib/2] >>  4) | ((h << 2) & 0x30);
-            h >>= 4;
-            const float d1 = d4d8*(ls1 - 32);
-            const float d2 = d4d8*(ls2 - 32);
-            int sumi1 = 0, sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d1 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
-            sumi1 = sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d2 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
-        }
-    }
-    *s = sumf;
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(nb);
+    ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
diff --git a/ggml/src/ggml-cpu/arch/x86/repack.cpp b/ggml/src/ggml-cpu/arch/x86/repack.cpp
index c00c1e541cb..1982cfef995 100644
--- a/ggml/src/ggml-cpu/arch/x86/repack.cpp
+++ b/ggml/src/ggml-cpu/arch/x86/repack.cpp
@@ -281,35 +281,9 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
     }
 
 #else
-    // scalar
-    const int blck_size_interleave = 8;
-    float srcv[4][QK8_0];
-    float id[4];
-
-    for (int i = 0; i < nb; i++) {
-        for (int row_iter = 0; row_iter < 4; row_iter++) {
-            float amax = 0.0f; // absolute max
-
-            for (int j = 0; j < QK8_0; j++) {
-                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
-                amax = MAX(amax, fabsf(srcv[row_iter][j]));
-            }
-
-            const float d = amax / ((1 << 7) - 1);
-            id[row_iter] = d ? 1.0f / d : 0.0f;
-
-            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
-        }
-
-        for (int j = 0; j < QK8_0 * 4; j++) {
-            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
-            int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
-            src_offset += (j % blck_size_interleave);
-
-            float x0 = srcv[src_id][src_offset] * id[src_id];
-            y[i].qs[j] = roundf(x0);
-        }
-    }
+    UNUSED(nb);
+    UNUSED(y);
+    ggml_quantize_mat_q8_0_4x8_generic(x, vy, k);
 #endif
 }
 
@@ -531,49 +505,9 @@ void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
     }
 
 #else
-
-    // scalar
-    const int blck_size_interleave = 8;
-    float srcv[4][QK_K];
-    float iscale[4];
-
-    for (int i = 0; i < nb; i++) {
-        for (int row_iter = 0; row_iter < 4; row_iter++) {
-            float amax = 0.0f; // absolute max
-            float max = 0;
-
-            for (int j = 0; j < QK_K; j++) {
-                srcv[row_iter][j] = x[row_iter * k + i * QK_K + j];
-                // Update the maximum value of the corresponding super block
-                if(amax < fabsf(srcv[row_iter][j])) {
-                    amax = fabsf(srcv[row_iter][j]);
-                    max = srcv[row_iter][j];
-                }
-            }
-
-            iscale[row_iter] = amax ? -127.f/max : 0;
-
-            y[i].d[row_iter] = amax ? 1/iscale[row_iter] : 0;
-        }
-
-        for (int j = 0; j < QK_K / 4; j++) {
-            y[i].bsums[j] = 0;
-        }
-
-        // Quants values are interleaved in sequence of eight bytes from corresponding super blocks
-        // Bsums values are interleaved in sequence of four bsums from each super block taken for interleaving
-        // i.e first four bsums from the first super block, followed by first four bsums from second super block and so on
-        for (int j = 0; j < QK_K * 4; j++) {
-            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
-            int src_id     = (j % (4 * blck_size_interleave)) / blck_size_interleave;
-            src_offset += (j % blck_size_interleave);
-            int index = (((j & 31) >> 3) << 2) + ((j >> 8) << 4) + ((j >> 6) & 3);
-
-            float x0 = srcv[src_id][src_offset] * iscale[src_id];
-            y[i].qs[j] = nearest_int(x0);
-            y[i].bsums[index] += y[i].qs[j];
-        }
-    }
+    UNUSED(nb);
+    UNUSED(y);
+    ggml_quantize_mat_q8_K_4x8_generic(x, vy, k);
 #endif
 }
 
@@ -689,31 +623,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     return;
 
 #endif
-    {
-        float sumf[8];
-        int sumi;
-
-        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
-
-            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int j = 0; j < ncols_interleaved; j++) {
-                        sumi = 0;
-                        for (int i = 0; i < blocklen; ++i) {
-                            const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
-                        }
-                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
-                    }
-                }
-            }
-            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
-        }
-    }
+    ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -932,61 +842,10 @@ void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     }
 
 #else
-
-    float sumf[8];
-    float sum_minf[8];
-    uint32_t utmp[32];
-    int sumi1;
-    int sumi2;
-    int sumi;
-
-    const block_q8_K * a_ptr = (const block_q8_K *) vy;
-    for (int x = 0; x < nc / ncols_interleaved; x++) {
-        const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb);
-
-        for (int j = 0; j < ncols_interleaved; j++) {
-            sumf[j] = 0.0;
-            sum_minf[j] = 0.0;
-        }
-        for (int l = 0; l < nb; l++) {
-            for (int sb = 0; sb < 8; sb++) {
-                memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
-                utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
-                const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
-                utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
-                utmp[sb * 4 + 2] = uaux_0;
-                utmp[sb * 4 + 0] &= kmask1;
-            }
-            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                uint8_t *scales_0 = (uint8_t*) utmp + (k / 4) * 32;
-                uint8_t *scales_1 = (uint8_t*) utmp + (k / 4) * 32 + 16;
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    sumi1 = 0;
-                    sumi2 = 0;
-                    sumi = 0;
-                    for (int i = 0; i < blocklen; ++i) {
-                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF);
-                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4);
-                        sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 64 + (k % 4) * blocklen + i]);
-                        sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 64 + (k % 4) * blocklen + i + 32]);
-                        sumi1 = sumi1 * scales_0[j];
-                        sumi2 = sumi2 * scales_1[j];
-                        sumi += sumi1 + sumi2;
-                    }
-                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
-                }
-            }
-            for (int sb = 0; sb < 8; sb++) {
-                uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16;
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
-                }
-            }
-        }
-        for (int j = 0; j < ncols_interleaved; j++) {
-            s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j];
-        }
-    }
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    ggml_gemv_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
 #endif
 }
 
@@ -1735,38 +1594,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     }
 
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
-    float sumf[4][8];
-    int sumi;
-
-    for (int y = 0; y < nr / 4; y++) {
-        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
-            }
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int m = 0; m < 4; m++) {
-                        for (int j = 0; j < ncols_interleaved; j++) {
-                            sumi = 0;
-                            for (int i = 0; i < blocklen; ++i) {
-                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
-                            }
-                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
-                        }
-                    }
-                }
-            }
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++)
-                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
-            }
-        }
-    }
+    ggml_gemm_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -3216,70 +3044,9 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     }
 
 #else
-
-    float sumf[4][8];
-    float sum_minf[4][8];
-    uint32_t utmp[32];
-    int sumi1;
-    int sumi2;
-    int sumi;
-
-    for (int y = 0; y < nr / 4; y++) {
-        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb);
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    sumf[m][j] = 0.0;
-                    sum_minf[m][j] = 0.0;
-                }
-            }
-            for (int l = 0; l < nb; l++) {
-                for (int sb = 0; sb < 8; sb++) {
-                    memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
-                    utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
-                    const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
-                    utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
-                    utmp[sb * 4 + 2] = uaux_0;
-                    utmp[sb * 4 + 0] &= kmask1;
-                }
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    uint8_t *scales_0 = (uint8_t*) utmp + (k / 4) * 32;
-                    uint8_t *scales_1 = (uint8_t*) utmp + (k / 4) * 32 + 16;
-                    for (int m = 0; m < 4; m++) {
-                        for (int j = 0; j < ncols_interleaved; j++) {
-                            sumi1 = 0;
-                            sumi2 = 0;
-                            sumi = 0;
-                            for (int i = 0; i < blocklen; ++i) {
-                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF);
-                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4);
-                                sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i]);
-                                sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i + 128]);
-                                sumi1 = sumi1 * scales_0[j];
-                                sumi2 = sumi2 * scales_1[j];
-                                sumi += sumi1 + sumi2;
-                            }
-                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
-                        }
-                    }
-                }
-                for (int sb = 0; sb < 8; sb++) {
-                    uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16;
-                    for(int m = 0; m < 4; m++) {
-                        const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
-                        for(int j = 0; j < ncols_interleaved; j++) {
-                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
-                        }
-                    }
-                }
-            }
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
-                }
-            }
-        }
-    }
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    ggml_gemm_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
 #endif
 }

From 60dbdabb8fce17711acfad3b5d3c8f6055ff8569 Mon Sep 17 00:00:00 2001
From: Aman Gupta <amangupta052@gmail.com>
Date: Tue, 29 Jul 2025 14:45:18 +0800
Subject: [PATCH 08/86] CUDA: add roll (llama/14919)

* CUDA: add roll

* Make everything const, use __restrict__
---
 ggml/src/ggml-cuda/ggml-cuda.cu |  9 +++++
 ggml/src/ggml-cuda/roll.cu      | 67 +++++++++++++++++++++++++++++++++
 ggml/src/ggml-cuda/roll.cuh     |  5 +++
 3 files changed, 81 insertions(+)
 create mode 100644 ggml/src/ggml-cuda/roll.cu
 create mode 100644 ggml/src/ggml-cuda/roll.cuh

diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index 03c380897cd..1f785796014 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -31,6 +31,7 @@
 #include "ggml-cuda/pool2d.cuh"
 #include "ggml-cuda/quantize.cuh"
 #include "ggml-cuda/rope.cuh"
+#include "ggml-cuda/roll.cuh"
 #include "ggml-cuda/scale.cuh"
 #include "ggml-cuda/softmax.cuh"
 #include "ggml-cuda/ssm-conv.cuh"
@@ -2419,6 +2420,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_ROPE_BACK:
             ggml_cuda_op_rope_back(ctx, dst);
             break;
+        case GGML_OP_ROLL:
+            ggml_cuda_op_roll(ctx, dst);
+            break;
         case GGML_OP_IM2COL:
             ggml_cuda_op_im2col(ctx, dst);
             break;
@@ -3411,6 +3415,11 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             memcpy(&max_bias, (const float *) op->op_params + 1, sizeof(float));
             return max_bias == 0.0f;
         }
+        case GGML_OP_ROLL:
+            if(op->src[0]->type == GGML_TYPE_F32) {
+                return true;
+            }
+            return false;
         case GGML_OP_ROPE:
         case GGML_OP_ROPE_BACK: {
             return op->src[0]->nb[0] == ggml_type_size(op->src[0]->type) && ggml_is_contiguous_2(op->src[0]);
diff --git a/ggml/src/ggml-cuda/roll.cu b/ggml/src/ggml-cuda/roll.cu
new file mode 100644
index 00000000000..a339dfc1ae0
--- /dev/null
+++ b/ggml/src/ggml-cuda/roll.cu
@@ -0,0 +1,67 @@
+#include "ggml-cuda/common.cuh"
+#include "roll.cuh"
+
+static __forceinline__ __device__ int64_t wrap_index(const int64_t idx, const int64_t ne) {
+    if (idx < 0) {
+        return idx + ne;
+    }
+    if (idx >= ne) {
+        return idx - ne;
+    }
+    return idx;
+}
+
+static __global__ void roll_f32_cuda(const float * __restrict__ src,
+                                     float * __restrict__ dst,
+                                     const int64_t ne00,
+                                     const int64_t ne01,
+                                     const int64_t ne02,
+                                     const int64_t ne03,
+                                     const int     s0,
+                                     const int     s1,
+                                     const int     s2,
+                                     const int     s3) {
+    const int64_t idx        = int64_t(blockDim.x) * blockIdx.x + threadIdx.x;
+    const int64_t n_elements = ne00 * ne01 * ne02 * ne03;
+
+    if (idx >= n_elements) {
+        return;
+    }
+
+    const int64_t i0 = idx % ne00;
+    const int64_t i1 = (idx / ne00) % ne01;
+    const int64_t i2 = (idx / (ne00 * ne01)) % ne02;
+    const int64_t i3 = (idx / (ne00 * ne01 * ne02)) % ne03;
+
+    const int64_t d0 = wrap_index(i0 - s0, ne00);
+    const int64_t d1 = wrap_index(i1 - s1, ne01);
+    const int64_t d2 = wrap_index(i2 - s2, ne02);
+    const int64_t d3 = wrap_index(i3 - s3, ne03);
+
+    dst[i3 * (ne00 * ne01 * ne02) + i2 * (ne01 * ne00) + i1 * ne00 + i0] =
+        src[d3 * (ne00 * ne01 * ne02) + d2 * (ne01 * ne00) + d1 * ne00 + d0];
+}
+
+void ggml_cuda_op_roll(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    int s0 = dst->op_params[0];
+    int s1 = dst->op_params[1];
+    int s2 = dst->op_params[2];
+    int s3 = dst->op_params[3];
+
+    const ggml_tensor * src0   = dst->src[0];
+    const float *       src0_d = (const float *) dst->src[0]->data;
+    float *             dst_d  = (float *) dst->data;
+
+    GGML_TENSOR_UNARY_OP_LOCALS;
+
+    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_are_same_shape(dst->src[0], dst));
+
+    cudaStream_t stream = ctx.stream();
+
+    int64_t sz         = (ne00 * ne01 * ne02 * ne03);
+    int64_t num_blocks = (sz + CUDA_ROLL_BLOCK_SIZE - 1) / CUDA_ROLL_BLOCK_SIZE;
+
+    roll_f32_cuda<<<num_blocks, CUDA_ROLL_BLOCK_SIZE, 0, stream>>>(
+        src0_d, dst_d, ne00, ne01, ne02, ne03, s0, s1, s2, s3);
+}
diff --git a/ggml/src/ggml-cuda/roll.cuh b/ggml/src/ggml-cuda/roll.cuh
new file mode 100644
index 00000000000..322d55436e2
--- /dev/null
+++ b/ggml/src/ggml-cuda/roll.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_ROLL_BLOCK_SIZE 256
+
+void ggml_cuda_op_roll(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

From b66cc8616ecb9048698df87c07bfb3386f474ddb Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= <sigbjorn.skjaeret@scala.com>
Date: Tue, 29 Jul 2025 14:22:03 +0200
Subject: [PATCH 09/86] cuda : add softcap fusion (llama/14907)

---
 ggml/src/ggml-cuda/ggml-cuda.cu | 49 +++++++++++++++++++++++++++++----
 ggml/src/ggml-cuda/softcap.cu   | 34 +++++++++++++++++++++++
 ggml/src/ggml-cuda/softcap.cuh  |  5 ++++
 3 files changed, 82 insertions(+), 6 deletions(-)
 create mode 100644 ggml/src/ggml-cuda/softcap.cu
 create mode 100644 ggml/src/ggml-cuda/softcap.cuh

diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index 1f785796014..819c2be78ec 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -33,6 +33,7 @@
 #include "ggml-cuda/rope.cuh"
 #include "ggml-cuda/roll.cuh"
 #include "ggml-cuda/scale.cuh"
+#include "ggml-cuda/softcap.cuh"
 #include "ggml-cuda/softmax.cuh"
 #include "ggml-cuda/ssm-conv.cuh"
 #include "ggml-cuda/ssm-scan.cuh"
@@ -2770,7 +2771,12 @@ static void update_cuda_graph_executable(ggml_backend_cuda_context * cuda_ctx) {
 }
 #endif
 
-static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, std::initializer_list<enum ggml_op> ops) {
+static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, std::initializer_list<enum ggml_op> ops, std::initializer_list<enum ggml_unary_op> unary_ops) {
+#ifndef NDEBUG
+    const size_t num_unary = std::count(ops.begin(), ops.end(), GGML_OP_UNARY);
+    GGML_ASSERT(unary_ops.size() == num_unary);
+#endif
+
     if (!ggml_can_fuse(cgraph, node_idx, ops)) {
         return false;
     }
@@ -2798,9 +2804,32 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
         if (!ggml_is_contiguous_rows(mul->src[0]) || !ggml_is_contiguous_rows(mul->src[1])) {
             return false;
         }
+
+        return true;
     }
 
-    return true;
+    if (ops.size() == 3 && ops.begin()[0] == GGML_OP_SCALE && ops.begin()[1] == GGML_OP_UNARY && ops.begin()[2] == GGML_OP_SCALE
+     && unary_ops.size() == 1 && unary_ops.begin()[0] == GGML_UNARY_OP_TANH) {
+        const ggml_tensor *scale  = cgraph->nodes[node_idx];
+        const ggml_tensor *tanh   = cgraph->nodes[node_idx+1];
+        const ggml_tensor *scale2 = cgraph->nodes[node_idx+2];
+
+        GGML_ASSERT(scale->src[0]->type == GGML_TYPE_F32);
+        GGML_ASSERT(scale->type == GGML_TYPE_F32);
+
+        if (ggml_get_unary_op(tanh) != GGML_UNARY_OP_TANH) {
+            return false;
+        }
+
+        // Check for bias
+        if (ggml_get_op_params_f32(scale, 1) != 0.0f || ggml_get_op_params_f32(scale2, 1) != 0.0f) {
+            return false;
+        }
+
+        return true;
+    }
+
+    return false;
 }
 
 static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
@@ -2821,10 +2850,18 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                 }
 
                 static bool disable_fusion = (getenv("GGML_CUDA_DISABLE_FUSION") != nullptr);
-                if (!disable_fusion && ggml_cuda_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
-                    ggml_cuda_op_rms_norm_fused(*cuda_ctx, node, cgraph->nodes[i+1]);
-                    i++;
-                    continue;
+                if (!disable_fusion) {
+                    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL }, {})) {
+                        ggml_cuda_op_rms_norm_fused(*cuda_ctx, node, cgraph->nodes[i+1]);
+                        i++;
+                        continue;
+                    }
+
+                    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_SCALE, GGML_OP_UNARY, GGML_OP_SCALE }, { GGML_UNARY_OP_TANH })) {
+                        i += 2;
+                        ggml_cuda_op_softcap(*cuda_ctx, cgraph->nodes[i], node);
+                        continue;
+                    }
                 }
 #ifndef NDEBUG
                 assert(node->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device));
diff --git a/ggml/src/ggml-cuda/softcap.cu b/ggml/src/ggml-cuda/softcap.cu
new file mode 100644
index 00000000000..40dfe45d65c
--- /dev/null
+++ b/ggml/src/ggml-cuda/softcap.cu
@@ -0,0 +1,34 @@
+#include "softcap.cuh"
+
+static __global__ void softcap_f32(const float * x, float * dst, const float scale, const float softcap, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    dst[i] = tanhf(scale * x[i]) * softcap;
+}
+
+static void softcap_f32_cuda(const float * x, float * dst, const float scale, const float softcap, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_SOFTCAP_BLOCK_SIZE - 1) / CUDA_SOFTCAP_BLOCK_SIZE;
+    softcap_f32<<<num_blocks, CUDA_SOFTCAP_BLOCK_SIZE, 0, stream>>>(x, dst, scale, softcap, k);
+}
+
+// fused GGML_OP_SCALE + GGML_UNARY_OP_TANH + GGML_OP_SCALE
+void ggml_cuda_op_softcap(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * src) {
+    const ggml_tensor * src0 = src->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    float scale;
+    float softcap;
+    memcpy(&scale,   (float *) src->op_params + 0, sizeof(float));
+    memcpy(&softcap, (float *) dst->op_params + 0, sizeof(float));
+
+    softcap_f32_cuda(src0_d, dst_d, scale, softcap, ggml_nelements(src0), stream);
+}
diff --git a/ggml/src/ggml-cuda/softcap.cuh b/ggml/src/ggml-cuda/softcap.cuh
new file mode 100644
index 00000000000..6d34fb2bee4
--- /dev/null
+++ b/ggml/src/ggml-cuda/softcap.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_SOFTCAP_BLOCK_SIZE 256
+
+void ggml_cuda_op_softcap(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * src);

From 2bd2afff2b349f1765e85adeca8b2e1c71a8b006 Mon Sep 17 00:00:00 2001
From: hipudding <huafengchun@gmail.com>
Date: Tue, 29 Jul 2025 22:36:43 +0800
Subject: [PATCH 10/86] CANN: Add ggml_set_rows (llama/14943)

---
 ggml/src/ggml-cann/aclnn_ops.cpp | 154 ++++++++++++++++++++++++-------
 ggml/src/ggml-cann/aclnn_ops.h   |  16 +++-
 ggml/src/ggml-cann/ggml-cann.cpp |  19 ++--
 3 files changed, 147 insertions(+), 42 deletions(-)

diff --git a/ggml/src/ggml-cann/aclnn_ops.cpp b/ggml/src/ggml-cann/aclnn_ops.cpp
index d616c491ae9..316055193d8 100755
--- a/ggml/src/ggml-cann/aclnn_ops.cpp
+++ b/ggml/src/ggml-cann/aclnn_ops.cpp
@@ -68,6 +68,8 @@
 #include <aclnnop/aclnn_grouped_matmul_v3.h>
 #include <aclnnop/aclnn_fused_infer_attention_score_v2.h>
 #include <aclnnop/aclnn_zero.h>
+#include <aclnnop/aclnn_index_copy.h>
+#include <aclnnop/aclnn_index_select.h>
 #include <float.h>
 
 #include <cmath>
@@ -1614,50 +1616,97 @@ void ggml_cann_softmax(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
 }
 
 /**
- * @brief Performs embedding operation on a 4D tensor using the CANN backend.
+ * @brief Performs index select operation on a 4D tensor using the CANN backend.
  *
- * This function extracts slices from the source tensor (`src_buffer`),
- * index tensor (`index`), and destination tensor (`dst`), and performs an
- * embedding operation on them. The embedding operation is applied by iterating
- * over the last two dimensions of the source tensor, creating the necessary
- * tensors for the source, index, and output, and executing the embedding operation.
+ * This function applies the `IndexSelect` operation along a specific dimension
+ * of the source tensor (`src_buffer`) using the indices from the index tensor (`index`).
+ * It iterates over the last two dimensions of the source tensor, creates the corresponding
+ * CANN tensors for the source, index, and output slices, and executes the `IndexSelect`
+ * operation for each slice.
  *
  * @param ctx The context for CANN backend operations.
- * @param src_buffer The source buffer holding the data for the source tensor.
+ * @param src_buffer The source buffer containing the 4D input tensor data.
  * @param src_ne The dimensions of the source tensor.
  * @param src_nb The strides (byte offsets) of the source tensor.
- * @param index The index tensor used in the embedding operation.
- * @param dst The destination tensor where the result will be stored.
+ * @param dst_buffer The destination buffer where the output tensor data will be written.
+ * @param dst_ne The dimensions of the destination tensor.
+ * @param dst_nb The strides (byte offsets) of the destination tensor.
+ * @param index The index tensor specifying the indices to select from the source tensor.
+ * @param type The data type of the source and destination tensors.
  */
-static void aclnn_embedding_4d(ggml_backend_cann_context& ctx, void* src_buffer,
-                            int64_t* src_ne, size_t* src_nb, ggml_tensor* index,
-                            ggml_tensor* dst) {
+static void aclnn_index_select_4d(ggml_backend_cann_context& ctx,
+                                void* src_buffer,int64_t* src_ne, size_t* src_nb,
+                                void* dst_buffer, int64_t* dst_ne, size_t* dst_nb,
+                                ggml_tensor* index, ggml_type type) {
     for (int64_t i = 0; i < src_ne[3]; i++) {
         for (int64_t j = 0; j < src_ne[2]; j++) {
             // src
-            int64_t acl_src_ne[2] = {src_ne[0], src_ne[1]};
-            size_t acl_src_nb[2] = {src_nb[0], src_nb[1]};
             aclTensor* acl_src_tensor = ggml_cann_create_tensor(
                 (char*)src_buffer + i * src_nb[3] + j * src_nb[2],
-                ggml_cann_type_mapping(dst->type), ggml_element_size(dst),
-                acl_src_ne, acl_src_nb, 2);
+                ggml_cann_type_mapping(type), ggml_type_size(type),
+                src_ne, src_nb, 2);
 
             // index
-            int64_t acl_index_ne[1] = {index->ne[0]};
-            size_t acl_index_nb[1] = {index->nb[0]};
             aclTensor* acl_index = ggml_cann_create_tensor(
-                (char*)index->data + i * index->nb[2] + j * index->nb[1],
+                (char*)index->data + (i % index->ne[2]) * index->nb[2] + (j % index->ne[1]) * index->nb[1],
                 ggml_cann_type_mapping(index->type), ggml_element_size(index),
-                acl_index_ne, acl_index_nb, 1);
+                index->ne, index->nb, 1);
 
             // out
-            int64_t acl_out_ne[2] = {dst->ne[0], dst->ne[1]};
-            size_t acl_out_nb[2] = {dst->nb[0], dst->nb[1]};
             aclTensor* acl_out = ggml_cann_create_tensor(
-                (char*)dst->data + i * dst->nb[3] + j * dst->nb[2],
-                ggml_cann_type_mapping(dst->type), ggml_element_size(dst),
-                acl_out_ne, acl_out_nb, 2);
-            GGML_CANN_CALL_ACLNN_OP(ctx, Embedding, acl_src_tensor, acl_index, acl_out);
+                (char*)dst_buffer + i * dst_nb[3] + j * dst_nb[2],
+                ggml_cann_type_mapping(type), ggml_type_size(type),
+                dst_ne, dst_nb, 2);
+            GGML_CANN_CALL_ACLNN_OP(ctx, IndexSelect, acl_src_tensor, 0, acl_index, acl_out);
+            ggml_cann_release_resources(ctx, acl_src_tensor, acl_index, acl_out);
+        }
+    }
+}
+
+/**
+ * @brief Performs inplace index copy operation on a 4D tensor using the CANN backend.
+ *
+ * This function applies the `IndexCopy` operation along a specific dimension of the
+ * destination tensor (`dst_buffer`) by copying elements from the source tensor (`src_buffer`)
+ * to positions specified by the index tensor (`index`).
+ * It iterates over the last two dimensions of the tensors, creates the corresponding
+ * CANN tensors for source, index, and destination slices, and performs the index copy
+ * operation for each slice.
+ *
+ * @param ctx The context for CANN backend operations.
+ * @param src_buffer The source buffer containing the 4D input tensor data to be copied.
+ * @param src_ne The dimensions of the source tensor.
+ * @param src_nb The strides (byte offsets) of the source tensor.
+ * @param dst_buffer The destination buffer where values will be copied to.
+ * @param dst_ne The dimensions of the destination tensor.
+ * @param dst_nb The strides (byte offsets) of the destination tensor.
+ * @param index The index tensor specifying target positions in the destination tensor.
+ * @param type The data type of the source and destination tensors.
+ */
+static void aclnn_index_copy_4d(ggml_backend_cann_context& ctx,
+                                void* src_buffer,int64_t* src_ne, size_t* src_nb,
+                                void* dst_buffer, int64_t* dst_ne, size_t* dst_nb,
+                                ggml_tensor* index, ggml_type type) {
+    for (int64_t i = 0; i < src_ne[3]; i++) {
+        for (int64_t j = 0; j < src_ne[2]; j++) {
+            // src
+            aclTensor* acl_src_tensor = ggml_cann_create_tensor(
+                (char*)src_buffer + i * src_nb[3] + j * src_nb[2],
+                ggml_cann_type_mapping(type), ggml_type_size(type),
+                src_ne, src_nb, 2);
+
+            // index
+            aclTensor* acl_index = ggml_cann_create_tensor(
+                (char*)index->data + (i % index->ne[2]) * index->nb[2] + (j % index->ne[1]) * index->nb[1],
+                ggml_cann_type_mapping(index->type), ggml_element_size(index),
+                index->ne, index->nb, 1);
+
+            // out
+            aclTensor* acl_out = ggml_cann_create_tensor(
+                (char*)dst_buffer + i * dst_nb[3] + j * dst_nb[2],
+                ggml_cann_type_mapping(type), ggml_type_size(type),
+                dst_ne, dst_nb, 2);
+            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceIndexCopy, acl_out, 0, acl_index, acl_src_tensor);
             ggml_cann_release_resources(ctx, acl_src_tensor, acl_index, acl_out);
         }
     }
@@ -1669,8 +1718,9 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
 
     switch (src0->type) {
         case GGML_TYPE_F32: {
-            aclnn_embedding_4d(ctx, src0->data, src0->ne, src0->nb, src1,
-                                   dst);
+            aclnn_index_select_4d(ctx, src0->data, src0->ne, src0->nb,
+                                dst->data, dst->ne, dst->nb,
+                                src1, dst->type);
             break;
         }
         case GGML_TYPE_F16: {
@@ -1687,8 +1737,9 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 src_trans_buffer, ACL_FLOAT, ggml_type_size(dst->type),
                 src0->ne, src_trans_nb, GGML_MAX_DIMS);
             aclnn_cast(ctx, acl_src0, src_trans_tensor, ggml_cann_type_mapping(dst->type));
-            aclnn_embedding_4d(ctx, src_trans_buffer, src0->ne,
-                                   src_trans_nb, src1, dst);
+            aclnn_index_select_4d(ctx, src_trans_buffer, src0->ne, src_trans_nb,
+                                dst->data, dst->ne, dst->nb,
+                                src1, dst->type);
             ggml_cann_release_resources(ctx, acl_src0, src_trans_tensor);
             break;
         }
@@ -1748,8 +1799,10 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 dequant_nb[i] = dequant_nb[i - 1] * src0->ne[i - 1];
             }
 
-            aclnn_embedding_4d(ctx, dequant_buffer_allocator.get(),
-                                   dequant_ne, dequant_nb, src1, dst);
+            aclnn_index_select_4d(ctx, dequant_buffer_allocator.get(),
+                                   dequant_ne, dequant_nb,
+                                   dst->data, dst->ne, dst->nb,
+                                   src1, dst->type);
 
             ggml_cann_release_resources(ctx, dequant_tensor);
             break;
@@ -1760,6 +1813,43 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     }
 }
 
+void ggml_cann_set_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src0 = dst->src[0];  // src
+    ggml_tensor* src1 = dst->src[1];  // index
+
+    switch (dst->type) {
+        case GGML_TYPE_F32: {
+            aclnn_index_copy_4d(ctx, src0->data, src0->ne, src0->nb,
+                                dst->data, dst->ne, dst->nb,
+                                src1, dst->type);
+            break;
+        }
+        case GGML_TYPE_F16: {
+            aclTensor* acl_src0 = ggml_cann_create_tensor(src0);
+            ggml_cann_pool_alloc src_buffer_allocator(
+                ctx.pool(), ggml_nelements(src0) * sizeof(uint16_t));
+            void* src_trans_buffer = src_buffer_allocator.get();
+            size_t src_trans_nb[GGML_MAX_DIMS];
+            src_trans_nb[0] = sizeof(uint16_t);
+            for (int i = 1; i < GGML_MAX_DIMS; i++) {
+                src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1];
+            }
+            aclTensor* src_trans_tensor = ggml_cann_create_tensor(
+                src_trans_buffer, ACL_FLOAT16, ggml_type_size(dst->type),
+                src0->ne, src_trans_nb, GGML_MAX_DIMS);
+            aclnn_cast(ctx, acl_src0, src_trans_tensor, ggml_cann_type_mapping(dst->type));
+            aclnn_index_copy_4d(ctx, src_trans_buffer, src0->ne, src_trans_nb,
+                                dst->data, dst->ne, dst->nb,
+                                src1, dst->type);
+            ggml_cann_release_resources(ctx, acl_src0, src_trans_tensor);
+            break;
+        }
+        default:
+            GGML_ABORT("Unsupported tensor type for GGML_OP_SET_ROWS");
+            break;
+    }
+}
+
 /**
  * @brief Repeats elements of a tensor along a specified dimension.
  *
diff --git a/ggml/src/ggml-cann/aclnn_ops.h b/ggml/src/ggml-cann/aclnn_ops.h
index 8deaf7ea1db..5c510cc9932 100755
--- a/ggml/src/ggml-cann/aclnn_ops.h
+++ b/ggml/src/ggml-cann/aclnn_ops.h
@@ -424,15 +424,25 @@ void ggml_cann_softmax(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  *
  * @details This function retrieves rows from a source tensor src0 according to
  *          the indices provided in another tensor src1 and stores the result in
- *          a destination tensor (\p dst). It supports different data types
- *          including F32, F16, Q4_0, and Q8_0.
+ *          a destination tensor (\p dst).
  *
  * @param ctx The backend CANN context for executing operations.
  * @param dst The destination tensor where the extracted rows will be stored.
- *            dst->op is `GGML_OP_GET_ROWS`.
  */
 void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
+/**
+ * @brief   Writes specific rows into a tensor at positions specified by indices.
+ *
+ * @details This function copies rows from a source tensor into a destination
+ *          tensor (\p dst) at the positions indicated by the indices in another
+ *          tensor.
+ *
+ * @param ctx The backend CANN context for executing operations.
+ * @param dst The destination tensor where the specified rows will be updated.
+ */
+void ggml_cann_set_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
 /**
  * @brief   Executes matrix multiplication for the given tensor.
  *
diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
index c6edb6b61bb..4dac2e8b789 100755
--- a/ggml/src/ggml-cann/ggml-cann.cpp
+++ b/ggml/src/ggml-cann/ggml-cann.cpp
@@ -1659,6 +1659,9 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
         case GGML_OP_GET_ROWS:
             ggml_cann_get_rows(ctx, dst);
             break;
+        case GGML_OP_SET_ROWS:
+            ggml_cann_set_rows(ctx, dst);
+            break;
         case GGML_OP_DUP:
             ggml_cann_dup(ctx, dst);
             break;
@@ -2191,13 +2194,15 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                     return false;
             }
         } break;
-        case GGML_OP_SET_ROWS:
-            {
-                // TODO: add support
-                // ref: https://github.com/ggml-org/llama.cpp/pull/14274
-#pragma message("TODO: implement F32, F16, BF16, Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, IQ4_NL support (https://github.com/ggml-org/llama.cpp/pull/14661)")
-                return false;
-            } break;
+        case GGML_OP_SET_ROWS: {
+            switch (op->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                    return true;
+                default:
+                    return false;
+            }
+        } break;
         case GGML_OP_CPY: {
             ggml_tensor *src = op->src[0];
             if ((op->type != GGML_TYPE_F32 && op->type != GGML_TYPE_F16) ||

From f7d3ac60de866f029fb0783d1256eb9bd1a121b0 Mon Sep 17 00:00:00 2001
From: uvos <carl@uvos.xyz>
Date: Tue, 29 Jul 2025 17:43:43 +0200
Subject: [PATCH 11/86] HIP: Ignore unsupported unroll transformation in
 fattn-vec (llama/14931)

llvm with the amdgcn target dose not support unrolling loops with conditional break statements, when those statements can not be resolved at compile time. Similar to other places in GGML lets simply ignore this warning.
---
 ggml/src/ggml-cuda/fattn-vec-f16.cuh | 9 +++++++++
 ggml/src/ggml-cuda/fattn-vec-f32.cuh | 9 +++++++++
 2 files changed, 18 insertions(+)

diff --git a/ggml/src/ggml-cuda/fattn-vec-f16.cuh b/ggml/src/ggml-cuda/fattn-vec-f16.cuh
index 109253838f2..afef815ceaf 100644
--- a/ggml/src/ggml-cuda/fattn-vec-f16.cuh
+++ b/ggml/src/ggml-cuda/fattn-vec-f16.cuh
@@ -1,6 +1,12 @@
 #include "common.cuh"
 #include "fattn-common.cuh"
 
+// Currenlty llvm with the amdgcn target dose not support unrolling loops
+// that contain a break that can not be resolved at compile time.
+#ifdef __clang__
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wpass-failed"
+#endif // __clang__
 template<int D, int ncols, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
 #ifndef GGML_USE_HIP
 __launch_bounds__(D, 1)
@@ -341,6 +347,9 @@ static __global__ void flash_attn_vec_ext_f16(
     NO_DEVICE_CODE;
 #endif // defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)
 }
+#ifdef __clang__
+#pragma clang diagnostic pop
+#endif // __clang__
 
 template <int D, int cols_per_block, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
 void ggml_cuda_flash_attn_ext_vec_f16_case_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
diff --git a/ggml/src/ggml-cuda/fattn-vec-f32.cuh b/ggml/src/ggml-cuda/fattn-vec-f32.cuh
index 2cf2e408e92..3595e29693a 100644
--- a/ggml/src/ggml-cuda/fattn-vec-f32.cuh
+++ b/ggml/src/ggml-cuda/fattn-vec-f32.cuh
@@ -1,6 +1,12 @@
 #include "common.cuh"
 #include "fattn-common.cuh"
 
+// Currenlty llvm with the amdgcn target dose not support unrolling loops
+// that contain a break that can not be resolved at compile time.
+#ifdef __clang__
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wpass-failed"
+#endif // __clang__
 template<int D, int ncols, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
 #ifndef GGML_USE_HIP
 __launch_bounds__(D, 1)
@@ -336,6 +342,9 @@ static __global__ void flash_attn_vec_ext_f32(
     NO_DEVICE_CODE;
 #endif // FLASH_ATTN_AVAILABLE
 }
+#ifdef __clang__
+#pragma clang diagnostic pop
+#endif // __clang__
 
 template <int D, int cols_per_block, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
 void ggml_cuda_flash_attn_ext_vec_f32_case_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {

From 01e161426578ff1fcee13651523eb331e7147027 Mon Sep 17 00:00:00 2001
From: uvos <carl@uvos.xyz>
Date: Tue, 29 Jul 2025 17:44:30 +0200
Subject: [PATCH 12/86] HIP: add GGML_HIP_MMQ_MFMA option to allow disableing
 the MFMA path. (llama/14930)

This is useful for testing for regressions on GCN with CDNA hardware.

With GGML_HIP_MMQ_MFMA=Off and GGML_CUDA_FORCE_MMQ=On we can conveniently test the GCN code path on CDNA. As CDNA is just GCN renamed with MFMA added and limited use ACC registers, this provides a good alternative for regression testing when GCN hardware is not available.
---
 ggml/CMakeLists.txt              | 1 +
 ggml/src/ggml-cuda/common.cuh    | 8 ++++++--
 ggml/src/ggml-hip/CMakeLists.txt | 4 ++++
 3 files changed, 11 insertions(+), 2 deletions(-)

diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt
index 20467c54da1..231250efce0 100644
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -174,6 +174,7 @@ option(GGML_HIP_GRAPHS                      "ggml: use HIP graph, experimental,
 option(GGML_HIP_NO_VMM                      "ggml: do not try to use HIP VMM"                 ON)
 option(GGML_HIP_ROCWMMA_FATTN               "ggml: enable rocWMMA for FlashAttention"         OFF)
 option(GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12   "ggml: enable rocWMMA FlashAttention on GFX12"    OFF)
+option(GGML_HIP_MMQ_MFMA                    "ggml: enable MFMA MMA for CDNA in MMQ"           ON)
 option(GGML_MUSA_GRAPHS                     "ggml: use MUSA graph, experimental, unstable"    OFF)
 option(GGML_MUSA_MUDNN_COPY                 "ggml: enable muDNN for accelerated copy"         OFF)
 option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
index cdc3bb5ae76..19fcc598224 100644
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -227,7 +227,7 @@ typedef float2 dfloat2;
 #define FP16_MMA_AVAILABLE
 #endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4)))
 
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && defined(CDNA3)
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && defined(CDNA3) && !defined(GGML_HIP_NO_MMQ_MFMA)
 #define AMD_MFMA_AVAILABLE
 #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && defined(CDNA3)
 
@@ -295,7 +295,11 @@ static bool fp32_mma_hardware_available(const int cc) {
 
 // AMD CDNA3 matrix cores.. Will add support for other CDNA generations later.
 static bool amd_mfma_available(const int cc) {
-    return cc >= GGML_CUDA_CC_OFFSET_AMD && GGML_CUDA_CC_IS_CDNA3(cc);
+#if !defined(GGML_HIP_NO_MMQ_MFMA)
+    return GGML_CUDA_CC_IS_CDNA3(cc);
+#else
+    return false;
+#endif //!defined(GGML_HIP_NO_MMQ_MFMA)
 }
 
 // Volta technically had FP16 tensor cores but they work very differently compared to Turing and later.
diff --git a/ggml/src/ggml-hip/CMakeLists.txt b/ggml/src/ggml-hip/CMakeLists.txt
index e29df98560e..e92ec7faa33 100644
--- a/ggml/src/ggml-hip/CMakeLists.txt
+++ b/ggml/src/ggml-hip/CMakeLists.txt
@@ -113,6 +113,10 @@ if (GGML_HIP_ROCWMMA_FATTN)
     add_compile_definitions(GGML_HIP_ROCWMMA_FATTN)
 endif()
 
+if (NOT GGML_HIP_MMQ_MFMA)
+    add_compile_definitions(GGML_HIP_NO_MMQ_MFMA)
+endif()
+
 if (GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12 OR ${hip_VERSION} VERSION_GREATER_EQUAL 7.0)
     add_compile_definitions(GGML_HIP_ROCWMMA_FATTN_GFX12)
 endif()

From 980d19eed38a4784c70b8fcb19aef17f92702979 Mon Sep 17 00:00:00 2001
From: uvos <carl@uvos.xyz>
Date: Tue, 29 Jul 2025 20:23:04 +0200
Subject: [PATCH 13/86] HIP: remove the use of __HIP_PLATFORM_AMD__, explicitly
 support only AMD targets (llama/14945)

---
 ggml/src/ggml-cuda/common.cuh        | 50 ++++++++++++++--------------
 ggml/src/ggml-cuda/fattn-common.cuh  |  4 +--
 ggml/src/ggml-cuda/fattn-mma-f16.cuh |  8 ++---
 ggml/src/ggml-cuda/fattn-tile-f16.cu |  4 +--
 ggml/src/ggml-cuda/fattn-tile-f32.cu |  4 +--
 ggml/src/ggml-cuda/fattn-wmma-f16.cu |  8 ++---
 ggml/src/ggml-cuda/ggml-cuda.cu      | 10 +++---
 ggml/src/ggml-cuda/mma.cuh           |  4 +--
 ggml/src/ggml-cuda/mmq.cuh           | 26 +++++++--------
 ggml/src/ggml-cuda/vendors/hip.h     | 14 ++++----
 10 files changed, 67 insertions(+), 65 deletions(-)

diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
index 19fcc598224..44daafbf78a 100644
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -176,7 +176,7 @@ static const char * cu_get_error_str(CUresult err) {
 #define CU_CHECK(err) CUDA_CHECK_GEN(err, CUDA_SUCCESS, cu_get_error_str)
 #endif
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
 #    define CUDA_SET_SHARED_MEMORY_LIMIT(kernel, nbytes)                                                       \
         do {                                                                                                   \
             static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = { false };                         \
@@ -191,7 +191,7 @@ static const char * cu_get_error_str(CUresult err) {
         do {                                             \
             GGML_UNUSED(nbytes);                         \
         } while (0)
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+#endif // !(defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
 
 #if CUDART_VERSION >= 11010 || defined(GGML_USE_MUSA)
 #define GGML_CUDA_ASSUME(x) __builtin_assume(x)
@@ -211,9 +211,9 @@ typedef float2 dfloat2;
 #define GGML_USE_VMM
 #endif // (!defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)) || (defined(GGML_USE_HIP) && !defined(GGML_HIP_NO_VMM))
 
-#if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
+#if defined(GGML_USE_HIP) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 #define FP16_AVAILABLE
-#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
+#endif // defined(GGML_USE_HIP) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 
 #if defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
 #define FAST_FP16_AVAILABLE
@@ -227,17 +227,17 @@ typedef float2 dfloat2;
 #define FP16_MMA_AVAILABLE
 #endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4)))
 
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && defined(CDNA3) && !defined(GGML_HIP_NO_MMQ_MFMA)
+#if defined(GGML_USE_HIP) && defined(CDNA3) && !defined(GGML_HIP_NO_MMQ_MFMA)
 #define AMD_MFMA_AVAILABLE
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && defined(CDNA3)
+#endif // defined(GGML_USE_HIP) && defined(CDNA3) && !defined(GGML_HIP_NO_MMQ_MFMA)
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
+#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
 #define NEW_MMA_AVAILABLE
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
+#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #define CP_ASYNC_AVAILABLE
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 
 #if !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220)
 #define FLASH_ATTN_AVAILABLE
@@ -259,7 +259,7 @@ static bool fast_fp16_hardware_available(const int cc) {
 
 // Any FP16 tensor core instructions are available for ggml code.
 static bool fp16_mma_available(const int cc) {
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN)
+#if defined(GGML_USE_HIP) && !defined(GGML_HIP_ROCWMMA_FATTN)
     return false;
 #else
     if ((GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) ||
@@ -275,7 +275,7 @@ static bool fp16_mma_available(const int cc) {
     } else {
         return false;
     }
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN)
+#endif // defined(GGML_USE_HIP) && !defined(GGML_HIP_ROCWMMA_FATTN)
 }
 
 // To be used for feature selection of external libraries, e.g. cuBLAS.
@@ -312,25 +312,25 @@ static bool cp_async_available(const int cc) {
 }
 
 static constexpr __device__ int ggml_cuda_get_physical_warp_size() {
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(__GFX9__) || defined(__GFX8__))
+#if defined(GGML_USE_HIP) && (defined(__GFX9__) || defined(__GFX8__))
     return 64;
 #else
     return 32;
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(__GFX9__) || defined(__GFX8__))
+#endif // defined(GGML_USE_HIP) && (defined(__GFX9__) || defined(__GFX8__))
 }
 
 [[noreturn]]
 static __device__ void no_device_code(
     const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) {
 
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP)
     printf("%s:%d: ERROR: HIP kernel %s has no device code compatible with HIP arch %d.\n",
            file_name, line, function_name, arch);
     GGML_UNUSED(arch_list);
 #else
     printf("%s:%d: ERROR: CUDA kernel %s has no device code compatible with CUDA arch %d. ggml-cuda.cu was compiled for: %s\n",
            file_name, line, function_name, arch, arch_list);
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP)
     __trap();
 
     GGML_UNUSED(no_device_code); // suppress unused function warning
@@ -367,7 +367,7 @@ struct ggml_cuda_unroll<1> {
 
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ int warp_reduce_sum(int x) {
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
     return __reduce_add_sync(0xffffffff, x);
 #else
 #pragma unroll
@@ -375,7 +375,7 @@ static __device__ __forceinline__ int warp_reduce_sum(int x) {
         x += __shfl_xor_sync(0xffffffff, x, offset, width);
     }
     return x;
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 }
 
 template<int width = WARP_SIZE>
@@ -444,11 +444,11 @@ static __device__ __forceinline__ float warp_reduce_max(float x) {
 static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b) {
 #ifdef FP16_AVAILABLE
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+#if !defined(GGML_USE_HIP) && CUDART_VERSION < CUDART_HMAX
     return __float2half(fmaxf(__half2float(a), __half2float(b)));
 #else
     return __hmax(a, b);
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+#endif // !defined(GGML_USE_HIP) && CUDART_VERSION < CUDART_HMAX
 
 #else
    NO_DEVICE_CODE;
@@ -476,7 +476,7 @@ static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const hal
 
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || (defined(GGML_USE_HIP) && HIP_VERSION >= 50700000)
+#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || (defined(GGML_USE_HIP) && HIP_VERSION >= 50700000)
 #pragma unroll
    for (int offset = width/2; offset > 0; offset >>= 1) {
        x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, offset, width));
@@ -485,7 +485,7 @@ static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
 #else
    GGML_UNUSED(x);
    NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || (defined(GGML_USE_HIP) && HIP_VERSION >= 50700000)
+#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || (defined(GGML_USE_HIP) && HIP_VERSION >= 50700000)
 }
 
 #if CUDART_VERSION < CUDART_HMASK
@@ -497,7 +497,7 @@ static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half
 #endif // CUDART_VERSION < CUDART_HMASK
 
 static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, int c) {
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP)
 #if defined(CDNA) || defined(RDNA2) || defined(__gfx906__)
     c = __builtin_amdgcn_sdot4(a, b, c, false);
 #elif defined(RDNA3) || defined(RDNA4)
@@ -523,7 +523,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
 #endif
     return c;
 
-#else // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#else // defined(GGML_USE_HIP)
 
 #if __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A || defined(GGML_USE_MUSA)
     return __dp4a(a, b, c);
@@ -533,7 +533,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
     return c + a8[0]*b8[0] + a8[1]*b8[1] + a8[2]*b8[2] + a8[3]*b8[3];
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A || defined(GGML_USE_MUSA)
 
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP)
 }
 
 typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);
diff --git a/ggml/src/ggml-cuda/fattn-common.cuh b/ggml/src/ggml-cuda/fattn-common.cuh
index 95e704e393c..0cc74f284a1 100644
--- a/ggml/src/ggml-cuda/fattn-common.cuh
+++ b/ggml/src/ggml-cuda/fattn-common.cuh
@@ -592,9 +592,9 @@ static __global__ void flash_attn_stream_k_fixup(
 }
 
 template<int D> // D == head size
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#if !defined(GGML_USE_HIP)
 __launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP)
 static __global__ void flash_attn_combine_results(
         const float  * __restrict__ VKQ_parts,
         const float2 * __restrict__ VKQ_meta,
diff --git a/ggml/src/ggml-cuda/fattn-mma-f16.cuh b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
index 83cf872f68a..8e847d361b4 100644
--- a/ggml/src/ggml-cuda/fattn-mma-f16.cuh
+++ b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
@@ -1391,24 +1391,24 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
         constexpr bool use_logit_softcap = false;
         fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla>;
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
         static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
         if (!shared_memory_limit_raised[id]) {
             CUDA_CHECK(cudaFuncSetAttribute(fattn_kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared_total));
             shared_memory_limit_raised[id] = true;
         }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
     } else {
         constexpr bool use_logit_softcap = true;
         fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla>;
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
         static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
         if (!shared_memory_limit_raised[id]) {
             CUDA_CHECK(cudaFuncSetAttribute(fattn_kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared_total));
             shared_memory_limit_raised[id] = true;
         }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
     }
 
     launch_fattn<DV, ncols1, ncols2>
diff --git a/ggml/src/ggml-cuda/fattn-tile-f16.cu b/ggml/src/ggml-cuda/fattn-tile-f16.cu
index 7661c21efbb..678288c13e3 100644
--- a/ggml/src/ggml-cuda/fattn-tile-f16.cu
+++ b/ggml/src/ggml-cuda/fattn-tile-f16.cu
@@ -5,9 +5,9 @@
 #define FATTN_KQ_STRIDE_TILE_F16 64
 
 template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#if !defined(GGML_USE_HIP)
 __launch_bounds__(nwarps*WARP_SIZE, 2)
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // !defined(GGML_USE_HIP)
 static __global__ void flash_attn_tile_ext_f16(
         const char * __restrict__ Q,
         const char * __restrict__ K,
diff --git a/ggml/src/ggml-cuda/fattn-tile-f32.cu b/ggml/src/ggml-cuda/fattn-tile-f32.cu
index 11778bb9611..bc283d9a7e4 100644
--- a/ggml/src/ggml-cuda/fattn-tile-f32.cu
+++ b/ggml/src/ggml-cuda/fattn-tile-f32.cu
@@ -5,9 +5,9 @@
 #define FATTN_KQ_STRIDE_TILE_F32 32
 
 template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#if !defined(GGML_USE_HIP)
 __launch_bounds__(nwarps*WARP_SIZE, 2)
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // !defined(GGML_USE_HIP)
 static __global__ void flash_attn_tile_ext_f32(
         const char * __restrict__ Q,
         const char * __restrict__ K,
diff --git a/ggml/src/ggml-cuda/fattn-wmma-f16.cu b/ggml/src/ggml-cuda/fattn-wmma-f16.cu
index c9b083bed01..f4393ec5711 100644
--- a/ggml/src/ggml-cuda/fattn-wmma-f16.cu
+++ b/ggml/src/ggml-cuda/fattn-wmma-f16.cu
@@ -7,7 +7,7 @@
 #include "fattn-wmma-f16.cuh"
 
 #ifdef FP16_MMA_AVAILABLE
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#if !defined(GGML_USE_HIP)
 #include <mma.h>
 #ifdef GGML_USE_MUSA
 namespace wmma = mtmusa::wmma;
@@ -18,7 +18,7 @@ namespace wmma = nvcuda::wmma;
 #undef HIP_ENABLE_WARP_SYNC_BUILTINS // conflicts with rocWMMA headers
 #include <rocwmma/rocwmma.hpp>
 namespace wmma = rocwmma;
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // !defined(GGML_USE_HIP)
 #endif // FP16_MMA_AVAILABLE
 
 // D == head size, VKQ_stride == num VKQ rows calculated in parallel:
@@ -546,7 +546,7 @@ void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, ggml_ten
         return;
     }
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#if !defined(GGML_USE_HIP)
     if (Q->ne[1] <= 8 && Q->ne[0] % warp_size == 0) {
         constexpr int cols_per_block = 8;
         switch (Q->ne[0]) {
@@ -568,7 +568,7 @@ void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, ggml_ten
         }
         return;
     }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // !defined(GGML_USE_HIP)
 
     if (Q->ne[1] <= 32) {
         constexpr int cols_per_block = 16;
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index 819c2be78ec..51792794673 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -128,7 +128,7 @@ static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device)
     return err;
 }
 
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP)
 static int ggml_cuda_parse_id(char devName[]) {
     // A list of possible Target IDs can be found under the rocclr/clr repo in device.cpp
     // these values are not stable so this is susceptible to breakage
@@ -175,10 +175,10 @@ static int ggml_cuda_parse_id(char devName[]) {
     archNum += archMinor;
     return archNum;
 }
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP)
 
 static ggml_cuda_device_info ggml_cuda_init() {
-#ifdef __HIP_PLATFORM_AMD__
+#if defined(GGML_USE_HIP)
     // Workaround for a rocBLAS bug when using multiple graphics cards:
     // https://github.com/ROCmSoftwarePlatform/rocBLAS/issues/1346
     {
@@ -251,7 +251,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
         info.devices[id].nsm        = prop.multiProcessorCount;
         info.devices[id].smpb       = prop.sharedMemPerBlock;
         info.devices[id].warp_size  = prop.warpSize;
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP)
         info.devices[id].smpbo = prop.sharedMemPerBlock;
 
         info.devices[id].cc = ggml_cuda_parse_id(prop.gcnArchName);
@@ -281,7 +281,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
         info.devices[id].cc = 100*prop.major + 10*prop.minor;
         GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s\n",
                         id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP)
     }
 
     for (int id = 0; id < info.device_count; ++id) {
diff --git a/ggml/src/ggml-cuda/mma.cuh b/ggml/src/ggml-cuda/mma.cuh
index d6817d804d2..a86365c6a06 100644
--- a/ggml/src/ggml-cuda/mma.cuh
+++ b/ggml/src/ggml-cuda/mma.cuh
@@ -68,7 +68,7 @@ namespace ggml_cuda_mma {
         static constexpr int I  = I_;
         static constexpr int J  = J_;
 
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP)
         static constexpr int ne = I * J / 64;
         T x[ne] = {0};
 
@@ -132,7 +132,7 @@ namespace ggml_cuda_mma {
                 static_assert(I == -1 && J == -1, "template specialization not implemented");
             }
         }
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP)
     };
 
     template <int I_, int J_>
diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh
index 36e84be154e..d8650dd70b9 100644
--- a/ggml/src/ggml-cuda/mmq.cuh
+++ b/ggml/src/ggml-cuda/mmq.cuh
@@ -104,9 +104,9 @@ static constexpr __device__ int get_mmq_x_max_device() {
     return 128;
 #else // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
 
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP)
     return 64;
-#else // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#else // defined(GGML_USE_HIP)
 
 #if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
 #ifdef GGML_CUDA_FORCE_MMQ
@@ -118,7 +118,7 @@ static constexpr __device__ int get_mmq_x_max_device() {
     return 64;
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
 
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP)
 #endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
 }
 
@@ -128,7 +128,7 @@ static int get_mmq_y_host(const int cc) {
 }
 
 static constexpr __device__ int get_mmq_y_device() {
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP)
 #if defined(RDNA1)
     return 64;
 #else
@@ -140,7 +140,7 @@ static constexpr __device__ int get_mmq_y_device() {
 #else
     return 64;
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP)
 }
 
 // Decouple shared memory tile sizes from WARP_SIZE to allow for different warp sizes.
@@ -250,7 +250,7 @@ static constexpr __device__ int mmq_get_granularity_device(const int /*mmq_x*/)
 }
 #endif // AMD_MFMA_AVAILABLE
 
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP)
 static int mmq_get_nwarps_host(const int cc) {
     return amd_mfma_available(cc) ? 8 : 4;
 }
@@ -258,10 +258,10 @@ static int mmq_get_nwarps_host(const int cc) {
 static int mmq_get_nwarps_host(const int /*cc*/) {
     return 8;
 }
-#endif // (GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // (GGML_USE_HIP)
 
 static constexpr __device__ int mmq_get_nwarps_device() {
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP)
 #if defined(AMD_MFMA_AVAILABLE)
     return 8;
 #else
@@ -269,7 +269,7 @@ static constexpr __device__ int mmq_get_nwarps_device() {
 #endif // AMD_MFMA_AVAILABLE
 #else
     return 8;
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP)
 }
 
 // ------------------------------------------------------------
@@ -3047,7 +3047,7 @@ static __device__ __forceinline__ void mul_mat_q_process_tile(
 // The mul_mat_q kernel implements "stream-k" work partitioning as described in https://arxiv.org/abs/2301.03598
 
 template <ggml_type type, int mmq_x, bool need_check>
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP)
 #if defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
     __launch_bounds__(ggml_cuda_get_physical_warp_size()*mmq_get_nwarps_device(), 2)
 #endif // defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
@@ -3057,7 +3057,7 @@ template <ggml_type type, int mmq_x, bool need_check>
 #else
     __launch_bounds__(ggml_cuda_get_physical_warp_size()*mmq_get_nwarps_device(), 2)
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP)
 static __global__ void mul_mat_q(
         const char * __restrict__ x, const int * __restrict__ y, const int32_t * __restrict__ ids_dst,
         const int32_t * __restrict__ expert_bounds, float * __restrict__ dst, float * __restrict__ tmp_fixup,
@@ -3097,7 +3097,7 @@ static __global__ void mul_mat_q(
     __syncthreads();
 
     // On AMD or old CUDA the performance with stream-k was worse, use conventional tiling instead:
-#if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(CDNA3)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA
+#if (defined(GGML_USE_HIP) && !defined(CDNA3)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA
     {
         const int wt = blockIdx.z / nchannels_y;
         const int zt = blockIdx.z - wt*nchannels_y;
@@ -3151,7 +3151,7 @@ static __global__ void mul_mat_q(
              tile_x_max_i, tile_y_max_j, 0, ncols_x/qk);
         return;
     }
-#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(CDNA3)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA
+#endif // (defined(GGML_USE_HIP) && !defined(CDNA3)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA
 
     const     int64_t blocks_per_ne00 = ncols_x / qk;
     constexpr int     blocks_per_iter = MMQ_ITER_K / qk;
diff --git a/ggml/src/ggml-cuda/vendors/hip.h b/ggml/src/ggml-cuda/vendors/hip.h
index 56e59a058f9..8b172e60f4b 100644
--- a/ggml/src/ggml-cuda/vendors/hip.h
+++ b/ggml/src/ggml-cuda/vendors/hip.h
@@ -5,10 +5,8 @@
 #include <hipblas/hipblas.h>
 #include <hip/hip_fp16.h>
 #include <hip/hip_bfloat16.h>
-#ifdef __HIP_PLATFORM_AMD__
 // for rocblas_initialize()
 #include "rocblas/rocblas.h"
-#endif // __HIP_PLATFORM_AMD__
 
 #define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
 #define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
@@ -139,7 +137,7 @@
 #define CUBLAS_STATUS_INTERNAL_ERROR HIPBLAS_STATUS_INTERNAL_ERROR
 #define CUBLAS_STATUS_NOT_SUPPORTED HIPBLAS_STATUS_NOT_SUPPORTED
 
-#if defined(__HIP_PLATFORM_AMD__) && HIP_VERSION >= 70000000
+#if HIP_VERSION >= 70000000
 #define CUBLAS_COMPUTE_16F HIPBLAS_COMPUTE_16F
 #define CUBLAS_COMPUTE_32F HIPBLAS_COMPUTE_32F
 #define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_COMPUTE_32F_FAST_16F
@@ -151,7 +149,11 @@
 #define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
 #define cublasComputeType_t hipblasDatatype_t
 #define cudaDataType_t hipblasDatatype_t
-#endif
+#endif // HIP_VERSION >= 7000000
+
+#if !defined(__HIP_PLATFORM_AMD__)
+#error "The HIP backend supports only AMD targets"
+#endif // !defined(__HIP_PLATFORM_AMD__)
 
 #define __CUDA_ARCH__ 1300
 
@@ -249,7 +251,7 @@ static __device__ __forceinline__ unsigned int __vcmpne4(unsigned int a, unsigne
     return c;
 }
 
-#if defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
+#if HIP_VERSION < 50600000
 // __shfl_xor() for half2 was added in ROCm 5.6
 static __device__ __forceinline__ half2 __shfl_xor(half2 var, int laneMask, int width) {
     typedef union half2_b32 {
@@ -261,4 +263,4 @@ static __device__ __forceinline__ half2 __shfl_xor(half2 var, int laneMask, int
     tmp.b32 = __shfl_xor(tmp.b32, laneMask, width);
     return tmp.val;
 }
-#endif // defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
+#endif // HIP_VERSION < 50600000

From c7162815594e521673ab2866f552d53cfd5a8084 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= <johannesg@5d6.de>
Date: Wed, 30 Jul 2025 15:46:13 +0200
Subject: [PATCH 14/86] CUDA: skip masked KV slices for all FA kernels
 (llama/14924)

---
 ggml/src/ggml-cuda/common.cuh        | 14 ++++++
 ggml/src/ggml-cuda/fattn-common.cuh  | 75 +++++++++++++++++++++++++++-
 ggml/src/ggml-cuda/fattn-mma-f16.cuh | 21 ++++++--
 ggml/src/ggml-cuda/fattn-tile-f16.cu |  4 +-
 ggml/src/ggml-cuda/fattn-tile-f32.cu |  4 +-
 ggml/src/ggml-cuda/fattn-vec-f16.cuh | 26 ++--------
 ggml/src/ggml-cuda/fattn-vec-f32.cuh | 25 ++--------
 ggml/src/ggml-cuda/fattn-wmma-f16.cu |  4 +-
 ggml/src/ggml-cuda/fattn.cu          |  3 +-
 9 files changed, 120 insertions(+), 56 deletions(-)

diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
index 44daafbf78a..4560d85c4cf 100644
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -432,6 +432,20 @@ static __global__ void reduce_rows_f32(const float * x, float * dst, const int n
     dst[row] = norm ? sum / ncols : sum;
 }
 
+template<int width = WARP_SIZE>
+static __device__ __forceinline__ int warp_reduce_all(int x) {
+#ifdef GGML_USE_HIP
+#pragma unroll
+    for (int offset = width/2; offset > 0; offset >>= 1) {
+        x = x && __shfl_xor_sync(0xffffffff, x, offset, width);
+    }
+    return x;
+#else
+    static_assert(width == WARP_SIZE, "width != WARP_SIZE not implemented");
+    return __all_sync(0xffffffff, x);
+#endif // GGML_USE_HIP
+}
+
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ float warp_reduce_max(float x) {
 #pragma unroll
diff --git a/ggml/src/ggml-cuda/fattn-common.cuh b/ggml/src/ggml-cuda/fattn-common.cuh
index 0cc74f284a1..b6db446c6fe 100644
--- a/ggml/src/ggml-cuda/fattn-common.cuh
+++ b/ggml/src/ggml-cuda/fattn-common.cuh
@@ -15,6 +15,7 @@ typedef void (* fattn_kernel_t)(
         const char * __restrict__ K,
         const char * __restrict__ V,
         const char * __restrict__ mask,
+        const int  * __restrict__ KV_max,
         float      * __restrict__ dst,
         float2     * __restrict__ dst_meta,
         const float scale,
@@ -500,6 +501,55 @@ constexpr __device__ dequantize_1_f32_t get_dequantize_1_f32(ggml_type type_V) {
         nullptr;
 }
 
+template <int ncols1>
+__launch_bounds__(FATTN_KQ_STRIDE/2, 1)
+static __global__ void flash_attn_mask_to_KV_max(
+        const half2 * __restrict__ mask, int * __restrict__ KV_max, const int ne30, const int s31, const int s33) {
+    const int ne31     = gridDim.x;
+    const int tid      = threadIdx.x;
+    const int sequence = blockIdx.y;
+    const int jt       = blockIdx.x;
+
+    mask += sequence*s33 + jt*ncols1*s31;
+
+    __shared__ int buf_iw[WARP_SIZE];
+    if (tid < WARP_SIZE) {
+        buf_iw[tid] = 1;
+    }
+    __syncthreads();
+
+    int KV_max_sj = (ne30 - 1) * FATTN_KQ_STRIDE;
+    for (; KV_max_sj >= 0; KV_max_sj -= FATTN_KQ_STRIDE) {
+        int all_inf = 1;
+
+#pragma unroll
+        for (int j = 0; j < ncols1; ++j) {
+            const float2 tmp = __half22float2(mask[j*s31 + KV_max_sj/2 + tid]);
+            all_inf = all_inf && int(isinf(tmp.x)) && int(isinf(tmp.y));
+        }
+
+        all_inf = warp_reduce_all(all_inf);
+        if (tid % WARP_SIZE == 0) {
+            buf_iw[tid / WARP_SIZE] = all_inf;
+        }
+        __syncthreads();
+        all_inf = buf_iw[tid % WARP_SIZE];
+        __syncthreads();
+        all_inf = warp_reduce_all(all_inf);
+
+        if (!all_inf) {
+            KV_max_sj += FATTN_KQ_STRIDE;
+            break;
+        }
+    }
+
+    if (threadIdx.x != 0) {
+        return;
+    }
+
+    KV_max[sequence*ne31 + jt] = KV_max_sj;
+}
+
 template<int D, int ncols1, int ncols2> // D == head size
 __launch_bounds__(D, 1)
 static __global__ void flash_attn_stream_k_fixup(
@@ -711,6 +761,7 @@ void launch_fattn(
 
     ggml_cuda_pool_alloc<half>   K_f16(pool);
     ggml_cuda_pool_alloc<half>   V_f16(pool);
+    ggml_cuda_pool_alloc<int>    KV_max(pool);
     ggml_cuda_pool_alloc<float>  dst_tmp(pool);
     ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
 
@@ -779,11 +830,30 @@ void launch_fattn(
         V_data = (char *) V_f16.ptr;
     }
 
-    int parallel_blocks = 1;
-
     const int ntiles_x = ((Q->ne[1] + ncols1 - 1) / ncols1);
     const int ntiles_total = ntiles_x * (Q->ne[2] / ncols2) * Q->ne[3];
 
+    // Optional optimization where the mask is scanned to determine whether part of the calculation can be skipped.
+    // Only worth the overhead if there is at lease one FATTN_KQ_STRIDE x FATTN_KQ_STRIDE square to be skipped or
+    //     multiple sequences of possibly different lengths.
+    if (mask && (Q->ne[1] >= 1024 || Q->ne[3] > 1)) {
+        const int s31 = mask->nb[1] / sizeof(half2);
+        const int s33 = mask->nb[3] / sizeof(half2);
+
+        const dim3 blocks_num_KV_max(ntiles_x, Q->ne[3], 1);
+        const dim3 block_dim_KV_max(FATTN_KQ_STRIDE/2, 1, 1);
+
+        const int ne_KV_max = blocks_num_KV_max.x*blocks_num_KV_max.y;
+        const int iter_k = K->ne[1] / FATTN_KQ_STRIDE;
+
+        KV_max.alloc(ne_KV_max);
+        flash_attn_mask_to_KV_max<ncols1><<<blocks_num_KV_max, block_dim_KV_max, 0, main_stream>>>
+            ((const half2 *) mask->data, KV_max.ptr, iter_k, s31, s33);
+        CUDA_CHECK(cudaGetLastError());
+    }
+
+    int parallel_blocks = 1;
+
     const dim3 block_dim(warp_size, nwarps, 1);
     int max_blocks_per_sm = 1; // Max. number of active blocks limited by occupancy.
     CUDA_CHECK(cudaOccupancyMaxActiveBlocksPerMultiprocessor(&max_blocks_per_sm, fattn_kernel, block_dim.x * block_dim.y * block_dim.z, nbytes_shared));
@@ -870,6 +940,7 @@ void launch_fattn(
         K_data,
         V_data,
         mask ? ((const char *) mask->data) : nullptr,
+        KV_max.ptr,
         !stream_k && parallel_blocks > 1 ? dst_tmp.ptr : (float *) KQV->data, dst_tmp_meta.ptr,
         scale, max_bias, m0, m1, n_head_log2, logit_softcap,
         Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3], Q->nb[1], Q->nb[2], Q->nb[3],
diff --git a/ggml/src/ggml-cuda/fattn-mma-f16.cuh b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
index 8e847d361b4..a86b95428f6 100644
--- a/ggml/src/ggml-cuda/fattn-mma-f16.cuh
+++ b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
@@ -392,7 +392,8 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_mask(
     }
 }
 
-template<int DKQ, int DV, int ncols1, int ncols2, int nwarps, int ntiles, bool use_logit_softcap, bool mla, bool needs_fixup, bool is_fixup, bool last_iter>
+template<int DKQ, int DV, int ncols1, int ncols2, int nwarps, int ntiles,
+    bool use_logit_softcap, bool mla, bool needs_fixup, bool is_fixup, bool last_iter>
 static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         const float2 * const __restrict__ Q_f2,
         const half2  * const __restrict__ K_h2,
@@ -922,7 +923,8 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
     }
 
     // Iterate over ne11 == previous tokens:
-    for (int kb0 = kb0_start; kb0 < kb0_stop-1; ++kb0) {
+    int kb0 = kb0_start;
+    for (; kb0 < kb0_stop-1; ++kb0) {
         constexpr bool last_iter = false;
         flash_attn_ext_f16_iter<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla, needs_fixup, is_fixup, last_iter>
             (Q_f2, K_h2, V_h2, mask_h2, dstk, dstk_fixup, scale, slope, logit_softcap,
@@ -932,7 +934,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         constexpr bool last_iter = true;
         flash_attn_ext_f16_iter<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla, needs_fixup, is_fixup, last_iter>
             (Q_f2, K_h2, V_h2, mask_h2, dstk, dstk_fixup, scale, slope, logit_softcap,
-             ne01, ne02, stride_K, stride_V, stride_mask, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0_stop-1);
+             ne01, ne02, stride_K, stride_V, stride_mask, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0);
     }
 
     // With multi-stage loading there is no __syncthreads at the end of the iter,
@@ -1204,6 +1206,7 @@ static __global__ void flash_attn_ext_f16(
         const char * __restrict__ K,
         const char * __restrict__ V,
         const char * __restrict__ mask,
+        const int  * __restrict__ KV_max,
         float      * __restrict__ dst,
         float2     * __restrict__ dst_meta,
         const float scale,
@@ -1280,7 +1283,11 @@ static __global__ void flash_attn_ext_f16(
         const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, head, n_head_log2, m0, m1) : 1.0f;
 
         const int kb0_start_kernel = kb0_start * kb_niter;
-        const int kb0_stop_kernel  = kb0_stop  * kb_niter;
+        int       kb0_stop_kernel  = kb0_stop  * kb_niter;
+
+        if (KV_max) {
+            kb0_stop_kernel = min(kb0_stop_kernel, KV_max[sequence*iter_j + jt] / c::nbatch_fa);
+        }
 
         constexpr bool is_fixup = false; // All but (potentially) the last iterations write their data to dst rather than the fixup buffer.
         if (kb0_start == 0) {
@@ -1321,7 +1328,11 @@ static __global__ void flash_attn_ext_f16(
     const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, head, n_head_log2, m0, m1) : 1.0f;
 
     const int kb0_start_kernel = kb0_start * kb_niter;
-    const int kb0_stop_kernel  = kb0_stop  * kb_niter;
+    int       kb0_stop_kernel  = kb0_stop  * kb_niter;
+
+    if (KV_max) {
+        kb0_stop_kernel = min(kb0_stop_kernel, KV_max[sequence*iter_j + jt] / c::nbatch_fa);
+    }
 
     constexpr bool is_fixup = true; // Last index writes its data to fixup buffer to avoid data races with other blocks.
     constexpr bool needs_fixup = false;
diff --git a/ggml/src/ggml-cuda/fattn-tile-f16.cu b/ggml/src/ggml-cuda/fattn-tile-f16.cu
index 678288c13e3..9d0b24ae7ec 100644
--- a/ggml/src/ggml-cuda/fattn-tile-f16.cu
+++ b/ggml/src/ggml-cuda/fattn-tile-f16.cu
@@ -13,6 +13,7 @@ static __global__ void flash_attn_tile_ext_f16(
         const char * __restrict__ K,
         const char * __restrict__ V,
         const char * __restrict__ mask,
+        const int  * __restrict__ KV_max,
         float      * __restrict__ dst,
         float2     * __restrict__ dst_meta,
         const float scale,
@@ -90,7 +91,8 @@ static __global__ void flash_attn_tile_ext_f16(
 
     __syncthreads();
 
-    for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE_TILE_F16; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE_TILE_F16) {
+    const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
+    for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE_TILE_F16; k_VKQ_0 < k_VKQ_max; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE_TILE_F16) {
         // Calculate KQ tile and keep track of new maximum KQ values:
 
         half kqmax_new[ncols/nwarps];
diff --git a/ggml/src/ggml-cuda/fattn-tile-f32.cu b/ggml/src/ggml-cuda/fattn-tile-f32.cu
index bc283d9a7e4..be72f76fb65 100644
--- a/ggml/src/ggml-cuda/fattn-tile-f32.cu
+++ b/ggml/src/ggml-cuda/fattn-tile-f32.cu
@@ -13,6 +13,7 @@ static __global__ void flash_attn_tile_ext_f32(
         const char * __restrict__ K,
         const char * __restrict__ V,
         const char * __restrict__ mask,
+        const int  * __restrict__ KV_max,
         float      * __restrict__ dst,
         float2     * __restrict__ dst_meta,
         const float scale,
@@ -99,7 +100,8 @@ static __global__ void flash_attn_tile_ext_f32(
 
     __syncthreads();
 
-    for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE_TILE_F32; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE_TILE_F32) {
+    const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
+    for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE_TILE_F32; k_VKQ_0 < k_VKQ_max; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE_TILE_F32) {
         // Calculate KQ tile and keep track of new maximum KQ values:
 
         float kqmax_new[ncols/nwarps];
diff --git a/ggml/src/ggml-cuda/fattn-vec-f16.cuh b/ggml/src/ggml-cuda/fattn-vec-f16.cuh
index afef815ceaf..a2df2f66be0 100644
--- a/ggml/src/ggml-cuda/fattn-vec-f16.cuh
+++ b/ggml/src/ggml-cuda/fattn-vec-f16.cuh
@@ -16,6 +16,7 @@ static __global__ void flash_attn_vec_ext_f16(
         const char * __restrict__ K,
         const char * __restrict__ V,
         const char * __restrict__ mask,
+        const int  * __restrict__ KV_max,
         float      * __restrict__ dst,
         float2     * __restrict__ dst_meta,
         const float scale,
@@ -177,10 +178,11 @@ static __global__ void flash_attn_vec_ext_f16(
 
     half2 VKQ[ncols] = {{0.0f, 0.0f}};
 
+    const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
     K     += blockIdx.y*D * nb11;
     V     += blockIdx.y*D * nb21;
     maskh += blockIdx.y*D;
-    for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*D,
+    for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < k_VKQ_max; k_VKQ_0 += gridDim.y*D,
              // Increment pointers after each loop:
              K += gridDim.y*D*nb11, V += gridDim.y*D*nb21, maskh += gridDim.y*D) {
 
@@ -191,29 +193,7 @@ static __global__ void flash_attn_vec_ext_f16(
             for (int j = 0; j < ncols; ++j) {
                 maskh_shared[j*D + tid] = slopeh*maskh[j*ne11 + tid];
             }
-
             __syncthreads();
-
-            // When using multiple parallel sequences in llama.cpp, some KV slices can be fully masked out.
-            // In such cases, skip the KV slice.
-            // On AMD __all_sync would not work correctly because it assumes a warp size of 64.
-#ifndef GGML_USE_HIP
-            bool skip = true;
-#pragma unroll
-            for (int j = 0; j < ncols; ++j) {
-#pragma unroll
-                for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
-                    const int i = i0 + threadIdx.x;
-
-                    const float2 tmp = __half22float2(((const half2 *) maskh_shared)[j*(D/2) + i]);
-                    skip = skip && isinf(tmp.x) && isinf(tmp.y);
-                }
-            }
-            if (__all_sync(0xFFFFFFFF, skip)) {
-                __syncthreads();
-                continue;
-            }
-#endif // GGML_USE_HIP
         }
 
         // For unknown reasons using a half array of size 1 for kqmax_new causes a performance regression,
diff --git a/ggml/src/ggml-cuda/fattn-vec-f32.cuh b/ggml/src/ggml-cuda/fattn-vec-f32.cuh
index 3595e29693a..9ab0fc133b7 100644
--- a/ggml/src/ggml-cuda/fattn-vec-f32.cuh
+++ b/ggml/src/ggml-cuda/fattn-vec-f32.cuh
@@ -16,6 +16,7 @@ static __global__ void flash_attn_vec_ext_f32(
         const char * __restrict__ K,
         const char * __restrict__ V,
         const char * __restrict__ mask,
+        const int  * __restrict__ KV_max,
         float      * __restrict__ dst,
         float2     * __restrict__ dst_meta,
         const float scale,
@@ -183,10 +184,11 @@ static __global__ void flash_attn_vec_ext_f32(
 
     float VKQ[ncols] = {0.0f};
 
+    const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
     K     += blockIdx.y*D * nb11;
     V     += blockIdx.y*D * nb21;
     maskh += blockIdx.y*D;
-    for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*D,
+    for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < k_VKQ_max; k_VKQ_0 += gridDim.y*D,
              // Increment pointers after each loop:
              K += gridDim.y*D*nb11, V += gridDim.y*D*nb21, maskh += gridDim.y*D) {
 
@@ -197,28 +199,7 @@ static __global__ void flash_attn_vec_ext_f32(
             for (int j = 0; j < ncols; ++j) {
                 maskf_shared[j*D + tid] = slope*__half2float(maskh[j*ne11 + tid]);
             }
-
             __syncthreads();
-
-            // When using multiple parallel sequences in llama.cpp, some KV slices can be fully masked out.
-            // In such cases, skip the KV slice.
-            // On AMD __all_sync would not work correctly because it assumes a warp size of 64.
-#ifndef GGML_USE_HIP
-            bool skip = true;
-#pragma unroll
-            for (int j = 0; j < ncols; ++j) {
-#pragma unroll
-                for (int i0 = 0; i0 < D; i0 += WARP_SIZE) {
-                    const int i = i0 + threadIdx.x;
-
-                    skip = skip && isinf(maskf_shared[j*D + i]);
-                }
-            }
-            if (__all_sync(0xFFFFFFFF, skip)) {
-                __syncthreads();
-                continue;
-            }
-#endif // GGML_USE_HIP
         }
 
         float kqmax_new_arr[ncols];
diff --git a/ggml/src/ggml-cuda/fattn-wmma-f16.cu b/ggml/src/ggml-cuda/fattn-wmma-f16.cu
index f4393ec5711..40554204b62 100644
--- a/ggml/src/ggml-cuda/fattn-wmma-f16.cu
+++ b/ggml/src/ggml-cuda/fattn-wmma-f16.cu
@@ -29,6 +29,7 @@ static __global__ void flash_attn_ext_f16(
         const char * __restrict__ K,
         const char * __restrict__ V,
         const char * __restrict__ mask,
+        const int  * __restrict__ KV_max,
         float      * __restrict__ dst,
         float2     * __restrict__ dst_meta,
         const float scale,
@@ -165,7 +166,8 @@ static __global__ void flash_attn_ext_f16(
     __syncthreads();
 
     // Iterate over ne11 == previous tokens:
-    for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE) {
+    const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
+    for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE; k_VKQ_0 < k_VKQ_max; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE) {
         // Calculate tile of KQ:
 #pragma unroll
         for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE; i_KQ_0 += KQ_stride_tc) {
diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu
index d9f1613051d..a51136f6b8a 100644
--- a/ggml/src/ggml-cuda/fattn.cu
+++ b/ggml/src/ggml-cuda/fattn.cu
@@ -315,7 +315,8 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
 
     const bool gqa_opt_applies = ((Q->ne[2] / K->ne[2]) % 2 == 0) && mask; // The mma-based kernels have GQA-specific optimizations
     const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16;
-    const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies && cc < GGML_CUDA_CC_ADA_LOVELACE && !mma_needs_data_conversion;
+    const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies &&
+        (Q->ne[3] > 1 || cc < GGML_CUDA_CC_ADA_LOVELACE) && !mma_needs_data_conversion;
     const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % (2*warp_size) == 0;
     if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1) {
         if (prec == GGML_PREC_DEFAULT) {

From 7ec71a0d17d4bba6a53e801b54e9d197ebddb6fc Mon Sep 17 00:00:00 2001
From: uvos <carl@uvos.xyz>
Date: Wed, 30 Jul 2025 17:38:06 +0200
Subject: [PATCH 15/86] HIP: enable mfma mmq on gfx908 and gfx90a for select
 datatypes and shapes (llama/14949)

---
 ggml/src/ggml-cuda/common.cuh |  7 +++----
 ggml/src/ggml-cuda/mmq.cu     | 24 ++++++++++++++++++++----
 ggml/src/ggml-cuda/mmq.cuh    |  4 ++--
 3 files changed, 25 insertions(+), 10 deletions(-)

diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
index 4560d85c4cf..7fb04d51b77 100644
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -227,9 +227,9 @@ typedef float2 dfloat2;
 #define FP16_MMA_AVAILABLE
 #endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4)))
 
-#if defined(GGML_USE_HIP) && defined(CDNA3) && !defined(GGML_HIP_NO_MMQ_MFMA)
+#if defined(GGML_USE_HIP) && defined(CDNA) && !defined(GGML_HIP_NO_MMQ_MFMA)
 #define AMD_MFMA_AVAILABLE
-#endif // defined(GGML_USE_HIP) && defined(CDNA3) && !defined(GGML_HIP_NO_MMQ_MFMA)
+#endif // defined(GGML_USE_HIP) && defined(CDNA) && !defined(GGML_HIP_NO_MMQ_MFMA)
 
 #if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
 #define NEW_MMA_AVAILABLE
@@ -293,10 +293,9 @@ static bool fp32_mma_hardware_available(const int cc) {
     return GGML_CUDA_CC_IS_CDNA(cc);
 }
 
-// AMD CDNA3 matrix cores.. Will add support for other CDNA generations later.
 static bool amd_mfma_available(const int cc) {
 #if !defined(GGML_HIP_NO_MMQ_MFMA)
-    return GGML_CUDA_CC_IS_CDNA3(cc);
+    return GGML_CUDA_CC_IS_CDNA(cc);
 #else
     return false;
 #endif //!defined(GGML_HIP_NO_MMQ_MFMA)
diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cu
index e2fd0c1c254..d4954fbe69e 100644
--- a/ggml/src/ggml-cuda/mmq.cu
+++ b/ggml/src/ggml-cuda/mmq.cu
@@ -109,8 +109,8 @@ void ggml_cuda_mul_mat_q(
     const int64_t s03 = src0->nb[3] / ts_src0;
     const int64_t s3  =  dst->nb[3] / ts_dst;
 
-    const bool use_stream_k = ((GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA)
-                            || (GGML_CUDA_CC_IS_AMD(cc) && GGML_CUDA_CC_IS_CDNA3(cc)));
+    const bool use_stream_k = (GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA)
+                            || GGML_CUDA_CC_IS_CDNA(cc);
 
     if (!ids) {
         const size_t nbytes_src1_q8_1 = ne13*ne12 * ne11*ne10_padded * sizeof(block_q8_1)/QK8_1 +
@@ -252,7 +252,7 @@ void ggml_cuda_op_mul_mat_q(
     // Also its fixup needs to allocate a temporary buffer in the memory pool.
     // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
     const bool use_stream_k = ((GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA)
-                            || (GGML_CUDA_CC_IS_AMD(cc) && GGML_CUDA_CC_IS_CDNA3(cc)))
+                            || GGML_CUDA_CC_IS_CDNA(cc))
                             && src1_ncols == ne11;
     const mmq_args args = {
         src0_dd_i, src0->type, (const int *) src1_ddq_i, nullptr, nullptr, dst_dd_i,
@@ -306,7 +306,7 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
         return false;
     }
 
-    if (new_mma_available(cc) || amd_mfma_available(cc)) {
+    if (new_mma_available(cc)) {
         return true;
     }
 
@@ -322,5 +322,21 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
         return !fp16_mma_hardware_available(cc) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
     }
 
+    if (amd_mfma_available(cc)) {
+        // As of ROCM 7.0 rocblas/tensile performs very poorly on CDNA3 and hipblaslt (via ROCBLAS_USE_HIPBLASLT)
+        // performs better but is currently suffering from a crash on this architecture.
+        // TODO: Revisit when hipblaslt is fixed on CDNA3
+        if (GGML_CUDA_CC_IS_CDNA3(cc)) {
+            return true;
+        }
+        if (ne11 <= 128 || type == GGML_TYPE_Q4_0 || type == GGML_TYPE_Q4_1 || type == GGML_TYPE_Q5_0 || type == GGML_TYPE_Q5_1) {
+            return true;
+        }
+        if (ne11 <= 256 && (type == GGML_TYPE_Q4_K || type == GGML_TYPE_Q5_K)) {
+            return true;
+        }
+        return false;
+    }
+
     return (!GGML_CUDA_CC_IS_RDNA4(cc) && !GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
 }
diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh
index d8650dd70b9..dd60529faa4 100644
--- a/ggml/src/ggml-cuda/mmq.cuh
+++ b/ggml/src/ggml-cuda/mmq.cuh
@@ -3096,8 +3096,8 @@ static __global__ void mul_mat_q(
     }
     __syncthreads();
 
-    // On AMD or old CUDA the performance with stream-k was worse, use conventional tiling instead:
-#if (defined(GGML_USE_HIP) && !defined(CDNA3)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA
+    // On non-CDNA AMD or old CUDA the performance with stream-k was worse, use conventional tiling instead:
+#if (defined(GGML_USE_HIP) && !defined(CDNA)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA
     {
         const int wt = blockIdx.z / nchannels_y;
         const int zt = blockIdx.z - wt*nchannels_y;

From cdfefa3d0722b704a7d30fa244132f84eaf9ec75 Mon Sep 17 00:00:00 2001
From: lhez <lih@qti.qualcomm.com>
Date: Wed, 30 Jul 2025 14:56:55 -0700
Subject: [PATCH 16/86] opencl: add `mul_mat_f32_f32_l4_lm` and
 `mul_mat_f16_f32_l4_lm` (llama/14809)

---
 ggml/src/ggml-opencl/CMakeLists.txt           |   2 +
 ggml/src/ggml-opencl/ggml-opencl.cpp          | 144 ++++++++++++++++--
 .../kernels/mul_mm_f16_f32_l4_lm.cl           | 132 ++++++++++++++++
 .../kernels/mul_mm_f32_f32_l4_lm.cl           | 133 ++++++++++++++++
 4 files changed, 399 insertions(+), 12 deletions(-)
 create mode 100644 ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl
 create mode 100644 ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl

diff --git a/ggml/src/ggml-opencl/CMakeLists.txt b/ggml/src/ggml-opencl/CMakeLists.txt
index 015fa8f0682..3adea836154 100644
--- a/ggml/src/ggml-opencl/CMakeLists.txt
+++ b/ggml/src/ggml-opencl/CMakeLists.txt
@@ -82,6 +82,8 @@ set(GGML_OPENCL_KERNELS
     mul_mv_q4_0_f32_1d_16x_flat
     mul_mv_q6_k
     mul_mv_id_q4_0_f32_8x_flat
+    mul_mm_f32_f32_l4_lm
+    mul_mm_f16_f32_l4_lm
     mul
     norm
     relu
diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
index c87a32383c8..984d35a2ecf 100644
--- a/ggml/src/ggml-opencl/ggml-opencl.cpp
+++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
@@ -33,6 +33,7 @@
 #undef MAX
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
+#define CEIL_DIV(M, N) (((M) + (N)-1) / (N))
 
 #define UNUSED(x) (void)(x)
 
@@ -396,6 +397,8 @@ struct ggml_backend_opencl_context {
     cl_program program_conv_2d_f16_f32;
     cl_program program_tsembd;
     cl_program program_mul_mv_id_q4_0_f32_8x_flat;
+    cl_program program_mul_mm_f32_f32_l4_lm;
+    cl_program program_mul_mm_f16_f32_l4_lm;
 
     cl_kernel kernel_add, kernel_add_row;
     cl_kernel kernel_mul, kernel_mul_row;
@@ -450,6 +453,8 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_conv_2d_f16_f32;
     cl_kernel kernel_timestep_embedding;
     cl_kernel kernel_mul_mv_id_q4_0_f32_8x_flat;
+    cl_kernel kernel_mul_mm_f32_f32_l4_lm;
+    cl_kernel kernel_mul_mm_f16_f32_l4_lm;
 
     std::vector<ProfilingInfo> profiling_info;
 
@@ -1040,6 +1045,38 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // mul_mm_f32_f32_l4_lm
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mm_f32_f32_l4_lm.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mm_f32_f32_l4_lm.cl");
+#endif
+        backend_ctx->program_mul_mm_f32_f32_l4_lm =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mm_f32_f32_l4_lm = clCreateKernel(backend_ctx->program_mul_mm_f32_f32_l4_lm, "kernel_mul_mm_f32_f32_l4_lm", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mm_f16_f32_l4_lm
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mm_f16_f32_l4_lm.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mm_f16_f32_l4_lm.cl");
+#endif
+        backend_ctx->program_mul_mm_f16_f32_l4_lm =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mm_f16_f32_l4_lm = clCreateKernel(backend_ctx->program_mul_mm_f16_f32_l4_lm, "kernel_mul_mm_f16_f32_l4_lm", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
     // mul
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -5297,18 +5334,6 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
 
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
-     if (src0t == GGML_TYPE_F16 && src1t == GGML_TYPE_F32 &&
-        src0->ne[1] > 32 &&   // M > 32
-        src1->ne[1] > 32 &&   // N > 32
-        src0->ne[0] > 32 &&   // K > 32
-        src0->ne[2] == 1 && src0->ne[3] == 1 &&
-        src1->ne[2] == 1 && src1->ne[3] == 1 &&
-        ggml_is_contiguous(src0) && ggml_is_contiguous(src1) &&
-        backend_ctx->kernel_mul_mat_f16_f32_tiled != NULL) {
-        ggml_cl_mul_mat_f16_f32_tiled(backend, src0, src1, dst);
-        return;
-    }
-
     ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
     ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
     ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
@@ -5655,6 +5680,101 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
     } // if (ne01 && ne1)
 #endif // GGML_OPENCL_USE_ADRENO_KERNELS
 
+    // GEMM using local memory
+    // Current BK = 16, so ne00 % 16 == 0
+    if (ggml_is_contiguous(src0) &&
+        ggml_is_contiguous(src1) &&
+        src1t == GGML_TYPE_F32 &&
+        ne00 % 16 == 0 &&
+        ne11 > 1) {
+        switch(src0t) {
+            case GGML_TYPE_F32: {
+                kernel = backend_ctx->kernel_mul_mm_f32_f32_l4_lm;
+                nth0 = 128; // calculated as (BM*BN)/(TM*TN)
+
+                int batch_stride_a = ne00*ne01;
+                int batch_stride_b = ne10*ne11;
+                int batch_stride_d = ne0*ne1;
+
+                CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+                CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+                CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+                CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+                CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+                CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+                CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne11));
+                CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+                CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne10)); // stride_a
+                CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne10)); // stride_b
+                CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne01)); // stride_d
+                CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &batch_stride_a));
+                CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &batch_stride_b));
+                CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &batch_stride_d));
+                CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r2));
+                CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &r3));
+
+                // 64 is block tile size BM and BN - change here when BM and BN in the kernel are changed.
+                size_t global_work_size[] = {(size_t)(CEIL_DIV(ne01, 64)*nth0), (size_t)(CEIL_DIV(ne11, 64)), (size_t)ne12*ne13};
+                size_t local_work_size[] = {(size_t)nth0, 1, 1};
+
+                backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+                return;
+            }
+            case GGML_TYPE_F16: {
+                kernel = backend_ctx->kernel_mul_mm_f16_f32_l4_lm;
+                nth0 = 128; // calculated as (BM*BN)/(TM*TN)
+
+                int batch_stride_a = ne00*ne01;
+                int batch_stride_b = ne10*ne11;
+                int batch_stride_d = ne0*ne1;
+
+                CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+                CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+                CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+                CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+                CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+                CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+                CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne11));
+                CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+                CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne10)); // stride_a
+                CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne10)); // stride_b
+                CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne01)); // stride_d
+                CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &batch_stride_a));
+                CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &batch_stride_b));
+                CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &batch_stride_d));
+                CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r2));
+                CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &r3));
+
+                // 64 is block tile size BM and BN - change here when BM and BN in the kernel are changed.
+                size_t global_work_size[] = {(size_t)(CEIL_DIV(ne01, 64)*nth0), (size_t)(CEIL_DIV(ne11, 64)), (size_t)ne12*ne13};
+                size_t local_work_size[] = {(size_t)nth0, 1, 1};
+
+                backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+                return;
+            }
+            default:
+                break;
+        }
+    }
+
+    if (src0t == GGML_TYPE_F16 && src1t == GGML_TYPE_F32 &&
+        src0->ne[1] > 32 &&   // M > 32
+        src1->ne[1] > 32 &&   // N > 32
+        src0->ne[0] > 32 &&   // K > 32
+        src0->ne[2] == 1 && src0->ne[3] == 1 &&
+        src1->ne[2] == 1 && src1->ne[3] == 1 &&
+        ggml_is_contiguous(src0) && ggml_is_contiguous(src1) &&
+        backend_ctx->kernel_mul_mat_f16_f32_tiled != NULL) {
+        ggml_cl_mul_mat_f16_f32_tiled(backend, src0, src1, dst);
+        return;
+    }
+
     if (!ggml_is_transposed(src0) &&
         !ggml_is_transposed(src1) &&
         src1t == GGML_TYPE_F32 &&
diff --git a/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl b/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl
new file mode 100644
index 00000000000..9599a0e1572
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl
@@ -0,0 +1,132 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define LOAD_VEC_A 4
+#define LOAD_VEC_B 4
+
+#define BM 64
+#define BN 64
+#define BK 16
+#define TM 4
+#define TN 8
+
+kernel void kernel_mul_mm_f16_f32_l4_lm(
+    global half4 * src0,
+    ulong offset0,
+    global float4 * src1,
+    ulong offset1,
+    global float * dst,
+    ulong offsetd,
+
+    int ne00,
+    int ne01,
+    int ne02,
+    int ne11,
+    int ne12,
+
+    int stride_a,
+    int stride_b,
+    int stride_d,
+
+    int batch_stride_a,
+    int batch_stride_b,
+    int batch_stride_d,
+
+    int r2,
+    int r3
+) {
+    src0 = (global half4*)((global char*)src0 + offset0);
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    local half  buf_a[BM * BK];
+    local float buf_b[BN * BK];
+
+    const int batch_idx = get_global_id(2);
+
+    const int i13 = batch_idx / ne12;
+    const int i12 = batch_idx % ne12;
+
+    const int i03 = i13 / r3;
+    const int i02 = i12 / r2;
+
+    const int batch_idx_a = i03 * ne02 + i02;
+
+    const int ir = get_group_id(0);
+    const int ic = get_group_id(1);
+
+    const int tid = get_local_id(0);
+    const int th_r  = tid % (BM / TM);
+    const int th_c  = tid / (BM / TM);
+
+    const int loadr_a = get_local_id(0) % (BK / LOAD_VEC_A);
+    const int loadc_a = get_local_id(0) / (BK / LOAD_VEC_A);
+    const int loadr_b = get_local_id(0) % (BK / LOAD_VEC_B);
+    const int loadc_b = get_local_id(0) / (BK / LOAD_VEC_B);
+
+    const int loadstride_a = get_local_size(0) * LOAD_VEC_A / BK;
+    const int loadstride_b = get_local_size(0) * LOAD_VEC_B / BK;
+
+    int pos_a = (batch_idx_a * batch_stride_a + ir * BM * stride_a) / LOAD_VEC_A;
+    int pos_b = (batch_idx   * batch_stride_b + ic * BN * stride_b) / LOAD_VEC_B;
+
+    float sums[TM * TN];
+    half  cache_a[TM];
+    float cache_b[TN];
+
+    for (int i = 0; i < TM * TN; i++) {
+        sums[i] = 0.0f;
+    }
+
+    for (int block = 0; block < ne00; block += BK) {
+        for (int l = 0; l < BM; l += loadstride_a) {
+            const int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a;
+            buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = src0[idx].s0;
+            buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = src0[idx].s1;
+            buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = src0[idx].s2;
+            buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = src0[idx].s3;
+        }
+
+        for (int l = 0; l < BN; l += loadstride_b) {
+            const int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
+            buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
+            buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;
+            buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = src1[idx].s2;
+            buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = src1[idx].s3;
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        pos_a += BK / LOAD_VEC_A;
+        pos_b += BK / LOAD_VEC_B;
+
+        for (int i = 0; i < BK; i++) {
+            for (int j = 0; j < TM; j++) {
+                cache_a[j] = buf_a[(i) * BM + th_r * TM + j];
+            }
+            for (int j = 0; j < TN; j++) {
+                cache_b[j] = buf_b[(i) * BN + th_c * TN + j];
+            }
+
+            for (int cc = 0; cc < TN; cc++) {
+                for (int cr = 0; cr < TM; cr++) {
+                    const int sums_idx = cc*TM + cr;
+                    sums[sums_idx] = mad(convert_float(cache_a[cr]), cache_b[cc], sums[sums_idx]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const int dr = ir * BM + th_r * TM;
+    const int dc = ic * BN + th_c * TN;
+
+    const int offsets = batch_idx * batch_stride_d;
+
+    for (int cc = 0; cc < TN; cc++) {
+        for (int cr = 0; cr < TM; cr++) {
+            if (dr + cr < ne01 && dc + cc < ne11) {
+                dst[offsets + (dc + cc) * stride_d + dr + cr] = sums[cc * TM + cr];
+            }
+        }
+    }
+}
diff --git a/ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl b/ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl
new file mode 100644
index 00000000000..58c5178e39c
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl
@@ -0,0 +1,133 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define LOAD_VEC_A 4
+#define LOAD_VEC_B 4
+
+#define BM 64
+#define BN 64
+#define BK 16
+#define TM 4
+#define TN 8
+
+kernel void kernel_mul_mm_f32_f32_l4_lm(
+    global float4 * src0,
+    ulong offset0,
+    global float4 * src1,
+    ulong offset1,
+    global float * dst,
+    ulong offsetd,
+
+    int ne00,
+    int ne01,
+    int ne02,
+    int ne11,
+    int ne12,
+
+    int stride_a,
+    int stride_b,
+    int stride_d,
+
+    int batch_stride_a,
+    int batch_stride_b,
+    int batch_stride_d,
+
+    int r2,
+    int r3
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    local float buf_a[BM * BK];
+    local float buf_b[BN * BK];
+
+    const int batch_idx = get_global_id(2);
+
+    const int i13 = batch_idx / ne12;
+    const int i12 = batch_idx % ne12;
+
+    const int i03 = i13 / r3;
+    const int i02 = i12 / r2;
+
+    const int batch_idx_a = i03 * ne02 + i02;
+
+    const int ir = get_group_id(0);
+    const int ic = get_group_id(1);
+
+    const int tid = get_local_id(0);
+    const int th_r  = tid % (BM / TM);
+    const int th_c  = tid / (BM / TM);
+
+    const int loadr_a = get_local_id(0) % (BK / LOAD_VEC_A);
+    const int loadc_a = get_local_id(0) / (BK / LOAD_VEC_A);
+    const int loadr_b = get_local_id(0) % (BK / LOAD_VEC_B);
+    const int loadc_b = get_local_id(0) / (BK / LOAD_VEC_B);
+
+    const int loadstride_a = get_local_size(0) * LOAD_VEC_A / BK;
+    const int loadstride_b = get_local_size(0) * LOAD_VEC_B / BK;
+
+    int pos_a = (batch_idx_a * batch_stride_a + ir * BM * stride_a) / LOAD_VEC_A;
+    int pos_b = (batch_idx   * batch_stride_b + ic * BN * stride_b) / LOAD_VEC_B;
+
+    float sums[TM * TN];
+    float cache_a[TM];
+    float cache_b[TN];
+
+    for (int i = 0; i < TM * TN; i++) {
+        sums[i] = 0.0f;
+    }
+
+    for (int block = 0; block < ne00; block += BK) {
+        for (int l = 0; l < BM; l += loadstride_a) {
+            const int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a;
+            buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = src0[idx].s0;
+            buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = src0[idx].s1;
+            buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = src0[idx].s2;
+            buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = src0[idx].s3;
+        }
+
+        for (int l = 0; l < BN; l += loadstride_b) {
+            const int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
+            buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
+            buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;
+            buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = src1[idx].s2;
+            buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = src1[idx].s3;
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        pos_a += BK / LOAD_VEC_A;
+        pos_b += BK / LOAD_VEC_B;
+
+        for (int i = 0; i < BK; i++) {
+            for (int j = 0; j < TM; j++) {
+                cache_a[j] = buf_a[(i) * BM + th_r * TM + j];
+            }
+
+            for (int j = 0; j < TN; j++) {
+                cache_b[j] = buf_b[(i) * BN + th_c * TN + j];
+            }
+
+            for (int cc = 0; cc < TN; cc++) {
+                for (int cr = 0; cr < TM; cr++) {
+                    const int sums_idx = cc*TM + cr;
+                    sums[sums_idx] = mad(cache_a[cr], cache_b[cc], sums[sums_idx]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const int dr = ir * BM + th_r * TM;
+    const int dc = ic * BN + th_c * TN;
+
+    const int offsets = batch_idx * batch_stride_d;
+
+    for (int cc = 0; cc < TN; cc++) {
+        for (int cr = 0; cr < TM; cr++) {
+            if (dr + cr < ne01 && dc + cc < ne11) {
+                dst[offsets + (dc + cc) * stride_d + dr + cr] = sums[cc * TM + cr];
+            }
+        }
+    }
+}

From f9cba92fd0494d85e2bd609ecb47fa883f3c561c Mon Sep 17 00:00:00 2001
From: hipudding <huafengchun@gmail.com>
Date: Thu, 31 Jul 2025 19:47:20 +0800
Subject: [PATCH 17/86] CANN: Improve loading efficiency after converting
 weights to NZ format. (llama/14985)

* CANN: Improve loading efficiency after converting weights to NZ format.

* CANN: fix typo
---
 ggml/src/ggml-cann/aclnn_ops.cpp |   8 +--
 ggml/src/ggml-cann/ggml-cann.cpp | 114 +++++++++++++++++--------------
 2 files changed, 66 insertions(+), 56 deletions(-)

diff --git a/ggml/src/ggml-cann/aclnn_ops.cpp b/ggml/src/ggml-cann/aclnn_ops.cpp
index 316055193d8..07d6b8b67d4 100755
--- a/ggml/src/ggml-cann/aclnn_ops.cpp
+++ b/ggml/src/ggml-cann/aclnn_ops.cpp
@@ -1913,11 +1913,9 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx,
                              bcast_weight_nb[4], bcast_weight_nb[5]};
     aclTensor* acl_weight_tensor;
 
-    bool weightToNZ = false;
-#ifdef ASCEND_310P
-    weightToNZ = (getenv("GGML_CANN_WEIGHT_NZ") != nullptr);
-#endif
-    if (weightToNZ && is_matmul_weight(weight)) {
+    // Only check env once.
+    static bool weight_to_nz = parse_bool(get_env("GGML_CANN_WEIGHT_NZ").value_or(""));
+    if (weight_to_nz && is_matmul_weight(weight)) {
         int64_t acl_stride[2] = {1, transpose_ne[1]};
 
         // Reverse ne.
diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
index 4dac2e8b789..49f55891d85 100755
--- a/ggml/src/ggml-cann/ggml-cann.cpp
+++ b/ggml/src/ggml-cann/ggml-cann.cpp
@@ -1116,61 +1116,59 @@ static enum ggml_status ggml_backend_cann_buffer_init_tensor(
     return GGML_STATUS_SUCCESS;
 }
 
-static int CreateAclTensorWeight(const void *hostData, const std::vector<int64_t> &shape, void **deviceAddr,
-                      aclDataType dataType, aclTensor **tensor)
-{
-    uint64_t size = 1;
-    for (auto i : shape) {
-        size *= i;
+// ND to NZ Workspace Cache Management. Thread-safety: Not guaranteed
+namespace {
+    void* g_nz_workspace = nullptr;
+    size_t g_nz_workspace_allocated = 0;
+
+    void release_nz_workspace() {
+        if (g_nz_workspace) {
+            aclrtFree(g_nz_workspace);
+            g_nz_workspace = nullptr;
+            g_nz_workspace_allocated = 0;
+        }
     }
 
-    const aclIntArray *mat2Size = aclCreateIntArray(shape.data(), shape.size());
-    ACL_CHECK(aclnnCalculateMatmulWeightSizeV2(mat2Size, dataType, &size));
-
-    size *= sizeof(int16_t);
-
-    ACL_CHECK(aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST));
-    aclrtMemcpy(*deviceAddr, size, hostData, size, ACL_MEMCPY_HOST_TO_DEVICE);
-
-    std::vector<int64_t> strides(shape.size(), 1);
-    for (int64_t i = shape.size() - 2; i >= 0; i--) {
-        strides[i] = shape[i + 1] * strides[i + 1];
+    void relloc_nz_workspace(size_t new_size) {
+        if (new_size > g_nz_workspace_allocated) {
+        if (g_nz_workspace) {
+            aclrtFree(g_nz_workspace);
+            g_nz_workspace = nullptr;
+        }
+        ACL_CHECK(aclrtMalloc(&g_nz_workspace, new_size, ACL_MEM_MALLOC_HUGE_FIRST));
+        g_nz_workspace_allocated = new_size;
+    }
     }
-
-    *tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND,
-                              shape.data(), shape.size(), *deviceAddr);
-    return 0;
 }
 
+/**
+ * @brief Convert tensor weights to NZ format using Ascend CANN API.
+ *
+ * This function creates a transposed tensor descriptor and performs the
+ * TransMatmulWeight operation. Converting tensor formats can significantly
+ * improve performance on certain hardware.
+ *
+ * @param tensor Pointer to the input ggml_tensor containing the weights.
+ * @param data Pointer to the raw data buffer for the tensor weights.
+ * @param offset Byte offset within the tensor data buffer where weights start.
+ *
+ * @note The workspace buffer used in this function is managed globally and reused
+ *       across calls. This reduces overhead from repeated memory allocation and deallocation.
+ */
 static void weight_format_to_nz(ggml_tensor *tensor, const void *data, size_t offset) {
-    aclrtStream stream;
-    ACL_CHECK(aclrtCreateStream(&stream));
-
-    std::vector<int64_t> weightTransposedShape = {tensor->ne[1], tensor->ne[0]};
-    void *weightTransposedDeviceAddr = nullptr;
-    aclTensor *weightTransposed = nullptr;
-    CreateAclTensorWeight(data, weightTransposedShape, &weightTransposedDeviceAddr,
-                          ggml_cann_type_mapping(tensor->type), &weightTransposed);
-
+    aclTensor* weightTransposed = ggml_cann_create_tensor(tensor, tensor->ne,
+                                    tensor->nb, 2, ACL_FORMAT_ND, offset);
     uint64_t workspaceSize = 0;
     aclOpExecutor *executor;
-    void *workspaceAddr = nullptr;
 
     // TransMatmulWeight
-    ACL_CHECK(aclnnTransMatmulWeightGetWorkspaceSize(weightTransposed, &workspaceSize, &executor));
-    std::unique_ptr<void, aclError (*)(void *)> workspaceAddrPtrTrans(nullptr, aclrtFree);
-    if (workspaceSize > 0) {
-        ACL_CHECK(aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST));
-        workspaceAddrPtrTrans.reset(workspaceAddr);
-    }
-    ACL_CHECK(aclnnTransMatmulWeight(workspaceAddr, workspaceSize, executor, stream));
+    ACL_CHECK(aclnnTransMatmulWeightGetWorkspaceSize(weightTransposed,
+                                                    &workspaceSize, &executor));
+    // Avoid frequent malloc/free of the workspace.
+    relloc_nz_workspace(workspaceSize);
 
-    size_t size = ggml_nelements(tensor) * ggml_element_size(tensor);
-
-    aclrtMemcpy((char *)tensor->data + offset, size,
-                weightTransposedDeviceAddr, size, ACL_MEMCPY_HOST_TO_DEVICE);
+    ACL_CHECK(aclnnTransMatmulWeight(g_nz_workspace, workspaceSize, executor, nullptr));
     ACL_CHECK(aclDestroyTensor(weightTransposed));
-    aclrtFree(weightTransposedDeviceAddr);
 }
 
 // TODO: need handle tensor which has paddings.
@@ -1197,14 +1195,14 @@ static void ggml_backend_cann_buffer_set_tensor(
     // For acl, synchronous functions use this default stream.
     // Why aclrtSynchronizeDevice?
 
-    bool weightToNZ = false;
-#ifdef ASCEND_310P
-    weightToNZ = (getenv("GGML_CANN_WEIGHT_NZ") != nullptr);
-#endif
+    // Only check env once.
+    static bool weight_to_nz = parse_bool(get_env("GGML_CANN_WEIGHT_NZ").value_or(""));
     if (!need_transform(tensor->type)) {
         ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size, data, size,
                               ACL_MEMCPY_HOST_TO_DEVICE));
-        if (weightToNZ && is_matmul_weight((const ggml_tensor*)tensor)) {
+        if (weight_to_nz && is_matmul_weight((const ggml_tensor*)tensor)) {
+            GGML_ASSERT(tensor->ne[2] == 1);
+            GGML_ASSERT(tensor->ne[3] == 1);
             weight_format_to_nz(tensor, data, offset);
         }
     } else {
@@ -1440,20 +1438,32 @@ static size_t ggml_backend_cann_buffer_type_get_alloc_size(
     size_t size = ggml_nbytes(tensor);
     int64_t ne0 = tensor->ne[0];
 
+    // Only check env once.
+    static bool weight_to_nz = parse_bool(get_env("GGML_CANN_WEIGHT_NZ").value_or(""));
+
     // last line must bigger than 32, because every single op deal at
     // least 32 bytes.
     // TODO: quantized type?
     // int64_t line_size = ne0 * ggml_element_size(tensor);
     // int64_t line_size_align_32 = (line_size + 31) & ~31;
     // size += (line_size_align_32 - line_size);
-
-    // TODO: not support quantized yet.
-    // TODO: consider un-continue tensor.
     if (ggml_is_quantized(tensor->type)) {
         if (ne0 % MATRIX_ROW_PADDING != 0) {
             size += ggml_row_size(
                 tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
         }
+    } else if (weight_to_nz && is_matmul_weight((const ggml_tensor*)tensor)) {
+        // NZ format weight are not support quantized yet.
+        // If ND tensor transform to NZ, size may changed.
+        int64_t shape[] = {tensor->ne[1], tensor->ne[0]};
+        GGML_ASSERT(tensor->ne[2] == 1);
+        GGML_ASSERT(tensor->ne[3] == 1);
+        const aclIntArray *acl_shape = aclCreateIntArray(shape, 2);
+        size_t new_size;
+        ACL_CHECK(aclnnCalculateMatmulWeightSizeV2(acl_shape,
+                    ggml_cann_type_mapping(tensor->type), &new_size));
+        ACL_CHECK(aclDestroyIntArray(acl_shape));
+        size = std::max(size, new_size);
     }
 
     return size;
@@ -2080,6 +2090,8 @@ static enum ggml_status ggml_backend_cann_graph_compute(
         (ggml_backend_cann_context*)backend->context;
 
     ggml_cann_set_device(cann_ctx->device);
+    //release temp buffer create by set tensor.
+    release_nz_workspace();
 
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_tensor* node = cgraph->nodes[i];

From 168581782f8adf91835ba2223ce67f4b6c5e5c97 Mon Sep 17 00:00:00 2001
From: Ruben Ortlam <picard12@live.de>
Date: Thu, 31 Jul 2025 17:46:54 +0200
Subject: [PATCH 18/86] Vulkan: Fix minor debug mode issues (llama/14899)

* vulkan: fix debug mode issues

* vulkan: remove broken check_results GGML_OP_SET_ROWS support
---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp | 17 +++++++++--------
 1 file changed, 9 insertions(+), 8 deletions(-)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index b1a1cff904b..75b58c26fc1 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -5225,9 +5225,9 @@ static void ggml_vk_quantize_q8_1(ggml_backend_vk_context * ctx, vk_context& sub
 }
 
 static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
-    VK_LOG_DEBUG("ggml_vk_mul_mat_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
-    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
-    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
+    VK_LOG_DEBUG("ggml_vk_mul_mat_q_f16((" << src0 << ", name=" << src0->name << ", type=" << ggml_type_name(src0->type) << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
+    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << ggml_type_name(src1->type) << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
+    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << ggml_type_name(dst->type) << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
     std::cerr << "), " << (dryrun ? "dryrun" : "") << ")");
     GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16);  // NOLINT
     GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);  // NOLINT
@@ -11168,7 +11168,7 @@ size_t comp_nb[GGML_MAX_DIMS];
 size_t check_counter = 0;
 static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, int tensor_idx) {
     ggml_tensor * tensor = cgraph->nodes[tensor_idx];
-    if (tensor->op == GGML_OP_TRANSPOSE) {
+    if (tensor->op == GGML_OP_TRANSPOSE || tensor->op == GGML_OP_SET_ROWS) {
         return;
     }
 
@@ -11288,7 +11288,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         tensor_clone = ggml_upscale_ext(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], (ggml_scale_mode) tensor->op_params[0]);
     } else if (tensor->op == GGML_OP_SCALE) {
         const float * params = (const float *)tensor->op_params;
-        tensor_clone = ggml_scale(ggml_ctx, src_clone[0], params[0]);
+        tensor_clone = ggml_scale_bias(ggml_ctx, src_clone[0], params[0], params[1]);
     } else if (tensor->op == GGML_OP_SQR) {
         tensor_clone = ggml_sqr(ggml_ctx, src_clone[0]);
     } else if (tensor->op == GGML_OP_SIN) {
@@ -11399,8 +11399,6 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         } else {
             tensor_clone = ggml_cpy(ggml_ctx, src_clone[0], src_clone[1]);
         }
-    } else if (tensor->op == GGML_OP_SET_ROWS) {
-        tensor_clone = ggml_set_rows(ggml_ctx, src_clone[0], src_clone[1]);
     } else if (tensor->op == GGML_OP_CONT) {
         tensor_clone = ggml_cont_4d(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
     } else if (tensor->op == GGML_OP_RESHAPE) {
@@ -11508,7 +11506,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
 
 static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, int tensor_idx) {
     ggml_tensor * tensor = cgraph->nodes[tensor_idx];
-    if (tensor->op == GGML_OP_TRANSPOSE) {
+    if (tensor->op == GGML_OP_TRANSPOSE || tensor->op == GGML_OP_SET_ROWS) {
         return;
     }
     bool fused_rms_norm_mul = false;
@@ -11568,6 +11566,9 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_cgraph *
                         } else if (tensor->type == GGML_TYPE_F16) {
                             correct = ggml_fp16_to_fp32(*(ggml_fp16_t *) ((char *) comp_result + i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0]));
                             result  = ggml_fp16_to_fp32(*(ggml_fp16_t *) ((char *) tensor_data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]));
+                        } else if (tensor->type == GGML_TYPE_BF16) {
+                            correct = ggml_bf16_to_fp32(*(ggml_bf16_t *) ((char *) comp_result + i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0]));
+                            result  = ggml_bf16_to_fp32(*(ggml_bf16_t *) ((char *) tensor_data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]));
                         } else if (tensor->type == GGML_TYPE_I32) {
                             correct = *(int32_t *) ((char *) comp_result + i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0]);
                             result  = *(int32_t *) ((char *) tensor_data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]);

From a3145e5c1a4fcbee967507baf35780766842ff4f Mon Sep 17 00:00:00 2001
From: diannao <55k@outlook.com>
Date: Fri, 1 Aug 2025 10:02:34 +0800
Subject: [PATCH 19/86] docker : add cann build pipline (llama/14591)

* docker: add cann build pipline

* docker: add cann build pipline

* docker: fix cann devops

* cann : fix multi card hccl

* Update ggml/src/ggml-cann/ggml-cann.cpp

Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com>

* Update ggml-cann.cpp

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com>
---
 ggml/src/ggml-cann/ggml-cann.cpp | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
index 49f55891d85..8eb8b1470bb 100755
--- a/ggml/src/ggml-cann/ggml-cann.cpp
+++ b/ggml/src/ggml-cann/ggml-cann.cpp
@@ -2016,6 +2016,9 @@ static bool ggml_backend_cann_cpy_tensor_async(
         (ggml_backend_cann_context*)backend_dst->context;
 
     size_t copy_size = ggml_nbytes(dst);
+    if (copy_size == 0) {
+        return true;
+    }
     if (backend_src != backend_dst) {
         ggml_backend_cann_buffer_context* buf_ctx_src =
             (ggml_backend_cann_buffer_context*)buf_src->context;

From 6eb52a2caffcf922f7a1cf3c6eb6ddd78e9c30f7 Mon Sep 17 00:00:00 2001
From: Srihari-mcw <96763064+Srihari-mcw@users.noreply.github.com>
Date: Fri, 1 Aug 2025 11:50:33 +0530
Subject: [PATCH 20/86] ggml : Q2k interleaving implementation - x86/x64 SIMD
 (llama/14373)

* Initial Q2_K Block Interleaving Implementation

* Addressed review comments and clean up of the code

* Post rebase fixes

* Initial CI/CD fixes

* Update declarations in arch-fallback.h

* Changes for GEMV Q2_K in arch-fallback.h

* Enable repacking only on AVX-512 machines

* Update comments in repack.cpp

* Address q2k comments

---------

Co-authored-by: Manogna-Sree <elisetti.manognasree@multicorewareinc.com>
---
 ggml/src/ggml-cpu/arch-fallback.h     |   14 +
 ggml/src/ggml-cpu/arch/x86/repack.cpp | 3196 +++++++++++++++++++++++++
 ggml/src/ggml-cpu/repack.cpp          |  263 ++
 ggml/src/ggml-cpu/repack.h            |   11 +
 4 files changed, 3484 insertions(+)

diff --git a/ggml/src/ggml-cpu/arch-fallback.h b/ggml/src/ggml-cpu/arch-fallback.h
index 10e5342516a..f02cfe8fa5c 100644
--- a/ggml/src/ggml-cpu/arch-fallback.h
+++ b/ggml/src/ggml-cpu/arch-fallback.h
@@ -37,17 +37,21 @@
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64)
 // repack.cpp
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
@@ -72,11 +76,13 @@
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #elif defined(__loongarch64)
 // quants.c
@@ -92,11 +98,13 @@
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #elif defined(__riscv)
 // quants.c
@@ -119,10 +127,12 @@
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #elif defined(__s390x__)
 // quants.c
@@ -147,11 +157,13 @@
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #elif defined(__wasm__)
 // quants.c
@@ -175,10 +187,12 @@
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #endif
diff --git a/ggml/src/ggml-cpu/arch/x86/repack.cpp b/ggml/src/ggml-cpu/arch/x86/repack.cpp
index 1982cfef995..37933a4b23f 100644
--- a/ggml/src/ggml-cpu/arch/x86/repack.cpp
+++ b/ggml/src/ggml-cpu/arch/x86/repack.cpp
@@ -849,6 +849,319 @@ void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 #endif
 }
 
+void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK_K;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__AVX2__)
+    // Lookup table to convert signed nibbles to signed bytes
+    __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
+    signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
+    // Shuffle masks to rearrange delta values to multiply with appropriate scales
+    __m128i deltamask = _mm_set_epi8(15, 14, 7, 6, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0);
+    // Permute mask used for easier vector processing at later stages
+    __m256i finalpermutemask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
+
+    const __m256i m3b = _mm256_set1_epi8(3);
+    const __m128i m4b_sse = _mm_set1_epi8(0xF);
+
+    //Mask to get appropriate scales
+    __m128i scalemask1 = _mm_set_epi8(14,14,6,6,12,12,4,4,10,10,2,2,8,8,0,0);
+    __m128i scalemask2 = _mm_set_epi8(15,15,7,7,13,13,5,5,11,11,3,3,9,9,1,1);
+
+    int64_t b_nb = n / QK_K;
+
+    const block_q2_Kx8 * b_ptr_start = (const block_q2_Kx8 *)vx;
+    const block_q8_K * a_ptr_start = (const block_q8_K *)vy;
+
+    // Process Q8_K blocks one by one
+    for (int64_t y = 0; y < nr; y++) {
+
+        // Pointers to LHS blocks of block_q8_K format
+        const block_q8_K * a_ptr = a_ptr_start + (y * nb);
+
+        // Take group of eight interleaved block_q2_K structures at each pass of the loop and perform dot product operation
+        for(int64_t x = 0; x < nc / 8; x++) {
+
+            // Pointers to RHS blocks
+            const block_q2_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
+
+            // Master FP accumulators
+            __m256 acc_row = _mm256_setzero_ps();
+            __m256 acc_min_rows = _mm256_setzero_ps();
+
+            for (int64_t b = 0; b < nb; b++) {
+
+                // Load and convert to FP32 delta from block_q8_K
+                const __m256 row_scale_f32 = _mm256_set1_ps((a_ptr[b].d));
+
+                // Load the delta values for the 8 blocks interleaved in block_q2_Kx8
+                // col_scale_f32 rearranged so as to multiply with appropriate quants
+                const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, deltamask);
+                const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin);
+
+                __m256i iacc_b = _mm256_setzero_si256();
+                __m256i iacc_min_b = _mm256_setzero_si256();
+
+                // Processes eight sub blocks from each Q2_K in each iteration
+                for(int sb = 0; sb < QK_K / 128; sb++) {
+
+                    // Load the eight block_q2_K for eight sub blocks quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
+                    const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + sb * 256));
+                    const __m256i rhs_raw_vec_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 32 + sb * 256));
+                    const __m256i rhs_raw_vec_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 64 + sb * 256));
+                    const __m256i rhs_raw_vec_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 96 + sb * 256));
+                    const __m256i rhs_raw_vec_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 128 + sb * 256));
+                    const __m256i rhs_raw_vec_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 160 + sb * 256));
+                    const __m256i rhs_raw_vec_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 192 + sb * 256));
+                    const __m256i rhs_raw_vec_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 224 + sb * 256));
+
+                    // 2-bit -> 8-bit
+                    // Values of the 0th,2nd,4th,6th sub blocks of eight block_q2_K structures for the sb loop
+                    const __m256i rhs_vec_0123_00 = _mm256_and_si256(rhs_raw_vec_0123_0, m3b); //B00(0-7) B01(0-7) B02(0-7) B03(0-7)
+                    const __m256i rhs_vec_0123_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 2), m3b); //B20(0-7) B21(0-7) B22(0-7) B23(0-7)
+                    const __m256i rhs_vec_0123_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 4), m3b); //B40(0-7) B41(0-7) B42(0-7) B43(0-7)
+                    const __m256i rhs_vec_0123_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 6), m3b); //B60(0-7) B61(0-7) B62(0-7) B63(0-7)
+
+                    const __m256i rhs_vec_4567_00 = _mm256_and_si256(rhs_raw_vec_4567_0, m3b); //B04(0-7) B05(0-7) B06(0-7) B07(0-7)
+                    const __m256i rhs_vec_4567_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 2), m3b); //B24(0-7) B25(0-7) B26(0-7) B27(0-7)
+                    const __m256i rhs_vec_4567_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 4), m3b); //B44(0-7) B45(0-7) B46(0-7) B47(0-7)
+                    const __m256i rhs_vec_4567_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 6), m3b); //B64(0-7) B65(0-7) B66(0-7) B67(0-7)
+
+                    const __m256i rhs_vec_0123_01 = _mm256_and_si256(rhs_raw_vec_0123_1, m3b); //B00(8-15) B01(8-15) B02(8-15) B03(8-15)
+                    const __m256i rhs_vec_0123_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 2), m3b); //B20(8-15) B21(8-15) B22(8-15) B23(8-15)
+                    const __m256i rhs_vec_0123_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m3b); //B40(8-15) B41(8-15) B42(8-15) B43(8-15)
+                    const __m256i rhs_vec_0123_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 6), m3b); //B60(8-15) B61(8-15) B62(8-15) B63(8-15)
+
+                    const __m256i rhs_vec_4567_01 = _mm256_and_si256(rhs_raw_vec_4567_1, m3b); //B04(8-15) B05(8-15) B06(8-15) B07(8-15)
+                    const __m256i rhs_vec_4567_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 2), m3b); //B24(8-15) B25(8-15) B26(8-15) B27(8-15)
+                    const __m256i rhs_vec_4567_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m3b); //B44(8-15) B45(8-15) B46(8-15) B47(8-15)
+                    const __m256i rhs_vec_4567_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 6), m3b); //B64(8-15) B65(8-15) B66(8-15) B67(8-15)
+
+                    // Values of the 1st,3rd,5th,7th sub blocks of eight block_q2_K structures for the sb loop
+                    const __m256i rhs_vec_0123_10 = _mm256_and_si256(rhs_raw_vec_0123_2, m3b); //B10(0-7) B11(0-7) B12(0-7) B13(0-7)
+                    const __m256i rhs_vec_0123_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 2), m3b); //B30(0-7) B31(0-7) B32(0-7) B33(0-7)
+                    const __m256i rhs_vec_0123_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 4), m3b); //B50(0-7) B51(0-7) B52(0-7) B53(0-7)
+                    const __m256i rhs_vec_0123_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 6), m3b); //B70(0-7) B71(0-7) B72(0-7) B73(0-7)
+
+                    const __m256i rhs_vec_4567_10 = _mm256_and_si256(rhs_raw_vec_4567_2, m3b); //B14(0-7) B15(0-7) B16(0-7) B17(0-7)
+                    const __m256i rhs_vec_4567_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 2), m3b); //B34(0-7) B35(0-7) B36(0-7) B37(0-7)
+                    const __m256i rhs_vec_4567_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 4), m3b); //B54(0-7) B55(0-7) B56(0-7) B57(0-7)
+                    const __m256i rhs_vec_4567_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 6), m3b); //B74(0-7) B75(0-7) B76(0-7) B77(0-7)
+
+                    const __m256i rhs_vec_0123_11 = _mm256_and_si256(rhs_raw_vec_0123_3, m3b); //B10(8-15) B11(8-15) B12(8-15) B13(8-15)
+                    const __m256i rhs_vec_0123_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 2), m3b); //B30(8-15) B31(8-15) B32(8-15) B33(8-15)
+                    const __m256i rhs_vec_0123_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 4), m3b); //B50(8-15) B51(8-15) B52(8-15) B53(8-15)
+                    const __m256i rhs_vec_0123_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 6), m3b); //B70(8-15) B71(8-15) B72(8-15) B73(8-15)
+
+                    const __m256i rhs_vec_4567_11 = _mm256_and_si256(rhs_raw_vec_4567_3, m3b); //B14(8-15) B15(8-15) B16(8-15) B17(8-15)
+                    const __m256i rhs_vec_4567_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 2), m3b); //B34(8-15) B35(8-15) B36(8-15) B37(8-15)
+                    const __m256i rhs_vec_4567_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 4), m3b); //B54(8-15) B55(8-15) B56(8-15) B57(8-15)
+                    const __m256i rhs_vec_4567_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 6), m3b); //B74(8-15) B75(8-15) B76(8-15) B77(8-15)
+
+                    //Scales and Mins of corresponding sub blocks from different Q2_K structures are stored together
+                    //s00 m00  s01 m01   s10 m10  s11 m11  s20 m20  s21 m21   s30 m30  s31 m31  s40 m40  s41 m41   s50 m50  s51 m51  s60 m60  s61 m61   s70 m70  s71 m71
+
+                    const __m128i mins_and_scales_01 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + sb * 64));
+                    const __m128i mins_and_scales_23 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 16 + sb * 64));
+                    const __m128i mins_and_scales_45 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 32 + sb * 64));
+                    const __m128i mins_and_scales_67 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 48 + sb * 64));
+
+                    // Extract scales which is lower half from mins_and_scales
+                    const __m128i scales_01 = _mm_and_si128(mins_and_scales_01, m4b_sse);
+                    const __m128i scales_23 = _mm_and_si128(mins_and_scales_23, m4b_sse);
+                    const __m128i scales_45 = _mm_and_si128(mins_and_scales_45, m4b_sse);
+                    const __m128i scales_67 = _mm_and_si128(mins_and_scales_67, m4b_sse);
+
+                    // Extract mins which is upper half from mins_and_scales
+                    const __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_01, 4), m4b_sse));
+                    const __m256i mins_23 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_23, 4), m4b_sse));
+                    const __m256i mins_45 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_45, 4), m4b_sse));
+                    const __m256i mins_67 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_67, 4), m4b_sse));
+
+                    // Scales of sub blocks in the sb loop
+                    // Scales of the 0th sub block from each super block
+                    __m128i scales_rearrange_0 = _mm_shuffle_epi8(scales_01, scalemask1);
+                    __m256i scales_0 = _mm256_cvtepu8_epi16(scales_rearrange_0);
+
+                    // Scales of the 1st sub block from each super block
+                    __m128i scales_rearrange_1 = _mm_shuffle_epi8(scales_01, scalemask2);
+                    __m256i scales_1 = _mm256_cvtepu8_epi16(scales_rearrange_1);
+
+                    // Scales of the 2nd sub block from each super block
+                    __m128i scales_rearrange_2 = _mm_shuffle_epi8(scales_23, scalemask1);
+                    __m256i scales_2 = _mm256_cvtepu8_epi16(scales_rearrange_2);
+
+                    // Scales of the 3rd sub block from each super block
+                    __m128i scales_rearrange_3 = _mm_shuffle_epi8(scales_23, scalemask2);
+                    __m256i scales_3 = _mm256_cvtepu8_epi16(scales_rearrange_3);
+
+                    // Scales of the 4th sub block from each super block
+                    __m128i scales_rearrange_4 = _mm_shuffle_epi8(scales_45, scalemask1);
+                    __m256i scales_4 = _mm256_cvtepu8_epi16(scales_rearrange_4);
+
+                    // Scales of the 5th sub block from each super block
+                    __m128i scales_rearrange_5 = _mm_shuffle_epi8(scales_45, scalemask2);
+                    __m256i scales_5 = _mm256_cvtepu8_epi16(scales_rearrange_5);
+
+                    // Scales of the 6th sub block from each super block
+                    __m128i scales_rearrange_6 = _mm_shuffle_epi8(scales_67, scalemask1);
+                    __m256i scales_6 = _mm256_cvtepu8_epi16(scales_rearrange_6);
+
+                    // Scales of the 7th sub block from each super block
+                    __m128i scales_rearrange_7 = _mm_shuffle_epi8(scales_67, scalemask2);
+                    __m256i scales_7 = _mm256_cvtepu8_epi16(scales_rearrange_7);
+
+                    // Load the sub block values corresponding to sb in block_q8_K in batches of 16 bytes and replicate the same across 256 bit vector
+                    __m256i lhs_vec_0 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + sb * 128)));
+                    __m256i lhs_vec_1 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 16 + sb * 128)));
+                    __m256i lhs_vec_2 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 32 + sb * 128)));
+                    __m256i lhs_vec_3 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 48 + sb * 128)));
+                    __m256i lhs_vec_4 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 64 + sb * 128)));
+                    __m256i lhs_vec_5 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 80 + sb * 128)));
+                    __m256i lhs_vec_6 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 96 + sb * 128)));
+                    __m256i lhs_vec_7 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 112 + sb * 128)));
+
+                    lhs_vec_0 = _mm256_permute2f128_si256(lhs_vec_0, lhs_vec_0, 0);
+                    lhs_vec_1 = _mm256_permute2f128_si256(lhs_vec_1, lhs_vec_1, 0);
+                    lhs_vec_2 = _mm256_permute2f128_si256(lhs_vec_2, lhs_vec_2, 0);
+                    lhs_vec_3 = _mm256_permute2f128_si256(lhs_vec_3, lhs_vec_3, 0);
+                    lhs_vec_4 = _mm256_permute2f128_si256(lhs_vec_4, lhs_vec_4, 0);
+                    lhs_vec_5 = _mm256_permute2f128_si256(lhs_vec_5, lhs_vec_5, 0);
+                    lhs_vec_6 = _mm256_permute2f128_si256(lhs_vec_6, lhs_vec_6, 0);
+                    lhs_vec_7 = _mm256_permute2f128_si256(lhs_vec_7, lhs_vec_7, 0);
+
+                    __m256i iacc_0 = _mm256_setzero_si256();
+                    __m256i iacc_1 = _mm256_setzero_si256();
+                    __m256i iacc_2 = _mm256_setzero_si256();
+                    __m256i iacc_3 = _mm256_setzero_si256();
+                    __m256i iacc_4 = _mm256_setzero_si256();
+                    __m256i iacc_5 = _mm256_setzero_si256();
+                    __m256i iacc_6 = _mm256_setzero_si256();
+                    __m256i iacc_7 = _mm256_setzero_si256();
+
+                    // Dot product done within 32 bit lanes and accumulated in the same vector
+                    // First done for 0th sub block and then for seven (1st - 7th) other sub blocks processed for each sb (sb < QK_K/128 loop)                    // B0(0-3) B4(0-3) B1(0-3) B5(0-3) B2(0-3) B6(0-3) B3(0-3) B7(0-3) with A0(0-3)
+                    // B0(4-7) B4(4-7) B1(4-7) B5(4-7) B2(4-7) B6(4-7) B3(4-7) B7(4-7) with A0(4-7)
+                    // B0(8-11) B4(8-11) B1(8-11) B5(8-11) B2(8-11) B6(8-11) B3(8-11) B7(8-11) with A0(8-11)
+                    // B0(12-15) B4(12-15) B1(12-15) B5(12-15) B2(12-15) B6(12-15) B3(12-15) B7(12-15) with A0(12-15)
+
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_00 ,_mm256_shuffle_epi32(rhs_vec_4567_00, 177), 170), _mm256_shuffle_epi32(lhs_vec_0, 0)));
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_00, 177) ,rhs_vec_4567_00, 170), _mm256_shuffle_epi32(lhs_vec_0, 85)));
+
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_01 ,_mm256_shuffle_epi32(rhs_vec_4567_01, 177), 170), _mm256_shuffle_epi32(lhs_vec_0, 170)));
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_01, 177) ,rhs_vec_4567_01, 170), _mm256_shuffle_epi32(lhs_vec_0, 255)));
+
+                    iacc_0 = _mm256_madd_epi16(iacc_0, scales_0);
+
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_10 ,_mm256_shuffle_epi32(rhs_vec_4567_10, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 0)));
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_10, 177) ,rhs_vec_4567_10, 170), _mm256_shuffle_epi32(lhs_vec_1, 85)));
+
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_11 ,_mm256_shuffle_epi32(rhs_vec_4567_11, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 170)));
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_11, 177) ,rhs_vec_4567_11, 170), _mm256_shuffle_epi32(lhs_vec_1, 255)));
+
+                    iacc_1 = _mm256_madd_epi16(iacc_1, scales_1);
+
+                    iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_20 ,_mm256_shuffle_epi32(rhs_vec_4567_20, 177), 170), _mm256_shuffle_epi32(lhs_vec_2, 0)));
+                    iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_20, 177) ,rhs_vec_4567_20, 170), _mm256_shuffle_epi32(lhs_vec_2, 85)));
+
+                    iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_21 ,_mm256_shuffle_epi32(rhs_vec_4567_21, 177), 170), _mm256_shuffle_epi32(lhs_vec_2, 170)));
+                    iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_21, 177) ,rhs_vec_4567_21, 170), _mm256_shuffle_epi32(lhs_vec_2, 255)));
+
+                    iacc_2 = _mm256_madd_epi16(iacc_2, scales_2);
+
+                    iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_30 ,_mm256_shuffle_epi32(rhs_vec_4567_30, 177), 170), _mm256_shuffle_epi32(lhs_vec_3, 0)));
+                    iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_30, 177) ,rhs_vec_4567_30, 170), _mm256_shuffle_epi32(lhs_vec_3, 85)));
+
+                    iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_31 ,_mm256_shuffle_epi32(rhs_vec_4567_31, 177), 170), _mm256_shuffle_epi32(lhs_vec_3, 170)));
+                    iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_31, 177) ,rhs_vec_4567_31, 170), _mm256_shuffle_epi32(lhs_vec_3, 255)));
+
+                    iacc_3 = _mm256_madd_epi16(iacc_3, scales_3);
+
+                    iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_40 ,_mm256_shuffle_epi32(rhs_vec_4567_40, 177), 170), _mm256_shuffle_epi32(lhs_vec_4, 0)));
+                    iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_40, 177) ,rhs_vec_4567_40, 170), _mm256_shuffle_epi32(lhs_vec_4, 85)));
+
+                    iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_41 ,_mm256_shuffle_epi32(rhs_vec_4567_41, 177), 170), _mm256_shuffle_epi32(lhs_vec_4, 170)));
+                    iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_41, 177) ,rhs_vec_4567_41, 170), _mm256_shuffle_epi32(lhs_vec_4, 255)));
+
+                    iacc_4 = _mm256_madd_epi16(iacc_4, scales_4);
+
+                    iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_50 ,_mm256_shuffle_epi32(rhs_vec_4567_50, 177), 170), _mm256_shuffle_epi32(lhs_vec_5, 0)));
+                    iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_50, 177) ,rhs_vec_4567_50, 170), _mm256_shuffle_epi32(lhs_vec_5, 85)));
+
+                    iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_51 ,_mm256_shuffle_epi32(rhs_vec_4567_51, 177), 170), _mm256_shuffle_epi32(lhs_vec_5, 170)));
+                    iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_51, 177) ,rhs_vec_4567_51, 170), _mm256_shuffle_epi32(lhs_vec_5, 255)));
+
+                    iacc_5 = _mm256_madd_epi16(iacc_5, scales_5);
+
+                    iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_60 ,_mm256_shuffle_epi32(rhs_vec_4567_60, 177), 170), _mm256_shuffle_epi32(lhs_vec_6, 0)));
+                    iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_60, 177) ,rhs_vec_4567_60, 170), _mm256_shuffle_epi32(lhs_vec_6, 85)));
+
+                    iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_61 ,_mm256_shuffle_epi32(rhs_vec_4567_61, 177), 170), _mm256_shuffle_epi32(lhs_vec_6, 170)));
+                    iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_61, 177) ,rhs_vec_4567_61, 170), _mm256_shuffle_epi32(lhs_vec_6, 255)));
+
+                    iacc_6 = _mm256_madd_epi16(iacc_6, scales_6);
+
+                    iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_70 ,_mm256_shuffle_epi32(rhs_vec_4567_70, 177), 170), _mm256_shuffle_epi32(lhs_vec_7, 0)));
+                    iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_70, 177) ,rhs_vec_4567_70, 170), _mm256_shuffle_epi32(lhs_vec_7, 85)));
+
+                    iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_71 ,_mm256_shuffle_epi32(rhs_vec_4567_71, 177), 170), _mm256_shuffle_epi32(lhs_vec_7, 170)));
+                    iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_71, 177) ,rhs_vec_4567_71, 170), _mm256_shuffle_epi32(lhs_vec_7, 255)));
+
+                    iacc_7 = _mm256_madd_epi16(iacc_7, scales_7);
+
+                    // Accumulate the iacc value for one sb
+                    __m256i iacc_sb = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_0, iacc_1), _mm256_add_epi32(iacc_2, iacc_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_4, iacc_5), _mm256_add_epi32(iacc_6, iacc_7)));
+
+                    __m128i q8sums = _mm_loadu_si128((const __m128i *)(a_ptr[b].bsums + sb * 8));
+                    __m256i q8s = _mm256_castsi128_si256(q8sums);
+                    q8s= _mm256_permute2f128_si256(q8s, q8s, 0);
+
+                    // Broadcast the bsums of the two corresponding subblocks of q8_k
+                    // Multiply-Add with corresponding mins of Q2_Kx8 with bsums
+                    __m256i iacc_min_sb_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 0), mins_01);
+                    __m256i iacc_min_sb_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 85), mins_23);
+                    __m256i iacc_min_sb_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 170), mins_45);
+                    __m256i iacc_min_sb_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 255), mins_67);
+
+                    __m256i iacc_min_sb = _mm256_add_epi32(_mm256_add_epi32(iacc_min_sb_01, iacc_min_sb_23), _mm256_add_epi32(iacc_min_sb_45,iacc_min_sb_67));
+
+                    // Accumulate for the complete block
+                    iacc_b = _mm256_add_epi32(iacc_b, iacc_sb);
+                    iacc_min_b = _mm256_add_epi32(iacc_min_b, iacc_min_sb);
+                }
+
+                //Multiply-Add with scale values for complete super block
+                acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_b), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row);
+                acc_min_rows = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_min_b), _mm256_mul_ps(col_dmin_f32, row_scale_f32), acc_min_rows);
+            }
+            // Accumulated output values permuted so as to be stored in appropriate order post accumulation
+            acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask);
+            _mm256_storeu_ps(s + (y * nr + x * 8), _mm256_sub_ps(acc_row, acc_min_rows));
+        }
+    }
+#else
+
+    ggml_gemv_q2_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+
+#endif
+}
+
 void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
@@ -3050,3 +3363,2886 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     ggml_gemm_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
 #endif
 }
+
+void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK_K;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__AVX2__) || defined(__AVX512F__)
+    const block_q2_Kx8 * b_ptr_start = (const block_q2_Kx8 * ) vx;
+    const block_q8_Kx4 * a_ptr_start = (const block_q8_Kx4 * ) vy;
+    int64_t b_nb = n / QK_K;
+    int64_t y = 0;
+
+    // Permute mask used for easier vector processing at later stages
+    __m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4);
+    int64_t xstart = 0;
+    int anr = nr - nr % 16; // Used to align nr with boundary of 16
+
+    // Mask to convert 2 bit and 4 bit values into a bytes
+    const __m256i m3b = _mm256_set1_epi8(3);
+    const __m128i m4b_sse = _mm_set1_epi8(0xF);
+
+    //Mask to get appropriate scales
+    __m128i scalesmask1_sse = _mm_set_epi8(14,14,12,12,10,10,8,8,6,6,4,4,2,2,0,0);
+    __m128i scalesmask2_sse = _mm_set_epi8(15,15,13,13,11,11,9,9,7,7,5,5,3,3,1,1);
+
+    __m256i scalesmask1 = _mm256_castsi128_si256(scalesmask1_sse);
+    scalesmask1 = _mm256_permute2f128_si256(scalesmask1, scalesmask1, 0);
+    __m256i scalesmask2 = _mm256_castsi128_si256(scalesmask2_sse);
+    scalesmask2 = _mm256_permute2f128_si256(scalesmask2, scalesmask2, 0);
+
+#ifdef __AVX512F__
+
+    int anc = nc - nc % 16; // Used to align nc with boundary of 16
+
+    // Mask to mask out nibbles from packed bytes
+    const __m256i m4b = _mm256_set1_epi8(0x0F);
+    // Mask to mask out nibbles from packed bytes expanded to 512 bit length
+    const __m512i m3bexpanded = _mm512_set1_epi8(3);
+    //Take group of four block_q8_Kx4 structures at each pass of the loop and perform dot product operation
+    for (; y < anr / 4; y += 4) {
+
+        const block_q8_Kx4 * a_ptrs[4];
+
+        a_ptrs[0] = a_ptr_start + (y * nb);
+        for (int i = 0; i < 3; ++i) {
+            a_ptrs[i + 1] = a_ptrs[i] + nb;
+        }
+
+        // Take group of eight block_q2_kx8 structures at each pass of the loop and perform dot product operation
+        for (int64_t x = 0; x < anc / 8; x += 2) {
+
+            const block_q2_Kx8 * b_ptr_0 = b_ptr_start + ((x) * b_nb);
+            const block_q2_Kx8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
+
+            // Master FP accumulators
+            __m512 acc_rows[16];
+            for (int i = 0; i < 16; i++) {
+                acc_rows[i] = _mm512_setzero_ps();
+            }
+
+            __m512 acc_min_rows[16];
+            for (int i = 0; i < 16; i++) {
+                acc_min_rows[i] = _mm512_setzero_ps();
+            }
+            // For super block
+            for (int64_t b = 0; b < nb; b++) {
+                // Delta values - Load the sixteen scale values from two block_q2_kx8 structures
+                const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
+
+                // dmin values - Load the sixteen dmin values from two block_q2_kx8 structures
+                const __m512 col_dmin_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].dmin, b_ptr_1[b].dmin);
+
+                // Loop to iterate over the sixteen sub blocks of a super block - eight sub blocks are processed per iteration
+                for (int sb = 0; sb < QK_K / 128; sb++) {
+
+                    // Load the eight block_q2_k for eight sub blocks quantized values interleaved with each other in chunks of eight bytes - B0,B1 ....B6,B7
+                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + sb * 256));
+                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 32 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 64 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 96 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 128 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 160 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 192 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 224 + sb * 256));
+
+                    const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 32 + sb * 256));
+                    const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 64 + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 96 + sb * 256));
+                    const __m256i rhs_raw_mat_89AB_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 128 + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 160 + sb * 256));
+                    const __m256i rhs_raw_mat_89AB_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 192 + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 224 + sb * 256));
+
+                    const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+                    const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+                    const __m256i rhs_raw_mat_0145_2 = _mm256_blend_epi32(rhs_raw_mat_0123_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_2, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_2, requiredOrder), rhs_raw_mat_4567_2, 240);
+                    const __m256i rhs_raw_mat_0145_3 = _mm256_blend_epi32(rhs_raw_mat_0123_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_3, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_3, requiredOrder), rhs_raw_mat_4567_3, 240);
+
+                    const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240);
+                    const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240);
+                    const __m256i rhs_raw_mat_89CD_2 = _mm256_blend_epi32(rhs_raw_mat_89AB_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_2, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_2, requiredOrder), rhs_raw_mat_CDEF_2, 240);
+                    const __m256i rhs_raw_mat_89CD_3 = _mm256_blend_epi32(rhs_raw_mat_89AB_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_3, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_3, requiredOrder), rhs_raw_mat_CDEF_3, 240);
+
+                    const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
+                    const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
+
+                    const __m512i rhs_raw_mat_014589CD_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_2), rhs_raw_mat_89CD_2, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_2), rhs_raw_mat_ABEF_2, 1);
+                    const __m512i rhs_raw_mat_014589CD_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_3), rhs_raw_mat_89CD_3, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_3), rhs_raw_mat_ABEF_3, 1);
+
+                    //2-bit -> 8-bit
+                    const __m512i rhs_mat_014589CD_00 = _mm512_and_si512(rhs_raw_mat_014589CD_0,m3bexpanded); //B00(0-7) B01(0-7) B04(0-7) B05(0-7) B08(0-7) B09(0-7) B0C(0-7) B0D(0-7)
+                    const __m512i rhs_mat_2367ABEF_00 = _mm512_and_si512(rhs_raw_mat_2367ABEF_0,m3bexpanded); //B02(0-7) B03(0-7) B06(0-7) B07(0-7) B0A(0-7) B0B(0-7) B0E(0-7) B0F(0-7)
+                    const __m512i rhs_mat_014589CD_01 = _mm512_and_si512(rhs_raw_mat_014589CD_1,m3bexpanded); //B00(8-15) B01(8-15) B04(8-15) B05(8-15) B08(8-15) B09(8-15) B0C(8-15) B0D(8-15)
+                    const __m512i rhs_mat_2367ABEF_01 = _mm512_and_si512(rhs_raw_mat_2367ABEF_1,m3bexpanded); //B02(8-15) B03(8-15) B06(8-15) B07(8-15) B0A(8-15) B0B(8-15) B0E(8-15) B0F(8-15)
+                    const __m512i rhs_mat_014589CD_10 = _mm512_and_si512(rhs_raw_mat_014589CD_2,m3bexpanded); //B10(0-7) B11(0-7) B14(0-7) B15(0-7) B18(0-7) B19(0-7) B1C(0-7) B1D(0-7)
+                    const __m512i rhs_mat_2367ABEF_10 = _mm512_and_si512(rhs_raw_mat_2367ABEF_2,m3bexpanded); //B12(0-7) B13(0-7) B16(0-7) B17(0-7) B1A(0-7) B1B(0-7) B1E(0-7) B1F(0-7)
+                    const __m512i rhs_mat_014589CD_11 = _mm512_and_si512(rhs_raw_mat_014589CD_3,m3bexpanded); //B10(8-15) B11(8-15) B14(8-15) B15(8-15) B18(8-15) B19(8-15) B1C(8-15) B1D(8-15)
+                    const __m512i rhs_mat_2367ABEF_11 = _mm512_and_si512(rhs_raw_mat_2367ABEF_3,m3bexpanded); //B12(8-15) B13(8-15) B16(8-15) B17(8-15) B1A(8-15) B1B(8-15) B1E(8-15) B1F(8-15)
+
+                    const __m512i rhs_mat_014589CD_20 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 2), m3bexpanded); //B20(0-7) B21(0-7) B24(0-7) B25(0-7) B28(0-7) B29(0-7) B2C(0-7) B2D(0-7)
+                    const __m512i rhs_mat_2367ABEF_20 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 2), m3bexpanded); //B22(0-7) B23(0-7) B26(0-7) B27(0-7) B2A(0-7) B2B(0-7) B2E(0-7) B2F(0-7)
+
+                    const __m512i rhs_mat_014589CD_21 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 2), m3bexpanded); //B20(8-15) B21(8-15) B24(8-15) B25(8-15) B28(8-15) B29(8-15) B2C(8-15) B2D(8-15)
+                    const __m512i rhs_mat_2367ABEF_21 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 2), m3bexpanded); //B22(8-15) B23(8-15) B26(8-15) B27(8-15) B2A(8-15) B2B(8-15) B2E(8-15) B2F(8-15)
+
+                    const __m512i rhs_mat_014589CD_30 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2, 2), m3bexpanded); //B30(0-7) B31(0-7) B34(0-7) B35(0-7) B38(0-7) B39(0-7) B3C(0-7) B3D(0-7)
+                    const __m512i rhs_mat_2367ABEF_30 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2, 2), m3bexpanded); //B32(0-7) B33(0-7) B36(0-7) B37(0-7) B3A(0-7) B3B(0-7) B3E(0-7) B3F(0-7)
+
+                    const __m512i rhs_mat_014589CD_31 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3, 2), m3bexpanded); //B30(8-15) B31(8-15) B34(8-15) B35(8-15) B38(8-15) B39(8-15) B3C(8-15) B3D(8-15)
+                    const __m512i rhs_mat_2367ABEF_31 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3, 2), m3bexpanded); //B32(8-15) B33(8-15) B36(8-15) B37(8-15) B3A(8-15) B3B(8-15) B3E(8-15) B3F(8-15)
+
+                    const __m512i rhs_mat_014589CD_40 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m3bexpanded); //B40(0-7) B41(0-7) B44(0-7) B45(0-7) B48(0-7) B49(0-7) B4C(0-7) B4D(0-7)
+                    const __m512i rhs_mat_2367ABEF_40 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m3bexpanded); //B42(0-7) B43(0-7) B46(0-7) B47(0-7) B4A(0-7) B4B(0-7) B4E(0-7) B4F(0-7)
+
+                    const __m512i rhs_mat_014589CD_41 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m3bexpanded); //B40(8-15) B41(8-15) B44(8-15) B45(8-15) B48(8-15) B49(8-15) B4C(8-15) B4D(8-15)
+                    const __m512i rhs_mat_2367ABEF_41 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m3bexpanded); //B42(8-15) B43(8-15) B46(8-15) B47(8-15) B4A(8-15) B4B(8-15) B4E(8-15) B4F(8-15)
+
+                    const __m512i rhs_mat_014589CD_50 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2, 4), m3bexpanded); //B50(0-7) B51(0-7) B54(0-7) B55(0-7) B58(0-7) B59(0-7) B5C(0-7) B5D(0-7)
+                    const __m512i rhs_mat_2367ABEF_50 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2, 4), m3bexpanded); //B52(0-7) B53(0-7) B56(0-7) B57(0-7) B5A(0-7) B5B(0-7) B5E(0-7) B5F(0-7)
+
+                    const __m512i rhs_mat_014589CD_51 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3, 4), m3bexpanded); //B50(8-15) B51(8-15) B54(8-15) B55(8-15) B58(8-15) B59(8-15) B5C(8-15) B5D(8-15)
+                    const __m512i rhs_mat_2367ABEF_51 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3, 4), m3bexpanded); //B52(8-15) B53(8-15) B56(8-15) B57(8-15) B5A(8-15) B5B(8-15) B5E(8-15) B5F(8-15)
+
+                    const __m512i rhs_mat_014589CD_60 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 6), m3bexpanded); //B60(0-7) B61(0-7) B64(0-7) B65(0-7) B68(0-7) B69(0-7) B6C(0-7) B6D(0-7)
+                    const __m512i rhs_mat_2367ABEF_60 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 6), m3bexpanded); //B62(0-7) B63(0-7) B66(0-7) B67(0-7) B6A(0-7) B6B(0-7) B6E(0-7) B6F(0-7)
+
+                    const __m512i rhs_mat_014589CD_61 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 6), m3bexpanded); //B60(8-15) B61(8-15) B64(8-15) B65(8-15) B68(8-15) B69(8-15) B6C(8-15) B6D(8-15)
+                    const __m512i rhs_mat_2367ABEF_61 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 6), m3bexpanded); //B62(8-15) B63(8-15) B66(8-15) B67(8-15) B6A(8-15) B6B(8-15) B6E(8-15) B6F(8-15)
+
+                    const __m512i rhs_mat_014589CD_70 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2, 6), m3bexpanded); //B70(0-7) B71(0-7) B74(0-7) B75(0-7) B78(0-7) B79(0-7) B7C(0-7) B7D(0-7)
+                    const __m512i rhs_mat_2367ABEF_70 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2, 6), m3bexpanded); //B72(0-7) B73(0-7) B76(0-7) B77(0-7) B7A(0-7) B7B(0-7) B7E(0-7) B7F(0-7)
+
+                    const __m512i rhs_mat_014589CD_71 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3, 6), m3bexpanded); //B70(8-15) B71(8-15) B74(8-15) B75(8-15) B78(8-15) B79(8-15) B7C(8-15) B7D(8-15)
+                    const __m512i rhs_mat_2367ABEF_71 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3, 6), m3bexpanded); //B72(8-15) B73(8-15) B76(8-15) B77(8-15) B7A(8-15) B7B(8-15) B7E(8-15) B7F(8-15)
+
+                    const __m512i rhs_mat_014589CD_00_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_00, (_MM_PERM_ENUM)136); //B00(0-3) B01(0-3) B00(0-3) B01(0-3) B04(0-3) B05(0-3) B04(0-3) B05(0-3) B08(0-3) B09(0-3) B08(0-3) B09(0-3) B0C(0-3) B0D(0-3) B0C(0-3) B0D(0-3)
+                    const __m512i rhs_mat_2367ABEF_00_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_00, (_MM_PERM_ENUM)136); //B02(0-3) B03(0-3) B02(0-3) B03(0-3) B06(0-3) B07(0-3) B06(0-3) B07(0-3) B0A(0-3) B0B(0-3) B0A(0-3) B0B(0-3) B0E(0-3) B0F(0-3) B0E(0-3) B0F(0-3)
+
+                    const __m512i rhs_mat_014589CD_01_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_01, (_MM_PERM_ENUM)136); //B00(8-11) B01(8-11) B00(8-11) B01(8-11) B04(8-11) B05(8-11) B04(8-11) B05(8-11) B08(8-11) B09(8-11) B08(8-11) B09(8-11) B0C(8-11) B0D(8-11) B0C(8-11) B0D(8-11)
+                    const __m512i rhs_mat_2367ABEF_01_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_01, (_MM_PERM_ENUM)136); //B02(8-11) B03(8-11) B02(8-11) B03(8-11) B06(8-11) B07(8-11) B06(8-11) B07(8-11) B0A(8-11) B0B(8-11) B0A(8-11) B0B(8-11) B0E(8-11) B0F(8-11) B0E(8-11) B0F(8-11)
+
+                    const __m512i rhs_mat_014589CD_10_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_10, (_MM_PERM_ENUM)136); //B10(0-3) B11(0-3) B10(0-3) B11(0-3) B14(0-3) B15(0-3) B14(0-3) B15(0-3) B18(0-3) B19(0-3) B18(0-3) B19(0-3) B1C(0-3) B1D(0-3) B1C(0-3) B1D(0-3)
+                    const __m512i rhs_mat_2367ABEF_10_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_10, (_MM_PERM_ENUM)136); //B12(0-3) B13(0-3) B12(0-3) B13(0-3) B16(0-3) B17(0-3) B16(0-3) B17(0-3) B1A(0-3) B1B(0-3) B1A(0-3) B1B(0-3) B1E(0-3) B1F(0-3) B1E(0-3) B1F(0-3)
+
+                    const __m512i rhs_mat_014589CD_11_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_11, (_MM_PERM_ENUM)136); //B10(8-11) B11(8-11) B10(8-11) B11(8-11) B14(8-11) B15(8-11) B14(8-11) B15(8-11) B18(8-11) B19(8-11) B18(8-11) B19(8-11) B1C(8-11) B1D(8-11) B1C(8-11) B1D(8-11)
+                    const __m512i rhs_mat_2367ABEF_11_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_11, (_MM_PERM_ENUM)136); //B12(8-11) B13(8-11) B12(8-11) B13(8-11) B16(8-11) B17(8-11) B16(8-11) B17(8-11) B1A(8-11) B1B(8-11) B1A(8-11) B1B(8-11) B1E(8-11) B1F(8-11) B1E(8-11) B1F(8-11)
+
+                    const __m512i rhs_mat_014589CD_20_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_20, (_MM_PERM_ENUM)136); //B20(0-3) B21(0-3) B20(0-3) B21(0-3) B24(0-3) B25(0-3) B24(0-3) B25(0-3) B28(0-3) B29(0-3) B28(0-3) B29(0-3) B2C(0-3) B2D(0-3) B2C(0-3) B2D(0-3)
+                    const __m512i rhs_mat_2367ABEF_20_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_20, (_MM_PERM_ENUM)136); //B22(0-3) B23(0-3) B22(0-3) B23(0-3) B26(0-3) B27(0-3) B26(0-3) B27(0-3) B2A(0-3) B2B(0-3) B2A(0-3) B2B(0-3) B2E(0-3) B2F(0-3) B2E(0-3) B2F(0-3)
+
+                    const __m512i rhs_mat_014589CD_21_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_21, (_MM_PERM_ENUM)136); //B20(8-11) B21(8-11) B20(8-11) B21(8-11) B24(8-11) B25(8-11) B24(8-11) B25(8-11) B28(8-11) B29(8-11) B28(8-11) B29(8-11) B2C(8-11) B2D(8-11) B2C(8-11) B2D(8-11)
+                    const __m512i rhs_mat_2367ABEF_21_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_21, (_MM_PERM_ENUM)136); //B22(8-11) B23(8-11) B22(8-11) B23(8-11) B26(8-11) B27(8-11) B26(8-11) B27(8-11) B2A(8-11) B2B(8-11) B2A(8-11) B2B(8-11) B2E(8-11) B2F(8-11) B2E(8-11) B2F(8-11)
+
+                    const __m512i rhs_mat_014589CD_30_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_30, (_MM_PERM_ENUM)136); ///B30(0-3) B31(0-3) B30(0-3) B31(0-3) B34(0-3) B35(0-3) B34(0-3) B35(0-3) B38(0-3) B39(0-3) B38(0-3) B39(0-3) B3C(0-3) B3D(0-3) B3C(0-3) B3D(0-3)
+                    const __m512i rhs_mat_2367ABEF_30_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_30, (_MM_PERM_ENUM)136); //B32(0-3) B33(0-3) B32(0-3) B33(0-3) B36(0-3) B37(0-3) B36(0-3) B37(0-3) B3A(0-3) B3B(0-3) B3A(0-3) B3B(0-3) B3E(0-3) B3F(0-3) B3E(0-3) B3F(0-3)
+
+                    const __m512i rhs_mat_014589CD_31_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_31, (_MM_PERM_ENUM)136); //B30(8-11) B31(8-11) B30(8-11) B31(8-11) B34(8-11) B35(8-11) B34(8-11) B35(8-11) B38(8-11) B39(8-11) B38(8-11) B39(8-11) B3C(8-11) B3D(8-11) B3C(8-11) B3D(8-11)
+                    const __m512i rhs_mat_2367ABEF_31_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_31, (_MM_PERM_ENUM)136); //B32(8-11) B33(8-11) B32(8-11) B33(8-11) B36(8-11) B37(8-11) B36(8-11) B37(8-11) B3A(8-11) B3B(8-11) B3A(8-11) B3B(8-11) B3E(8-11) B3F(8-11) B3E(8-11) B3F(8-11)
+
+                    const __m512i rhs_mat_014589CD_40_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_40, (_MM_PERM_ENUM)136); //B40(0-3) B41(0-3) B40(0-3) B41(0-3) B44(0-3) B45(0-3) B44(0-3) B45(0-3) B48(0-3) B49(0-3) B48(0-3) B49(0-3) B4C(0-3) B4D(0-3) B4C(0-3) B4D(0-3)
+                    const __m512i rhs_mat_2367ABEF_40_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_40, (_MM_PERM_ENUM)136); //B42(0-3) B43(0-3) B42(0-3) B43(0-3) B46(0-3) B47(0-3) B46(0-3) B47(0-3) B4A(0-3) B4B(0-3) B4A(0-3) B4B(0-3) B4E(0-3) B4F(0-3) B4E(0-3) B4F(0-3)
+
+                    const __m512i rhs_mat_014589CD_41_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_41, (_MM_PERM_ENUM)136); //B40(8-11) B41(8-11) B40(8-11) B41(8-11) B44(8-11) B45(8-11) B44(8-11) B45(8-11) B48(8-11) B49(8-11) B48(8-11) B49(8-11) B4C(8-11) B4D(8-11) B4C(8-11) B4D(8-11)
+                    const __m512i rhs_mat_2367ABEF_41_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_41, (_MM_PERM_ENUM)136); //B42(8-11) B43(8-11) B42(8-11) B43(8-11) B46(8-11) B47(8-11) B46(8-11) B47(8-11) B4A(8-11) B4B(8-11) B4A(8-11) B4B(8-11) B4E(8-11) B4F(8-11) B4E(8-11) B4F(8-11)
+
+                    const __m512i rhs_mat_014589CD_50_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_50, (_MM_PERM_ENUM)136); //B50(0-3) B51(0-3) B50(0-3) B51(0-3) B54(0-3) B55(0-3) B54(0-3) B55(0-3) B58(0-3) B59(0-3) B58(0-3) B59(0-3) B5C(0-3) B5D(0-3) B5C(0-3) B5D(0-3)
+                    const __m512i rhs_mat_2367ABEF_50_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_50, (_MM_PERM_ENUM)136); //B52(0-3) B53(0-3) B52(0-3) B53(0-3) B56(0-3) B57(0-3) B56(0-3) B57(0-3) B5A(0-3) B5B(0-3) B5A(0-3) B5B(0-3) B5E(0-3) B5F(0-3) B5E(0-3) B5F(0-3)
+
+                    const __m512i rhs_mat_014589CD_51_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_51, (_MM_PERM_ENUM)136); //B50(8-11) B51(8-11) B50(8-11) B51(8-11) B54(8-11) B55(8-11) B54(8-11) B55(8-11) B58(8-11) B59(8-11) B58(8-11) B59(8-11) B5C(8-11) B5D(8-11) B5C(8-11) B5D(8-11)
+                    const __m512i rhs_mat_2367ABEF_51_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_51, (_MM_PERM_ENUM)136); //B52(8-11) B53(8-11) B52(8-11) B53(8-11) B56(8-11) B57(8-11) B56(8-11) B57(8-11) B5A(8-11) B5B(8-11) B5A(8-11) B5B(8-11) B5E(8-11) B5F(8-11) B5E(8-11) B5F(8-11)
+
+                    const __m512i rhs_mat_014589CD_60_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_60, (_MM_PERM_ENUM)136); //B60(0-3) B61(0-3) B60(0-3) B61(0-3) B64(0-3) B65(0-3) B64(0-3) B65(0-3) B68(0-3) B69(0-3) B68(0-3) B69(0-3) B6C(0-3) B6D(0-3) B6C(0-3) B6D(0-3)
+                    const __m512i rhs_mat_2367ABEF_60_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_60, (_MM_PERM_ENUM)136); //B62(0-3) B63(0-3) B62(0-3) B63(0-3) B66(0-3) B67(0-3) B66(0-3) B67(0-3) B6A(0-3) B6B(0-3) B6A(0-3) B6B(0-3) B6E(0-3) B6F(0-3) B6E(0-3) B6F(0-3)
+
+                    const __m512i rhs_mat_014589CD_61_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_61, (_MM_PERM_ENUM)136); //B60(8-11) B61(8-11) B60(8-11) B61(8-11) B64(8-11) B65(8-11) B64(8-11) B65(8-11) B68(8-11) B69(8-11) B68(8-11) B69(8-11) B6C(8-11) B6D(8-11) B6C(8-11) B6D(8-11)
+                    const __m512i rhs_mat_2367ABEF_61_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_61, (_MM_PERM_ENUM)136); //B62(8-11) B63(8-11) B62(8-11) B63(8-11) B66(8-11) B67(8-11) B66(8-11) B67(8-11) B6A(8-11) B6B(8-11) B6A(8-11) B6B(8-11) B6E(8-11) B6F(8-11) B6E(8-11) B6F(8-11)
+
+                    const __m512i rhs_mat_014589CD_70_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_70, (_MM_PERM_ENUM)136); //B70(0-3) B71(0-3) B70(0-3) B71(0-3) B74(0-3) B75(0-3) B74(0-3) B75(0-3) B78(0-3) B79(0-3) B78(0-3) B79(0-3) B7C(0-3) B7D(0-3) B7C(0-3) B7D(0-3)
+                    const __m512i rhs_mat_2367ABEF_70_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_70, (_MM_PERM_ENUM)136); //B72(0-3) B73(0-3) B72(0-3) B73(0-3) B76(0-3) B77(0-3) B76(0-3) B77(0-3) B7A(0-3) B7B(0-3) B7A(0-3) B7B(0-3) B7E(0-3) B7F(0-3) B7E(0-3) B7F(0-3)
+
+                    const __m512i rhs_mat_014589CD_71_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_71, (_MM_PERM_ENUM)136); //B00(8-11) B01(8-11) B00(8-11) B01(8-11) B04(8-11) B05(8-11) B04(8-11) B05(8-11) B08(8-11) B09(8-11) B08(8-11) B09(8-11) B0C(8-11) B0D(8-11) B0C(8-11) B0D(8-11)
+                    const __m512i rhs_mat_2367ABEF_71_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_71, (_MM_PERM_ENUM)136); //B72(8-11) B73(8-11) B72(8-11) B73(8-11) B76(8-11) B77(8-11) B76(8-11) B77(8-11) B7A(8-11) B7B(8-11) B7A(8-11) B7B(8-11) B7E(8-11) B7F(8-11) B7E(8-11) B7F(8-11)
+
+                    const __m512i rhs_mat_014589CD_00_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_00, (_MM_PERM_ENUM)221); //B00(4-7) B01(4-7) B00(4-7) B01(4-7) B04(4-7) B05(4-7) B04(4-7) B05(4-7) B08(4-7) B09(4-7) B08(4-7) B09(4-7) B0C(4-7) B0D(4-7) B0C(4-7) B0D(4-7)
+                    const __m512i rhs_mat_2367ABEF_00_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_00, (_MM_PERM_ENUM)221); //B02(4-7) B03(4-7) B02(4-7) B03(4-7) B06(4-7) B07(4-7) B06(4-7) B07(4-7) B0A(4-7) B0B(4-7) B0A(4-7) B0B(4-7) B0E(4-7) B0F(4-7) B0E(4-7) B0F(4-7)
+
+                    const __m512i rhs_mat_014589CD_01_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_01, (_MM_PERM_ENUM)221); //B00(12-15) B01(12-15) B00(12-15) B01(12-15) B04(12-15) B05(12-15) B04(12-15) B05(12-15) B08(12-15) B09(12-15) B08(12-15) B09(12-15) B0C(12-15) B0D(12-15) B0C(12-15) B0D(12-15)
+                    const __m512i rhs_mat_2367ABEF_01_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_01, (_MM_PERM_ENUM)221); //B02(12-15) B03(12-15) B02(12-15) B03(12-15) B06(12-15) B07(12-15) B06(12-15) B07(12-15) B0A(12-15) B0B(12-15) B0A(12-15) B0B(12-15) B0E(12-15) B0F(12-15) B0E(12-15) B0F(12-15)
+
+                    const __m512i rhs_mat_014589CD_10_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_10, (_MM_PERM_ENUM)221); //B10(4-7) B11(4-7) B10(4-7) B11(4-7) B14(4-7) B15(4-7) B14(4-7) B15(4-7) B18(4-7) B19(4-7) B18(4-7) B19(4-7) B1C(4-7) B1D(4-7) B1C(4-7) B1D(4-7)
+                    const __m512i rhs_mat_2367ABEF_10_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_10, (_MM_PERM_ENUM)221); //B12(4-7) B13(4-7) B12(4-7) B13(4-7) B16(4-7) B17(4-7) B16(4-7) B17(4-7) B1A(4-7) B1B(4-7) B1A(4-7) B1B(4-7) B1E(4-7) B1F(4-7) B1E(4-7) B1F(4-7)
+
+                    const __m512i rhs_mat_014589CD_11_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_11, (_MM_PERM_ENUM)221); //B10(12-15) B11(12-15) B10(12-15) B11(12-15) B14(12-15) B15(12-15) B14(12-15) B15(12-15) B18(12-15) B19(12-15) B18(12-15) B19(12-15) B1C(12-15) B1D(12-15) B1C(12-15) B1D(12-15)
+                    const __m512i rhs_mat_2367ABEF_11_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_11, (_MM_PERM_ENUM)221); //B12(12-15) B13(12-15) B12(12-15) B13(12-15) B16(12-15) B17(12-15) B16(12-15) B17(12-15) B1A(12-15) B1B(12-15) B1A(12-15) B1B(12-15) B1E(12-15) B1F(12-15) B1E(12-15) B1F(12-15)
+
+                    const __m512i rhs_mat_014589CD_20_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_20, (_MM_PERM_ENUM)221); //B20(4-7) B21(4-7) B20(4-7) B21(4-7) B24(4-7) B25(4-7) B24(4-7) B25(4-7) B28(4-7) B29(4-7) B28(4-7) B29(4-7) B2C(4-7) B2D(4-7) B2C(4-7) B2D(4-7)
+                    const __m512i rhs_mat_2367ABEF_20_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_20, (_MM_PERM_ENUM)221); //B22(4-7) B23(4-7) B22(4-7) B23(4-7) B26(4-7) B27(4-7) B26(4-7) B27(4-7) B2A(4-7) B2B(4-7) B2A(4-7) B2B(4-7) B2E(4-7) B2F(4-7) B2E(4-7) B2F(4-7)
+
+                    const __m512i rhs_mat_014589CD_21_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_21, (_MM_PERM_ENUM)221); //B20(12-15) B21(12-15) B20(12-15) B21(12-15) B24(12-15) B25(12-15) B24(12-15) B25(12-15) B28(12-15) B29(12-15) B28(12-15) B29(12-15) B2C(12-15) B2D(12-15) B2C(12-15) B2D(12-15)
+                    const __m512i rhs_mat_2367ABEF_21_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_21, (_MM_PERM_ENUM)221); //B22(12-15) B23(12-15) B22(12-15) B23(12-15) B26(12-15) B27(12-15) B26(12-15) B27(12-15) B2A(12-15) B2B(12-15) B2A(12-15) B2B(12-15) B2E(12-15) B2F(12-15) B2E(12-15) B2F(12-15)
+
+                    const __m512i rhs_mat_014589CD_30_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_30, (_MM_PERM_ENUM)221); //B30(4-7) B31(4-7) B30(4-7) B31(4-7) B34(4-7) B35(4-7) B34(4-7) B35(4-7) B38(4-7) B39(4-7) B38(4-7) B39(4-7) B3C(4-7) B3D(4-7) B3C(4-7) B3D(4-7)
+                    const __m512i rhs_mat_2367ABEF_30_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_30, (_MM_PERM_ENUM)221); //B32(4-7) B33(4-7) B32(4-7) B33(4-7) B36(4-7) B37(4-7) B36(4-7) B37(4-7) B3A(4-7) B3B(4-7) B3A(4-7) B3B(4-7) B3E(4-7) B3F(4-7) B3E(4-7) B3F(4-7)
+
+                    const __m512i rhs_mat_014589CD_31_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_31, (_MM_PERM_ENUM)221); //B30(12-15) B31(12-15) B30(12-15) B31(12-15) B34(12-15) B35(12-15) B34(12-15) B35(12-15) B38(12-15) B39(12-15) B38(12-15) B39(12-15) B3C(12-15) B3D(12-15) B3C(12-15) B3D(12-15)
+                    const __m512i rhs_mat_2367ABEF_31_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_31, (_MM_PERM_ENUM)221); //B32(12-15) B33(12-15) B32(12-15) B33(12-15) B36(12-15) B37(12-15) B36(12-15) B37(12-15) B3A(12-15) B3B(12-15) B3A(12-15) B3B(12-15) B3E(12-15) B3F(12-15) B3E(12-15) B3F(12-15)
+
+                    const __m512i rhs_mat_014589CD_40_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_40, (_MM_PERM_ENUM)221); //B40(4-7) B41(4-7) B40(4-7) B41(4-7) B44(4-7) B45(4-7) B44(4-7) B45(4-7) B48(4-7) B49(4-7) B48(4-7) B49(4-7) B4C(4-7) B4D(4-7) B4C(4-7) B4D(4-7)
+                    const __m512i rhs_mat_2367ABEF_40_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_40, (_MM_PERM_ENUM)221); //B42(4-7) B43(4-7) B42(4-7) B43(4-7) B46(4-7) B47(4-7) B46(4-7) B47(4-7) B4A(4-7) B4B(4-7) B4A(4-7) B4B(4-7) B4E(4-7) B4F(4-7) B4E(4-7) B4F(4-7)
+
+                    const __m512i rhs_mat_014589CD_41_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_41, (_MM_PERM_ENUM)221); //B40(12-15) B41(12-15) B40(12-15) B41(12-15) B44(12-15) B45(12-15) B44(12-15) B45(12-15) B48(12-15) B49(12-15) B48(12-15) B49(12-15) B4C(12-15) B4D(12-15) B4C(12-15) B4D(12-15)
+                    const __m512i rhs_mat_2367ABEF_41_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_41, (_MM_PERM_ENUM)221); //B42(12-15) B43(12-15) B42(12-15) B43(12-15) B46(12-15) B47(12-15) B46(12-15) B47(12-15) B4A(12-15) B4B(12-15) B4A(12-15) B4B(12-15) B4E(12-15) B4F(12-15) B4E(12-15) B4F(12-15)
+
+                    const __m512i rhs_mat_014589CD_50_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_50, (_MM_PERM_ENUM)221); //B50(4-7) B51(4-7) B50(4-7) B51(4-7) B54(4-7) B55(4-7) B54(4-7) B55(4-7) B58(4-7) B59(4-7) B58(4-7) B59(4-7) B5C(4-7) B5D(4-7) B5C(4-7) B5D(4-7)
+                    const __m512i rhs_mat_2367ABEF_50_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_50, (_MM_PERM_ENUM)221); //B52(4-7) B53(4-7) B52(4-7) B53(4-7) B56(4-7) B57(4-7) B56(4-7) B57(4-7) B5A(4-7) B5B(4-7) B5A(4-7) B5B(4-7) B5E(4-7) B5F(4-7) B5E(4-7) B5F(4-7)
+
+                    const __m512i rhs_mat_014589CD_51_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_51, (_MM_PERM_ENUM)221); //B50(12-15) B51(12-15) B50(12-15) B51(12-15) B54(12-15) B55(12-15) B54(12-15) B55(12-15) B58(12-15) B59(12-15) B58(12-15) B59(12-15) B5C(12-15) B5D(12-15) B5C(12-15) B5D(12-15)
+                    const __m512i rhs_mat_2367ABEF_51_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_51, (_MM_PERM_ENUM)221); //B52(12-15) B53(12-15) B52(12-15) B53(12-15) B56(12-15) B57(12-15) B56(12-15) B57(12-15) B5A(12-15) B5B(12-15) B5A(12-15) B5B(12-15) B5E(12-15) B5F(12-15) B5E(12-15) B5F(12-15)
+
+                    const __m512i rhs_mat_014589CD_60_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_60, (_MM_PERM_ENUM)221); //B60(4-7) B61(4-7) B60(4-7) B61(4-7) B64(4-7) B65(4-7) B64(4-7) B65(4-7) B68(4-7) B69(4-7) B68(4-7) B69(4-7) B6C(4-7) B6D(4-7) B6C(4-7) B6D(4-7)
+                    const __m512i rhs_mat_2367ABEF_60_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_60, (_MM_PERM_ENUM)221); //B62(4-7) B63(4-7) B62(4-7) B63(4-7) B66(4-7) B67(4-7) B66(4-7) B67(4-7) B6A(4-7) B6B(4-7) B6A(4-7) B6B(4-7) B6E(4-7) B6F(4-7) B6E(4-7) B6F(4-7)
+
+                    const __m512i rhs_mat_014589CD_61_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_61, (_MM_PERM_ENUM)221); //B60(12-15) B61(12-15) B60(12-15) B61(12-15) B64(12-15) B65(12-15) B64(12-15) B65(12-15) B68(12-15) B69(12-15) B68(12-15) B69(12-15) B6C(12-15) B6D(12-15) B6C(12-15) B6D(12-15)
+                    const __m512i rhs_mat_2367ABEF_61_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_61, (_MM_PERM_ENUM)221); //B62(12-15) B63(12-15) B62(12-15) B63(12-15) B66(12-15) B67(12-15) B66(12-15) B67(12-15) B6A(12-15) B6B(12-15) B6A(12-15) B6B(12-15) B6E(12-15) B6F(12-15) B6E(12-15) B6F(12-15)
+
+                    const __m512i rhs_mat_014589CD_70_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_70, (_MM_PERM_ENUM)221); //B70(4-7) B71(4-7) B70(4-7) B71(4-7) B74(4-7) B75(4-7) B74(4-7) B75(4-7) B78(4-7) B79(4-7) B78(4-7) B79(4-7) B7C(4-7) B7D(4-7) B7C(4-7) B7D(4-7)
+                    const __m512i rhs_mat_2367ABEF_70_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_70, (_MM_PERM_ENUM)221); //B72(4-7) B73(4-7) B72(4-7) B73(4-7) B76(4-7) B77(4-7) B76(4-7) B77(4-7) B7A(4-7) B7B(4-7) B7A(4-7) B7B(4-7) B7E(4-7) B7F(4-7) B7E(4-7) B7F(4-7)
+
+                    const __m512i rhs_mat_014589CD_71_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_71, (_MM_PERM_ENUM)221); //B70(12-15) B71(12-15) B70(12-15) B71(12-15) B74(12-15) B75(12-15) B74(12-15) B75(12-15) B78(12-15) B79(12-15) B78(12-15) B79(12-15) B7C(12-15) B7D(12-15) B7C(12-15) B7D(12-15)
+                    const __m512i rhs_mat_2367ABEF_71_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_71, (_MM_PERM_ENUM)221); //B72(12-15) B73(12-15) B72(12-15) B73(12-15) B76(12-15) B77(12-15) B76(12-15) B77(12-15) B7A(12-15) B7B(12-15) B7A(12-15) B7B(12-15) B7E(12-15) B7F(12-15) B7E(12-15) B7F(12-15)
+
+                    //notation:superblock subblock
+                    //s00 m00  s01 m01   s10 m10  s11 m11  s20 m20  s21 m21   s30 m30  s31 m31  s40 m40  s41 m41   s50 m50  s51 m51  s60 m60  s61 m61   s70 m70  s71 m71
+
+                    const __m128i mins_and_scales_01_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + sb * 64));
+                    const __m128i mins_and_scales_23_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + 16 + sb * 64));
+                    const __m128i mins_and_scales_45_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + 32 + sb * 64));
+                    const __m128i mins_and_scales_67_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + 48 + sb * 64));
+
+                    const __m128i mins_and_scales_01_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + sb * 64));
+                    const __m128i mins_and_scales_23_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + 16 + sb * 64));
+                    const __m128i mins_and_scales_45_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + 32 + sb * 64));
+                    const __m128i mins_and_scales_67_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + 48 + sb * 64));
+
+                    // Combine mins and scales for sub-blocks: 0-1, 2-3, 4-5, 6-7 in the sb loop
+                    const __m256i mins_and_scales_01 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_01_0), mins_and_scales_01_1, 1);
+                    const __m256i mins_and_scales_23 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_23_0), mins_and_scales_23_1, 1);
+                    const __m256i mins_and_scales_45 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_45_0), mins_and_scales_45_1, 1);
+                    const __m256i mins_and_scales_67 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_67_0), mins_and_scales_67_1, 1);
+
+                    // Extract scales which is lower half from mins_and_scales
+                    const __m256i scales_01 = _mm256_and_si256(mins_and_scales_01, m4b);
+                    const __m256i scales_23 = _mm256_and_si256(mins_and_scales_23, m4b);
+                    const __m256i scales_45 = _mm256_and_si256(mins_and_scales_45, m4b);
+                    const __m256i scales_67 = _mm256_and_si256(mins_and_scales_67, m4b);
+
+                    // Extract mins which is upper half from mins_and_scales
+                    const __m512i mins_01 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_01, 4), m4b));
+                    const __m512i mins_23 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_23, 4), m4b));
+                    const __m512i mins_45 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_45, 4), m4b));
+                    const __m512i mins_67 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_67, 4), m4b));
+
+                    const __m512i scales_0 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_01,scalesmask1));
+                    const __m512i scales_1 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_01,scalesmask2));
+                    const __m512i scales_2 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_23,scalesmask1));
+                    const __m512i scales_3 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_23,scalesmask2));
+                    const __m512i scales_4 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_45,scalesmask1));
+                    const __m512i scales_5 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_45,scalesmask2));
+                    const __m512i scales_6 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_67,scalesmask1));
+                    const __m512i scales_7 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_67,scalesmask2));
+
+                    const __m512i scale_014589CD_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)238);
+
+                    const __m512i scale_014589CD_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)238);
+
+                    const __m512i scale_014589CD_2 = _mm512_shuffle_epi32(scales_2, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_2 = _mm512_shuffle_epi32(scales_2, (_MM_PERM_ENUM)238);
+
+                    const __m512i scale_014589CD_3 = _mm512_shuffle_epi32(scales_3, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_3 = _mm512_shuffle_epi32(scales_3, (_MM_PERM_ENUM)238);
+
+                    const __m512i scale_014589CD_4 = _mm512_shuffle_epi32(scales_4, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_4 = _mm512_shuffle_epi32(scales_4, (_MM_PERM_ENUM)238);
+
+                    const __m512i scale_014589CD_5 = _mm512_shuffle_epi32(scales_5, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_5 = _mm512_shuffle_epi32(scales_5, (_MM_PERM_ENUM)238);
+
+                    const __m512i scale_014589CD_6 = _mm512_shuffle_epi32(scales_6, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_6 = _mm512_shuffle_epi32(scales_6, (_MM_PERM_ENUM)238);
+
+                    const __m512i scale_014589CD_7 = _mm512_shuffle_epi32(scales_7, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_7 = _mm512_shuffle_epi32(scales_7, (_MM_PERM_ENUM)238);
+
+
+                    for (int rp = 0; rp < 4; rp++) {
+
+                        // Load the four block_q8_k quantized values interleaved with each other in chunks of eight bytes - A0,A1,A2,A3
+                        // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector
+                        __m256i lhs_mat_ymm_0123_00 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 512 * sb)));
+                        __m256i lhs_mat_ymm_01_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 0);
+                        __m256i lhs_mat_ymm_23_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 17);
+                        __m256i lhs_mat_ymm_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 32 + 512 * sb)));
+                        __m256i lhs_mat_ymm_01_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 0);
+                        __m256i lhs_mat_ymm_23_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 17);
+                        __m256i lhs_mat_ymm_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 64 + 512 * sb)));
+                        __m256i lhs_mat_ymm_01_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 0);
+                        __m256i lhs_mat_ymm_23_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 17);
+                        __m256i lhs_mat_ymm_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 96 + 512 * sb)));
+                        __m256i lhs_mat_ymm_01_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 0);
+                        __m256i lhs_mat_ymm_23_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 17);
+                        __m256i lhs_mat_ymm_0123_20 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 128 + 512 * sb)));
+                        __m256i lhs_mat_ymm_01_20 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_20, lhs_mat_ymm_0123_20, 0);
+                        __m256i lhs_mat_ymm_23_20 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_20, lhs_mat_ymm_0123_20, 17);
+                        __m256i lhs_mat_ymm_0123_21 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 160 + 512 * sb)));
+                        __m256i lhs_mat_ymm_01_21 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_21, lhs_mat_ymm_0123_21, 0);
+                        __m256i lhs_mat_ymm_23_21 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_21, lhs_mat_ymm_0123_21, 17);
+                        __m256i lhs_mat_ymm_0123_30 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 192 + 512 * sb)));
+                        __m256i lhs_mat_ymm_01_30 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_30, lhs_mat_ymm_0123_30, 0);
+                        __m256i lhs_mat_ymm_23_30 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_30, lhs_mat_ymm_0123_30, 17);
+                        __m256i lhs_mat_ymm_0123_31 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 224 + 512 * sb)));
+                        __m256i lhs_mat_ymm_01_31 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_31, lhs_mat_ymm_0123_31, 0);
+                        __m256i lhs_mat_ymm_23_31 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_31, lhs_mat_ymm_0123_31, 17);
+
+                        __m256i lhs_mat_ymm_0123_40 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 256 + 512 * sb)));
+                        __m256i lhs_mat_ymm_01_40 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_40, lhs_mat_ymm_0123_40, 0);
+                        __m256i lhs_mat_ymm_23_40 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_40, lhs_mat_ymm_0123_40, 17);
+                        __m256i lhs_mat_ymm_0123_41 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 288 + 512 * sb)));
+                        __m256i lhs_mat_ymm_01_41 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_41, lhs_mat_ymm_0123_41, 0);
+                        __m256i lhs_mat_ymm_23_41 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_41, lhs_mat_ymm_0123_41, 17);
+                        __m256i lhs_mat_ymm_0123_50 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 320 + 512 * sb)));
+                        __m256i lhs_mat_ymm_01_50 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_50, lhs_mat_ymm_0123_50, 0);
+                        __m256i lhs_mat_ymm_23_50 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_50, lhs_mat_ymm_0123_50, 17);
+                        __m256i lhs_mat_ymm_0123_51 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 352 + 512 * sb)));
+                        __m256i lhs_mat_ymm_01_51 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_51, lhs_mat_ymm_0123_51, 0);
+                        __m256i lhs_mat_ymm_23_51 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_51, lhs_mat_ymm_0123_51, 17);
+                        __m256i lhs_mat_ymm_0123_60 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 384 + 512 * sb)));
+                        __m256i lhs_mat_ymm_01_60 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_60, lhs_mat_ymm_0123_60, 0);
+                        __m256i lhs_mat_ymm_23_60 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_60, lhs_mat_ymm_0123_60, 17);
+                        __m256i lhs_mat_ymm_0123_61 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 416 + 512 * sb)));
+                        __m256i lhs_mat_ymm_01_61 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_61, lhs_mat_ymm_0123_61, 0);
+                        __m256i lhs_mat_ymm_23_61 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_61, lhs_mat_ymm_0123_61, 17);
+                        __m256i lhs_mat_ymm_0123_70 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 448 + 512 * sb)));
+                        __m256i lhs_mat_ymm_01_70 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_70, lhs_mat_ymm_0123_70, 0);
+                        __m256i lhs_mat_ymm_23_70 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_70, lhs_mat_ymm_0123_70, 17);
+                        __m256i lhs_mat_ymm_0123_71 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 480 + 512 * sb)));
+                        __m256i lhs_mat_ymm_01_71 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_71, lhs_mat_ymm_0123_71, 0);
+                        __m256i lhs_mat_ymm_23_71 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_71, lhs_mat_ymm_0123_71, 17);
+
+
+                        __m512i lhs_mat_01_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_00), lhs_mat_ymm_01_00, 1);
+                        __m512i lhs_mat_23_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_00), lhs_mat_ymm_23_00, 1);
+                        __m512i lhs_mat_01_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_01), lhs_mat_ymm_01_01, 1);
+                        __m512i lhs_mat_23_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_01), lhs_mat_ymm_23_01, 1);
+
+                        __m512i lhs_mat_01_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_10), lhs_mat_ymm_01_10, 1);
+                        __m512i lhs_mat_23_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_10), lhs_mat_ymm_23_10, 1);
+                        __m512i lhs_mat_01_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_11), lhs_mat_ymm_01_11, 1);
+                        __m512i lhs_mat_23_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_11), lhs_mat_ymm_23_11, 1);
+
+                        __m512i lhs_mat_01_20 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_20), lhs_mat_ymm_01_20, 1);
+                        __m512i lhs_mat_23_20 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_20), lhs_mat_ymm_23_20, 1);
+                        __m512i lhs_mat_01_21 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_21), lhs_mat_ymm_01_21, 1);
+                        __m512i lhs_mat_23_21 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_21), lhs_mat_ymm_23_21, 1);
+
+                        __m512i lhs_mat_01_30 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_30), lhs_mat_ymm_01_30, 1);
+                        __m512i lhs_mat_23_30 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_30), lhs_mat_ymm_23_30, 1);
+                        __m512i lhs_mat_01_31 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_31), lhs_mat_ymm_01_31, 1);
+                        __m512i lhs_mat_23_31 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_31), lhs_mat_ymm_23_31, 1);
+
+                        __m512i lhs_mat_01_40 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_40), lhs_mat_ymm_01_40, 1);
+                        __m512i lhs_mat_23_40 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_40), lhs_mat_ymm_23_40, 1);
+                        __m512i lhs_mat_01_41 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_41), lhs_mat_ymm_01_41, 1);
+                        __m512i lhs_mat_23_41 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_41), lhs_mat_ymm_23_41, 1);
+
+                        __m512i lhs_mat_01_50 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_50), lhs_mat_ymm_01_50, 1);
+                        __m512i lhs_mat_23_50 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_50), lhs_mat_ymm_23_50, 1);
+                        __m512i lhs_mat_01_51 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_51), lhs_mat_ymm_01_51, 1);
+                        __m512i lhs_mat_23_51 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_51), lhs_mat_ymm_23_51, 1);
+
+                        __m512i lhs_mat_01_60 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_60), lhs_mat_ymm_01_60, 1);
+                        __m512i lhs_mat_23_60 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_60), lhs_mat_ymm_23_60, 1);
+                        __m512i lhs_mat_01_61 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_61), lhs_mat_ymm_01_61, 1);
+                        __m512i lhs_mat_23_61 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_61), lhs_mat_ymm_23_61, 1);
+
+                        __m512i lhs_mat_01_70 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_70), lhs_mat_ymm_01_70, 1);
+                        __m512i lhs_mat_23_70 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_70), lhs_mat_ymm_23_70, 1);
+                        __m512i lhs_mat_01_71 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_71), lhs_mat_ymm_01_71, 1);
+                        __m512i lhs_mat_23_71 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_71), lhs_mat_ymm_23_71, 1);
+
+                        // Bsums are loaded for the different Q8_K blocks
+                        __m128i lhs_raw_bsums_01_0123 = _mm_loadu_si128((const __m128i *)((a_ptrs[rp][b].bsums + 32 * sb)));
+                        __m128i lhs_raw_bsums_23_0123 = _mm_loadu_si128((const __m128i *)(a_ptrs[rp][b].bsums + 8 + 32 * sb));
+                        __m128i lhs_raw_bsums_01_4567 = _mm_loadu_si128((const __m128i *)((a_ptrs[rp][b].bsums + 16 + 32 * sb)));
+                        __m128i lhs_raw_bsums_23_4567 = _mm_loadu_si128((const __m128i *)(a_ptrs[rp][b].bsums + 24 + 32 * sb));
+
+                        __m256i lhs_bsums_ymm_01_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_01_0123), lhs_raw_bsums_01_0123, 1);
+                        __m512i lhs_bsums_01_0123 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_ymm_01_0123), lhs_bsums_ymm_01_0123, 1);
+                        __m256i lhs_bsums_ymm_23_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_0123), lhs_raw_bsums_23_0123, 1);
+                        __m512i lhs_bsums_23_0123 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_ymm_23_0123), lhs_bsums_ymm_23_0123, 1);                        __m256i lhs_bsums_ymm_01_4567 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_01_4567), lhs_raw_bsums_01_4567, 1);
+                        __m512i lhs_bsums_01_4567 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_ymm_01_4567), lhs_bsums_ymm_01_4567, 1);
+                        __m256i lhs_bsums_ymm_23_4567 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_4567), lhs_raw_bsums_23_4567, 1);
+                        __m512i lhs_bsums_23_4567 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_ymm_23_4567), lhs_bsums_ymm_23_4567, 1);
+
+                        // Shuffle pattern one - left side input
+                        const __m512i lhs_mat_01_00_sp1 = _mm512_shuffle_epi32(lhs_mat_01_00, (_MM_PERM_ENUM)160); //A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3)
+                        const __m512i lhs_mat_23_00_sp1 = _mm512_shuffle_epi32(lhs_mat_23_00, (_MM_PERM_ENUM)160); //A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3)
+
+                        const __m512i lhs_mat_01_01_sp1 = _mm512_shuffle_epi32(lhs_mat_01_01, (_MM_PERM_ENUM)160); //A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11)
+                        const __m512i lhs_mat_23_01_sp1 = _mm512_shuffle_epi32(lhs_mat_23_01, (_MM_PERM_ENUM)160); //A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11)
+
+                        const __m512i lhs_mat_01_10_sp1 = _mm512_shuffle_epi32(lhs_mat_01_10, (_MM_PERM_ENUM)160); //A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3)
+                        const __m512i lhs_mat_23_10_sp1 = _mm512_shuffle_epi32(lhs_mat_23_10, (_MM_PERM_ENUM)160); //A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3)
+
+                        const __m512i lhs_mat_01_11_sp1 = _mm512_shuffle_epi32(lhs_mat_01_11, (_MM_PERM_ENUM)160); //A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11)
+                        const __m512i lhs_mat_23_11_sp1 = _mm512_shuffle_epi32(lhs_mat_23_11, (_MM_PERM_ENUM)160); //A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11)
+
+                        const __m512i lhs_mat_01_20_sp1 = _mm512_shuffle_epi32(lhs_mat_01_20, (_MM_PERM_ENUM)160); //A20(0-3) A20(0-3) A21(0-3) A21(0-3) A20(0-3) A20(0-3) A21(0-3) A21(0-3) A20(0-3) A20(0-3) A21(0-3) A21(0-3) A20(0-3) A20(0-3) A21(0-3) A21(0-3)
+                        const __m512i lhs_mat_23_20_sp1 = _mm512_shuffle_epi32(lhs_mat_23_20, (_MM_PERM_ENUM)160); //A22(0-3) A22(0-3) A23(0-3) A23(0-3) A22(0-3) A22(0-3) A23(0-3) A23(0-3) A22(0-3) A22(0-3) A23(0-3) A23(0-3) A22(0-3) A22(0-3) A23(0-3) A23(0-3)
+
+                        const __m512i lhs_mat_01_21_sp1 = _mm512_shuffle_epi32(lhs_mat_01_21, (_MM_PERM_ENUM)160); //A20(8-11) A20(8-11) A21(8-11) A21(8-11) A20(8-11) A20(8-11) A21(8-11) A21(8-11) A20(8-11) A20(8-11) A21(8-11) A21(8-11) A20(8-11) A20(8-11) A21(8-11) A21(8-11)
+                        const __m512i lhs_mat_23_21_sp1 = _mm512_shuffle_epi32(lhs_mat_23_21, (_MM_PERM_ENUM)160); //A22(8-11) A22(8-11) A23(8-11) A23(8-11) A22(8-11) A22(8-11) A23(8-11) A23(8-11) A22(8-11) A22(8-11) A23(8-11) A23(8-11) A22(8-11) A22(8-11) A23(8-11) A23(8-11)
+
+                        const __m512i lhs_mat_01_30_sp1 = _mm512_shuffle_epi32(lhs_mat_01_30, (_MM_PERM_ENUM)160); //A30(0-3) A30(0-3) A31(0-3) A31(0-3) A30(0-3) A30(0-3) A31(0-3) A31(0-3) A30(0-3) A30(0-3) A31(0-3) A31(0-3) A30(0-3) A30(0-3) A31(0-3) A31(0-3)
+                        const __m512i lhs_mat_23_30_sp1 = _mm512_shuffle_epi32(lhs_mat_23_30, (_MM_PERM_ENUM)160); //A32(0-3) A32(0-3) A33(0-3) A33(0-3) A32(0-3) A32(0-3) A33(0-3) A33(0-3) A32(0-3) A32(0-3) A33(0-3) A33(0-3) A32(0-3) A32(0-3) A33(0-3) A33(0-3)
+
+                        const __m512i lhs_mat_01_31_sp1 = _mm512_shuffle_epi32(lhs_mat_01_31, (_MM_PERM_ENUM)160); //A30(8-11) A30(8-11) A31(8-11) A31(8-11) A30(8-11) A30(8-11) A31(8-11) A31(8-11) A30(8-11) A30(8-11) A31(8-11) A31(8-11) A30(8-11) A30(8-11) A31(8-11) A31(8-11)
+                        const __m512i lhs_mat_23_31_sp1 = _mm512_shuffle_epi32(lhs_mat_23_31, (_MM_PERM_ENUM)160); //A32(8-11) A32(8-11) A33(8-11) A33(8-11) A32(8-11) A32(8-11) A33(8-11) A33(8-11) A32(8-11) A32(8-11) A33(8-11) A33(8-11) A32(8-11) A32(8-11) A33(8-11) A33(8-11)
+
+                        const __m512i lhs_mat_01_40_sp1 = _mm512_shuffle_epi32(lhs_mat_01_40, (_MM_PERM_ENUM)160); //A40(0-3) A40(0-3) A41(0-3) A41(0-3) A40(0-3) A40(0-3) A41(0-3) A41(0-3) A40(0-3) A40(0-3) A41(0-3) A41(0-3) A40(0-3) A40(0-3) A41(0-3) A41(0-3)
+                        const __m512i lhs_mat_23_40_sp1 = _mm512_shuffle_epi32(lhs_mat_23_40, (_MM_PERM_ENUM)160); //A42(0-3) A42(0-3) A43(0-3) A43(0-3) A42(0-3) A42(0-3) A43(0-3) A43(0-3) A42(0-3) A42(0-3) A43(0-3) A43(0-3) A42(0-3) A42(0-3) A43(0-3) A43(0-3)
+
+                        const __m512i lhs_mat_01_41_sp1 = _mm512_shuffle_epi32(lhs_mat_01_41, (_MM_PERM_ENUM)160); //A40(8-11) A40(8-11) A41(8-11) A41(8-11) A40(8-11) A40(8-11) A41(8-11) A41(8-11) A40(8-11) A40(8-11) A41(8-11) A41(8-11) A40(8-11) A40(8-11) A41(8-11) A41(8-11)
+                        const __m512i lhs_mat_23_41_sp1 = _mm512_shuffle_epi32(lhs_mat_23_41, (_MM_PERM_ENUM)160); //A42(8-11) A42(8-11) A43(8-11) A43(8-11) A42(8-11) A42(8-11) A43(8-11) A43(8-11) A42(8-11) A42(8-11) A43(8-11) A43(8-11) A42(8-11) A42(8-11) A43(8-11) A43(8-11)
+
+                        const __m512i lhs_mat_01_50_sp1 = _mm512_shuffle_epi32(lhs_mat_01_50, (_MM_PERM_ENUM)160); //A50(0-3) A50(0-3) A51(0-3) A51(0-3) A50(0-3) A50(0-3) A51(0-3) A51(0-3) A50(0-3) A50(0-3) A51(0-3) A51(0-3) A50(0-3) A50(0-3) A51(0-3) A51(0-3)
+                        const __m512i lhs_mat_23_50_sp1 = _mm512_shuffle_epi32(lhs_mat_23_50, (_MM_PERM_ENUM)160); //A52(0-3) A52(0-3) A53(0-3) A53(0-3) A52(0-3) A52(0-3) A53(0-3) A53(0-3) A52(0-3) A52(0-3) A53(0-3) A53(0-3) A52(0-3) A52(0-3) A53(0-3) A53(0-3)
+
+                        const __m512i lhs_mat_01_51_sp1 = _mm512_shuffle_epi32(lhs_mat_01_51, (_MM_PERM_ENUM)160); //A50(8-11) A50(8-11) A51(8-11) A51(8-11) A50(8-11) A50(8-11) A51(8-11) A51(8-11) A50(8-11) A50(8-11) A51(8-11) A51(8-11) A50(8-11) A50(8-11) A51(8-11) A51(8-11)
+                        const __m512i lhs_mat_23_51_sp1 = _mm512_shuffle_epi32(lhs_mat_23_51, (_MM_PERM_ENUM)160); //A52(8-11) A52(8-11) A53(8-11) A53(8-11) A52(8-11) A52(8-11) A53(8-11) A53(8-11) A52(8-11) A52(8-11) A53(8-11) A53(8-11) A52(8-11) A52(8-11) A53(8-11) A53(8-11)
+
+                        const __m512i lhs_mat_01_60_sp1 = _mm512_shuffle_epi32(lhs_mat_01_60, (_MM_PERM_ENUM)160); //A60(0-3) A60(0-3) A61(0-3) A61(0-3) A60(0-3) A60(0-3) A61(0-3) A61(0-3) A60(0-3) A60(0-3) A61(0-3) A61(0-3) A60(0-3) A60(0-3) A61(0-3) A61(0-3)
+                        const __m512i lhs_mat_23_60_sp1 = _mm512_shuffle_epi32(lhs_mat_23_60, (_MM_PERM_ENUM)160); //A62(0-3) A62(0-3) A63(0-3) A63(0-3) A62(0-3) A62(0-3) A63(0-3) A63(0-3) A62(0-3) A62(0-3) A63(0-3) A63(0-3) A62(0-3) A62(0-3) A63(0-3) A63(0-3)
+
+                        const __m512i lhs_mat_01_61_sp1 = _mm512_shuffle_epi32(lhs_mat_01_61, (_MM_PERM_ENUM)160); //A60(8-11) A60(8-11) A61(8-11) A61(8-11) A60(8-11) A60(8-11) A61(8-11) A61(8-11) A60(8-11) A60(8-11) A61(8-11) A61(8-11) A60(8-11) A60(8-11) A61(8-11) A61(8-11)
+                        const __m512i lhs_mat_23_61_sp1 = _mm512_shuffle_epi32(lhs_mat_23_61, (_MM_PERM_ENUM)160); //A62(8-11) A62(8-11) A63(8-11) A63(8-11) A62(8-11) A62(8-11) A63(8-11) A63(8-11) A62(8-11) A62(8-11) A63(8-11) A63(8-11) A62(8-11) A62(8-11) A63(8-11) A63(8-11)
+
+                        const __m512i lhs_mat_01_70_sp1 = _mm512_shuffle_epi32(lhs_mat_01_70, (_MM_PERM_ENUM)160); //A70(0-3) A70(0-3) A71(0-3) A71(0-3) A70(0-3) A70(0-3) A71(0-3) A71(0-3) A70(0-3) A70(0-3) A71(0-3) A71(0-3) A70(0-3) A70(0-3) A71(0-3) A71(0-3)
+                        const __m512i lhs_mat_23_70_sp1 = _mm512_shuffle_epi32(lhs_mat_23_70, (_MM_PERM_ENUM)160); //A72(0-3) A72(0-3) A73(0-3) A73(0-3) A72(0-3) A72(0-3) A73(0-3) A73(0-3) A72(0-3) A72(0-3) A73(0-3) A73(0-3) A72(0-3) A72(0-3) A73(0-3) A73(0-3)
+
+                        const __m512i lhs_mat_01_71_sp1 = _mm512_shuffle_epi32(lhs_mat_01_71, (_MM_PERM_ENUM)160); //A70(8-11) A70(8-11) A71(8-11) A71(8-11) A70(8-11) A70(8-11) A71(8-11) A71(8-11) A70(8-11) A70(8-11) A71(8-11) A71(8-11) A70(8-11) A70(8-11) A71(8-11) A71(8-11)
+                        const __m512i lhs_mat_23_71_sp1 = _mm512_shuffle_epi32(lhs_mat_23_71, (_MM_PERM_ENUM)160); //A72(8-11) A72(8-11) A73(8-11) A73(8-11) A72(8-11) A72(8-11) A73(8-11) A73(8-11) A72(8-11) A72(8-11) A73(8-11) A73(8-11) A72(8-11) A72(8-11) A73(8-11) A73(8-11)
+
+                        const __m512i lhs_mat_01_00_sp2 = _mm512_shuffle_epi32(lhs_mat_01_00, (_MM_PERM_ENUM)245); //A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7)
+                        const __m512i lhs_mat_23_00_sp2 = _mm512_shuffle_epi32(lhs_mat_23_00, (_MM_PERM_ENUM)245); //A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7)
+
+                        const __m512i lhs_mat_01_01_sp2 = _mm512_shuffle_epi32(lhs_mat_01_01, (_MM_PERM_ENUM)245); //A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15)
+                        const __m512i lhs_mat_23_01_sp2 = _mm512_shuffle_epi32(lhs_mat_23_01, (_MM_PERM_ENUM)245); //A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15)
+
+                        const __m512i lhs_mat_01_10_sp2 = _mm512_shuffle_epi32(lhs_mat_01_10, (_MM_PERM_ENUM)245); //A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7)
+                        const __m512i lhs_mat_23_10_sp2 = _mm512_shuffle_epi32(lhs_mat_23_10, (_MM_PERM_ENUM)245); //A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7)
+
+                        const __m512i lhs_mat_01_11_sp2 = _mm512_shuffle_epi32(lhs_mat_01_11, (_MM_PERM_ENUM)245); //A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15)
+                        const __m512i lhs_mat_23_11_sp2 = _mm512_shuffle_epi32(lhs_mat_23_11, (_MM_PERM_ENUM)245); //A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15)
+
+                        const __m512i lhs_mat_01_20_sp2 = _mm512_shuffle_epi32(lhs_mat_01_20, (_MM_PERM_ENUM)245); //A20(4-7) A20(4-7) A21(4-7) A21(4-7) A20(4-7) A20(4-7) A21(4-7) A21(4-7) A20(4-7) A20(4-7) A21(4-7) A21(4-7) A20(4-7) A20(4-7) A21(4-7) A21(4-7)
+                        const __m512i lhs_mat_23_20_sp2 = _mm512_shuffle_epi32(lhs_mat_23_20, (_MM_PERM_ENUM)245); //A22(4-7) A22(4-7) A23(4-7) A23(4-7) A22(4-7) A22(4-7) A23(4-7) A23(4-7) A22(4-7) A22(4-7) A23(4-7) A23(4-7) A22(4-7) A22(4-7) A23(4-7) A23(4-7)
+
+                        const __m512i lhs_mat_01_21_sp2 = _mm512_shuffle_epi32(lhs_mat_01_21, (_MM_PERM_ENUM)245); //A20(12-15) A20(12-15) A21(12-15) A21(12-15) A20(12-15) A20(12-15) A21(12-15) A21(12-15) A20(12-15) A20(12-15) A21(12-15) A21(12-15) A20(12-15) A20(12-15) A21(12-15) A21(12-15)
+                        const __m512i lhs_mat_23_21_sp2 = _mm512_shuffle_epi32(lhs_mat_23_21, (_MM_PERM_ENUM)245); //A22(12-15) A22(12-15) A23(12-15) A23(12-15) A22(12-15) A22(12-15) A23(12-15) A23(12-15) A22(12-15) A22(12-15) A23(12-15) A23(12-15) A22(12-15) A22(12-15) A23(12-15) A23(12-15)
+
+                        const __m512i lhs_mat_01_30_sp2 = _mm512_shuffle_epi32(lhs_mat_01_30, (_MM_PERM_ENUM)245); //A30(4-7) A30(4-7) A31(4-7) A31(4-7) A30(4-7) A30(4-7) A31(4-7) A31(4-7) A30(4-7) A30(4-7) A31(4-7) A31(4-7) A30(4-7) A30(4-7) A31(4-7) A31(4-7)
+                        const __m512i lhs_mat_23_30_sp2 = _mm512_shuffle_epi32(lhs_mat_23_30, (_MM_PERM_ENUM)245); //A32(4-7) A32(4-7) A33(4-7) A33(4-7) A32(4-7) A32(4-7) A33(4-7) A33(4-7) A32(4-7) A32(4-7) A33(4-7) A33(4-7) A32(4-7) A32(4-7) A33(4-7) A33(4-7)
+
+                        const __m512i lhs_mat_01_31_sp2 = _mm512_shuffle_epi32(lhs_mat_01_31, (_MM_PERM_ENUM)245); //A30(12-15) A30(12-15) A31(12-15) A31(12-15) A30(12-15) A30(12-15) A31(12-15) A31(12-15) A30(12-15) A30(12-15) A31(12-15) A31(12-15) A30(12-15) A30(12-15) A31(12-15) A31(12-15)
+                        const __m512i lhs_mat_23_31_sp2 = _mm512_shuffle_epi32(lhs_mat_23_31, (_MM_PERM_ENUM)245); //A32(12-15) A32(12-15) A33(12-15) A33(12-15) A32(12-15) A32(12-15) A33(12-15) A33(12-15) A32(12-15) A32(12-15) A33(12-15) A33(12-15) A32(12-15) A32(12-15) A33(12-15) A33(12-15)
+
+                        const __m512i lhs_mat_01_40_sp2 = _mm512_shuffle_epi32(lhs_mat_01_40, (_MM_PERM_ENUM)245); //A40(4-7) A40(4-7) A41(4-7) A41(4-7) A40(4-7) A40(4-7) A41(4-7) A41(4-7) A40(4-7) A40(4-7) A41(4-7) A41(4-7) A40(4-7) A40(4-7) A41(4-7) A41(4-7)
+                        const __m512i lhs_mat_23_40_sp2 = _mm512_shuffle_epi32(lhs_mat_23_40, (_MM_PERM_ENUM)245); //A42(4-7) A42(4-7) A43(4-7) A43(4-7) A42(4-7) A42(4-7) A43(4-7) A43(4-7) A42(4-7) A42(4-7) A43(4-7) A43(4-7) A42(4-7) A42(4-7) A43(4-7) A43(4-7)
+
+                        const __m512i lhs_mat_01_41_sp2 = _mm512_shuffle_epi32(lhs_mat_01_41, (_MM_PERM_ENUM)245); //A40(12-15) A40(12-15) A41(12-15) A41(12-15) A40(12-15) A40(12-15) A41(12-15) A41(12-15) A40(12-15) A40(12-15) A41(12-15) A41(12-15) A40(12-15) A40(12-15) A41(12-15) A41(12-15)
+                        const __m512i lhs_mat_23_41_sp2 = _mm512_shuffle_epi32(lhs_mat_23_41, (_MM_PERM_ENUM)245); //A42(12-15) A42(12-15) A43(12-15) A43(12-15) A42(12-15) A42(12-15) A43(12-15) A43(12-15) A42(12-15) A42(12-15) A43(12-15) A43(12-15) A42(12-15) A42(12-15) A43(12-15) A43(12-15)
+
+                        const __m512i lhs_mat_01_50_sp2 = _mm512_shuffle_epi32(lhs_mat_01_50, (_MM_PERM_ENUM)245); //A50(4-7) A50(4-7) A51(4-7) A51(4-7) A50(4-7) A50(4-7) A51(4-7) A51(4-7) A50(4-7) A50(4-7) A51(4-7) A51(4-7) A50(4-7) A50(4-7) A51(4-7) A51(4-7)
+                        const __m512i lhs_mat_23_50_sp2 = _mm512_shuffle_epi32(lhs_mat_23_50, (_MM_PERM_ENUM)245); //A52(4-7) A52(4-7) A53(4-7) A53(4-7) A52(4-7) A52(4-7) A53(4-7) A53(4-7) A52(4-7) A52(4-7) A53(4-7) A53(4-7) A52(4-7) A52(4-7) A53(4-7) A53(4-7)
+
+                        const __m512i lhs_mat_01_51_sp2 = _mm512_shuffle_epi32(lhs_mat_01_51, (_MM_PERM_ENUM)245); //A50(12-15) A50(12-15) A51(12-15) A51(12-15) A50(12-15) A50(12-15) A51(12-15) A51(12-15) A50(12-15) A50(12-15) A51(12-15) A51(12-15) A50(12-15) A50(12-15) A51(12-15) A51(12-15)
+                        const __m512i lhs_mat_23_51_sp2 = _mm512_shuffle_epi32(lhs_mat_23_51, (_MM_PERM_ENUM)245); //A52(12-15) A52(12-15) A53(12-15) A53(12-15) A52(12-15) A52(12-15) A53(12-15) A53(12-15) A52(12-15) A52(12-15) A53(12-15) A53(12-15) A52(12-15) A52(12-15) A53(12-15) A53(12-15)
+
+                        const __m512i lhs_mat_01_60_sp2 = _mm512_shuffle_epi32(lhs_mat_01_60, (_MM_PERM_ENUM)245); //A60(4-7) A60(4-7) A61(4-7) A61(4-7) A60(4-7) A60(4-7) A61(4-7) A61(4-7) A60(4-7) A60(4-7) A61(4-7) A61(4-7) A60(4-7) A60(4-7) A61(4-7) A61(4-7)
+                        const __m512i lhs_mat_23_60_sp2 = _mm512_shuffle_epi32(lhs_mat_23_60, (_MM_PERM_ENUM)245); //A62(4-7) A62(4-7) A63(4-7) A63(4-7) A62(4-7) A62(4-7) A63(4-7) A63(4-7) A62(4-7) A62(4-7) A63(4-7) A63(4-7) A62(4-7) A62(4-7) A63(4-7) A63(4-7)
+
+                        const __m512i lhs_mat_01_61_sp2 = _mm512_shuffle_epi32(lhs_mat_01_61, (_MM_PERM_ENUM)245); //A60(12-15) A60(12-15) A61(12-15) A61(12-15) A60(12-15) A60(12-15) A61(12-15) A61(12-15) A60(12-15) A60(12-15) A61(12-15) A61(12-15) A60(12-15) A60(12-15) A61(12-15) A61(12-15)
+                        const __m512i lhs_mat_23_61_sp2 = _mm512_shuffle_epi32(lhs_mat_23_61, (_MM_PERM_ENUM)245); //A62(12-15) A62(12-15) A63(12-15) A63(12-15) A62(12-15) A62(12-15) A63(12-15) A63(12-15) A62(12-15) A62(12-15) A63(12-15) A63(12-15) A62(12-15) A62(12-15) A63(12-15) A63(12-15)
+
+                        const __m512i lhs_mat_01_70_sp2 = _mm512_shuffle_epi32(lhs_mat_01_70, (_MM_PERM_ENUM)245); //A70(4-7) A70(4-7) A71(4-7) A71(4-7) A70(4-7) A70(4-7) A71(4-7) A71(4-7) A70(4-7) A70(4-7) A71(4-7) A71(4-7) A70(4-7) A70(4-7) A71(4-7) A71(4-7)
+                        const __m512i lhs_mat_23_70_sp2 = _mm512_shuffle_epi32(lhs_mat_23_70, (_MM_PERM_ENUM)245); //A72(4-7) A72(4-7) A73(4-7) A73(4-7) A72(4-7) A72(4-7) A73(4-7) A73(4-7) A72(4-7) A72(4-7) A73(4-7) A73(4-7) A72(4-7) A72(4-7) A73(4-7) A73(4-7)
+
+                        const __m512i lhs_mat_01_71_sp2 = _mm512_shuffle_epi32(lhs_mat_01_71, (_MM_PERM_ENUM)245); //A70(12-15) A70(12-15) A71(12-15) A71(12-15) A70(12-15) A70(12-15) A71(12-15) A71(12-15) A70(12-15) A70(12-15) A71(12-15) A71(12-15) A70(12-15) A70(12-15) A71(12-15) A71(12-15)
+                        const __m512i lhs_mat_23_71_sp2 = _mm512_shuffle_epi32(lhs_mat_23_71, (_MM_PERM_ENUM)245); //A72(12-15) A72(12-15) A73(12-15) A73(12-15) A72(12-15) A72(12-15) A73(12-15) A73(12-15) A72(12-15) A72(12-15) A73(12-15) A73(12-15) A72(12-15) A72(12-15) A73(12-15) A73(12-15)
+
+                        // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+                        __m512i iacc_mat_00_0_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_01_00_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_01_01_sp1));
+                        __m512i iacc_mat_01_0_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_01_00_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_01_01_sp1));
+
+                        __m512i iacc_mat_10_0_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_23_00_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_23_01_sp1));
+                        __m512i iacc_mat_11_0_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_23_00_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_23_01_sp1));
+
+                        __m512i iacc_mat_00_1_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_01_10_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_01_11_sp1));
+                        __m512i iacc_mat_01_1_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_01_10_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_01_11_sp1));
+
+                        __m512i iacc_mat_10_1_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_23_10_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_23_11_sp1));
+                        __m512i iacc_mat_11_1_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_23_10_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_23_11_sp1));
+
+                        __m512i iacc_mat_00_2_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_20_sp1, lhs_mat_01_20_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_21_sp1, lhs_mat_01_21_sp1));
+                        __m512i iacc_mat_01_2_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_20_sp1, lhs_mat_01_20_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_21_sp1, lhs_mat_01_21_sp1));
+
+                        __m512i iacc_mat_10_2_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_20_sp1, lhs_mat_23_20_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_21_sp1, lhs_mat_23_21_sp1));
+                        __m512i iacc_mat_11_2_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_20_sp1, lhs_mat_23_20_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_21_sp1, lhs_mat_23_21_sp1));
+
+                        __m512i iacc_mat_00_3_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_30_sp1, lhs_mat_01_30_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_31_sp1, lhs_mat_01_31_sp1));
+                        __m512i iacc_mat_01_3_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_30_sp1, lhs_mat_01_30_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_31_sp1, lhs_mat_01_31_sp1));
+
+                        __m512i iacc_mat_10_3_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_30_sp1, lhs_mat_23_30_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_31_sp1, lhs_mat_23_31_sp1));
+                        __m512i iacc_mat_11_3_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_30_sp1, lhs_mat_23_30_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_31_sp1, lhs_mat_23_31_sp1));
+
+                        __m512i iacc_mat_00_4_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_40_sp1, lhs_mat_01_40_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_41_sp1, lhs_mat_01_41_sp1));
+                        __m512i iacc_mat_01_4_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_40_sp1, lhs_mat_01_40_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_41_sp1, lhs_mat_01_41_sp1));
+
+                        __m512i iacc_mat_10_4_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_40_sp1, lhs_mat_23_40_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_41_sp1, lhs_mat_23_41_sp1));
+                        __m512i iacc_mat_11_4_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_40_sp1, lhs_mat_23_40_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_41_sp1, lhs_mat_23_41_sp1));
+
+                        __m512i iacc_mat_00_5_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_50_sp1, lhs_mat_01_50_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_51_sp1, lhs_mat_01_51_sp1));
+                        __m512i iacc_mat_01_5_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_50_sp1, lhs_mat_01_50_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_51_sp1, lhs_mat_01_51_sp1));
+
+                        __m512i iacc_mat_10_5_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_50_sp1, lhs_mat_23_50_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_51_sp1, lhs_mat_23_51_sp1));
+                        __m512i iacc_mat_11_5_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_50_sp1, lhs_mat_23_50_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_51_sp1, lhs_mat_23_51_sp1));
+
+                        __m512i iacc_mat_00_6_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_60_sp1, lhs_mat_01_60_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_61_sp1, lhs_mat_01_61_sp1));
+                        __m512i iacc_mat_01_6_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_60_sp1, lhs_mat_01_60_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_61_sp1, lhs_mat_01_61_sp1));
+
+                        __m512i iacc_mat_10_6_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_60_sp1, lhs_mat_23_60_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_61_sp1, lhs_mat_23_61_sp1));
+                        __m512i iacc_mat_11_6_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_60_sp1, lhs_mat_23_60_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_61_sp1, lhs_mat_23_61_sp1));
+
+                        __m512i iacc_mat_00_7_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_70_sp1, lhs_mat_01_70_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_71_sp1, lhs_mat_01_71_sp1));
+                        __m512i iacc_mat_01_7_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_70_sp1, lhs_mat_01_70_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_71_sp1, lhs_mat_01_71_sp1));
+
+                        __m512i iacc_mat_10_7_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_70_sp1, lhs_mat_23_70_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_71_sp1, lhs_mat_23_71_sp1));
+                        __m512i iacc_mat_11_7_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_70_sp1, lhs_mat_23_70_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_71_sp1, lhs_mat_23_71_sp1));
+
+
+                        __m512i iacc_mat_00_0_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_01_00_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_01_01_sp2));
+                        __m512i iacc_mat_01_0_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_01_00_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_01_01_sp2));
+
+                        __m512i iacc_mat_10_0_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_23_00_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_23_01_sp2));
+                        __m512i iacc_mat_11_0_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_23_00_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_23_01_sp2));
+
+                        __m512i iacc_mat_00_1_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_01_10_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_01_11_sp2));
+                        __m512i iacc_mat_01_1_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_01_10_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_01_11_sp2));
+
+                        __m512i iacc_mat_10_1_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_23_10_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_23_11_sp2));
+                        __m512i iacc_mat_11_1_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_23_10_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_23_11_sp2));
+
+                        __m512i iacc_mat_00_2_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_20_sp2, lhs_mat_01_20_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_21_sp2, lhs_mat_01_21_sp2));
+                        __m512i iacc_mat_01_2_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_20_sp2, lhs_mat_01_20_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_21_sp2, lhs_mat_01_21_sp2));
+
+                        __m512i iacc_mat_10_2_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_20_sp2, lhs_mat_23_20_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_21_sp2, lhs_mat_23_21_sp2));
+                        __m512i iacc_mat_11_2_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_20_sp2, lhs_mat_23_20_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_21_sp2, lhs_mat_23_21_sp2));
+
+                        __m512i iacc_mat_00_3_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_30_sp2, lhs_mat_01_30_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_31_sp2, lhs_mat_01_31_sp2));
+                        __m512i iacc_mat_01_3_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_30_sp2, lhs_mat_01_30_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_31_sp2, lhs_mat_01_31_sp2));
+
+                        __m512i iacc_mat_10_3_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_30_sp2, lhs_mat_23_30_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_31_sp2, lhs_mat_23_31_sp2));
+                        __m512i iacc_mat_11_3_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_30_sp2, lhs_mat_23_30_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_31_sp2, lhs_mat_23_31_sp2));
+
+                        __m512i iacc_mat_00_4_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_40_sp2, lhs_mat_01_40_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_41_sp2, lhs_mat_01_41_sp2));
+                        __m512i iacc_mat_01_4_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_40_sp2, lhs_mat_01_40_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_41_sp2, lhs_mat_01_41_sp2));
+
+                        __m512i iacc_mat_10_4_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_40_sp2, lhs_mat_23_40_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_41_sp2, lhs_mat_23_41_sp2));
+                        __m512i iacc_mat_11_4_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_40_sp2, lhs_mat_23_40_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_41_sp2, lhs_mat_23_41_sp2));
+
+                        __m512i iacc_mat_00_5_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_50_sp2, lhs_mat_01_50_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_51_sp2, lhs_mat_01_51_sp2));
+                        __m512i iacc_mat_01_5_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_50_sp2, lhs_mat_01_50_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_51_sp2, lhs_mat_01_51_sp2));
+
+                        __m512i iacc_mat_10_5_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_50_sp2, lhs_mat_23_50_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_51_sp2, lhs_mat_23_51_sp2));
+                        __m512i iacc_mat_11_5_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_50_sp2, lhs_mat_23_50_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_51_sp2, lhs_mat_23_51_sp2));
+
+                        __m512i iacc_mat_00_6_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_60_sp2, lhs_mat_01_60_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_61_sp2, lhs_mat_01_61_sp2));
+                        __m512i iacc_mat_01_6_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_60_sp2, lhs_mat_01_60_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_61_sp2, lhs_mat_01_61_sp2));
+
+                        __m512i iacc_mat_10_6_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_60_sp2, lhs_mat_23_60_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_61_sp2, lhs_mat_23_61_sp2));
+                        __m512i iacc_mat_11_6_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_60_sp2, lhs_mat_23_60_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_61_sp2, lhs_mat_23_61_sp2));
+
+                        __m512i iacc_mat_00_7_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_70_sp2, lhs_mat_01_70_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_71_sp2, lhs_mat_01_71_sp2));
+                        __m512i iacc_mat_01_7_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_70_sp2, lhs_mat_01_70_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_71_sp2, lhs_mat_01_71_sp2));
+
+                        __m512i iacc_mat_10_7_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_70_sp2, lhs_mat_23_70_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_71_sp2, lhs_mat_23_71_sp2));
+                        __m512i iacc_mat_11_7_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_70_sp2, lhs_mat_23_70_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_71_sp2, lhs_mat_23_71_sp2));
+
+                        // Combine results from both shuffle patterns for each output block
+                        __m512i iacc_mat_00_0 = _mm512_add_epi16(iacc_mat_00_0_sp1, iacc_mat_00_0_sp2);
+                        __m512i iacc_mat_01_0 = _mm512_add_epi16(iacc_mat_01_0_sp1, iacc_mat_01_0_sp2);
+                        __m512i iacc_mat_10_0 = _mm512_add_epi16(iacc_mat_10_0_sp1, iacc_mat_10_0_sp2);
+                        __m512i iacc_mat_11_0 = _mm512_add_epi16(iacc_mat_11_0_sp1, iacc_mat_11_0_sp2);
+
+                        __m512i iacc_mat_00_1 = _mm512_add_epi16(iacc_mat_00_1_sp1, iacc_mat_00_1_sp2);
+                        __m512i iacc_mat_01_1 = _mm512_add_epi16(iacc_mat_01_1_sp1, iacc_mat_01_1_sp2);
+                        __m512i iacc_mat_10_1 = _mm512_add_epi16(iacc_mat_10_1_sp1, iacc_mat_10_1_sp2);
+                        __m512i iacc_mat_11_1 = _mm512_add_epi16(iacc_mat_11_1_sp1, iacc_mat_11_1_sp2);
+
+                        __m512i iacc_mat_00_2 = _mm512_add_epi16(iacc_mat_00_2_sp1, iacc_mat_00_2_sp2);
+                        __m512i iacc_mat_01_2 = _mm512_add_epi16(iacc_mat_01_2_sp1, iacc_mat_01_2_sp2);
+                        __m512i iacc_mat_10_2 = _mm512_add_epi16(iacc_mat_10_2_sp1, iacc_mat_10_2_sp2);
+                        __m512i iacc_mat_11_2 = _mm512_add_epi16(iacc_mat_11_2_sp1, iacc_mat_11_2_sp2);
+
+                        __m512i iacc_mat_00_3 = _mm512_add_epi16(iacc_mat_00_3_sp1, iacc_mat_00_3_sp2);
+                        __m512i iacc_mat_01_3 = _mm512_add_epi16(iacc_mat_01_3_sp1, iacc_mat_01_3_sp2);
+                        __m512i iacc_mat_10_3 = _mm512_add_epi16(iacc_mat_10_3_sp1, iacc_mat_10_3_sp2);
+                        __m512i iacc_mat_11_3 = _mm512_add_epi16(iacc_mat_11_3_sp1, iacc_mat_11_3_sp2);
+
+                        __m512i iacc_mat_00_4 = _mm512_add_epi16(iacc_mat_00_4_sp1, iacc_mat_00_4_sp2);
+                        __m512i iacc_mat_01_4 = _mm512_add_epi16(iacc_mat_01_4_sp1, iacc_mat_01_4_sp2);
+                        __m512i iacc_mat_10_4 = _mm512_add_epi16(iacc_mat_10_4_sp1, iacc_mat_10_4_sp2);
+                        __m512i iacc_mat_11_4 = _mm512_add_epi16(iacc_mat_11_4_sp1, iacc_mat_11_4_sp2);
+
+                        __m512i iacc_mat_00_5 = _mm512_add_epi16(iacc_mat_00_5_sp1, iacc_mat_00_5_sp2);
+                        __m512i iacc_mat_01_5 = _mm512_add_epi16(iacc_mat_01_5_sp1, iacc_mat_01_5_sp2);
+                        __m512i iacc_mat_10_5 = _mm512_add_epi16(iacc_mat_10_5_sp1, iacc_mat_10_5_sp2);
+                        __m512i iacc_mat_11_5 = _mm512_add_epi16(iacc_mat_11_5_sp1, iacc_mat_11_5_sp2);
+
+                        __m512i iacc_mat_00_6 = _mm512_add_epi16(iacc_mat_00_6_sp1, iacc_mat_00_6_sp2);
+                        __m512i iacc_mat_01_6 = _mm512_add_epi16(iacc_mat_01_6_sp1, iacc_mat_01_6_sp2);
+                        __m512i iacc_mat_10_6 = _mm512_add_epi16(iacc_mat_10_6_sp1, iacc_mat_10_6_sp2);
+                        __m512i iacc_mat_11_6 = _mm512_add_epi16(iacc_mat_11_6_sp1, iacc_mat_11_6_sp2);
+
+                        __m512i iacc_mat_00_7 = _mm512_add_epi16(iacc_mat_00_7_sp1, iacc_mat_00_7_sp2);
+                        __m512i iacc_mat_01_7 = _mm512_add_epi16(iacc_mat_01_7_sp1, iacc_mat_01_7_sp2);
+                        __m512i iacc_mat_10_7 = _mm512_add_epi16(iacc_mat_10_7_sp1, iacc_mat_10_7_sp2);
+                        __m512i iacc_mat_11_7 = _mm512_add_epi16(iacc_mat_11_7_sp1, iacc_mat_11_7_sp2);
+
+                        // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+                        iacc_mat_00_0 = _mm512_madd_epi16(iacc_mat_00_0, scale_014589CD_0);
+                        iacc_mat_01_0 = _mm512_madd_epi16(iacc_mat_01_0, scale_2367ABEF_0);
+                        iacc_mat_10_0 = _mm512_madd_epi16(iacc_mat_10_0, scale_014589CD_0);
+                        iacc_mat_11_0 = _mm512_madd_epi16(iacc_mat_11_0, scale_2367ABEF_0);
+
+                        iacc_mat_00_1 = _mm512_madd_epi16(iacc_mat_00_1, scale_014589CD_1);
+                        iacc_mat_01_1 = _mm512_madd_epi16(iacc_mat_01_1, scale_2367ABEF_1);
+                        iacc_mat_10_1 = _mm512_madd_epi16(iacc_mat_10_1, scale_014589CD_1);
+                        iacc_mat_11_1 = _mm512_madd_epi16(iacc_mat_11_1, scale_2367ABEF_1);
+
+                        iacc_mat_00_2 = _mm512_madd_epi16(iacc_mat_00_2, scale_014589CD_2);
+                        iacc_mat_01_2 = _mm512_madd_epi16(iacc_mat_01_2, scale_2367ABEF_2);
+                        iacc_mat_10_2 = _mm512_madd_epi16(iacc_mat_10_2, scale_014589CD_2);
+                        iacc_mat_11_2 = _mm512_madd_epi16(iacc_mat_11_2, scale_2367ABEF_2);
+
+                        iacc_mat_00_3 = _mm512_madd_epi16(iacc_mat_00_3, scale_014589CD_3);
+                        iacc_mat_01_3 = _mm512_madd_epi16(iacc_mat_01_3, scale_2367ABEF_3);
+                        iacc_mat_10_3 = _mm512_madd_epi16(iacc_mat_10_3, scale_014589CD_3);
+                        iacc_mat_11_3 = _mm512_madd_epi16(iacc_mat_11_3, scale_2367ABEF_3);
+
+                        iacc_mat_00_4 = _mm512_madd_epi16(iacc_mat_00_4, scale_014589CD_4);
+                        iacc_mat_01_4 = _mm512_madd_epi16(iacc_mat_01_4, scale_2367ABEF_4);
+                        iacc_mat_10_4 = _mm512_madd_epi16(iacc_mat_10_4, scale_014589CD_4);
+                        iacc_mat_11_4 = _mm512_madd_epi16(iacc_mat_11_4, scale_2367ABEF_4);
+
+                        iacc_mat_00_5 = _mm512_madd_epi16(iacc_mat_00_5, scale_014589CD_5);
+                        iacc_mat_01_5 = _mm512_madd_epi16(iacc_mat_01_5, scale_2367ABEF_5);
+                        iacc_mat_10_5 = _mm512_madd_epi16(iacc_mat_10_5, scale_014589CD_5);
+                        iacc_mat_11_5 = _mm512_madd_epi16(iacc_mat_11_5, scale_2367ABEF_5);
+
+                        iacc_mat_00_6 = _mm512_madd_epi16(iacc_mat_00_6, scale_014589CD_6);
+                        iacc_mat_01_6 = _mm512_madd_epi16(iacc_mat_01_6, scale_2367ABEF_6);
+                        iacc_mat_10_6 = _mm512_madd_epi16(iacc_mat_10_6, scale_014589CD_6);
+                        iacc_mat_11_6 = _mm512_madd_epi16(iacc_mat_11_6, scale_2367ABEF_6);
+
+                        iacc_mat_00_7 = _mm512_madd_epi16(iacc_mat_00_7, scale_014589CD_7);
+                        iacc_mat_01_7 = _mm512_madd_epi16(iacc_mat_01_7, scale_2367ABEF_7);
+                        iacc_mat_10_7 = _mm512_madd_epi16(iacc_mat_10_7, scale_014589CD_7);
+                        iacc_mat_11_7 = _mm512_madd_epi16(iacc_mat_11_7, scale_2367ABEF_7);
+
+                        __m512i iacc_mat_00 = _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(iacc_mat_00_0, iacc_mat_00_1), _mm512_add_epi32(iacc_mat_00_2, iacc_mat_00_3)), _mm512_add_epi32(_mm512_add_epi32(iacc_mat_00_4, iacc_mat_00_5), _mm512_add_epi32(iacc_mat_00_6, iacc_mat_00_7)));
+                        __m512i iacc_mat_01 = _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(iacc_mat_01_0, iacc_mat_01_1), _mm512_add_epi32(iacc_mat_01_2, iacc_mat_01_3)), _mm512_add_epi32(_mm512_add_epi32(iacc_mat_01_4, iacc_mat_01_5), _mm512_add_epi32(iacc_mat_01_6, iacc_mat_01_7)));
+                        __m512i iacc_mat_10 = _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(iacc_mat_10_0, iacc_mat_10_1), _mm512_add_epi32(iacc_mat_10_2, iacc_mat_10_3)), _mm512_add_epi32(_mm512_add_epi32(iacc_mat_10_4, iacc_mat_10_5), _mm512_add_epi32(iacc_mat_10_6, iacc_mat_10_7)));
+                        __m512i iacc_mat_11 = _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(iacc_mat_11_0, iacc_mat_11_1), _mm512_add_epi32(iacc_mat_11_2, iacc_mat_11_3)), _mm512_add_epi32(_mm512_add_epi32(iacc_mat_11_4, iacc_mat_11_5), _mm512_add_epi32(iacc_mat_11_6, iacc_mat_11_7)));
+
+                        // Straighten out to make 4 row vectors
+                        __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
+                        __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
+                        __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
+                        __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
+
+                        // Load the scale(d) values for all the 4 Q8_k blocks and repeat it across lanes
+                        const __m128 row_scale_f32_sse = _mm_load_ps(a_ptrs[rp][b].d);
+                        const __m256 row_scale_f32_ymm = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
+                        const __m512 row_scale_f32 = _mm512_insertf32x8(_mm512_castps256_ps512(row_scale_f32_ymm), row_scale_f32_ymm, 1);
+
+                        // Multiply with appropiate scales and accumulate (for both d and dmin) below
+                        acc_rows[rp * 4] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
+                        acc_rows[rp * 4  + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
+                        acc_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
+                        acc_rows[rp * 4 + 3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[rp * 4 + 3]);
+
+                        // Take two bsums from two Q8_Ks at a time and multiply with corresponding mins values from each Q2_K
+                        __m512i iacc_row_min_0_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)0), mins_01);
+                        __m512i iacc_row_min_1_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)170), mins_01);
+                        __m512i iacc_row_min_2_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)0), mins_01);
+                        __m512i iacc_row_min_3_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)170), mins_01);
+
+                        __m512i iacc_row_min_0_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)85), mins_23);
+                        __m512i iacc_row_min_1_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)255), mins_23);
+                        __m512i iacc_row_min_2_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)85), mins_23);
+                        __m512i iacc_row_min_3_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)255), mins_23);
+
+                        __m512i iacc_row_min_0_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)0), mins_45);
+                        __m512i iacc_row_min_1_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)170), mins_45);
+                        __m512i iacc_row_min_2_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)0), mins_45);
+                        __m512i iacc_row_min_3_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)170), mins_45);
+
+                        __m512i iacc_row_min_0_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)85), mins_67);
+                        __m512i iacc_row_min_1_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)255), mins_67);
+                        __m512i iacc_row_min_2_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)85), mins_67);
+                        __m512i iacc_row_min_3_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)255), mins_67);
+
+                        __m512i iacc_row_min_0 = _mm512_add_epi32(_mm512_add_epi32(iacc_row_min_0_01, iacc_row_min_0_23), _mm512_add_epi32(iacc_row_min_0_45,iacc_row_min_0_67));
+                        __m512i iacc_row_min_1 = _mm512_add_epi32(_mm512_add_epi32(iacc_row_min_1_01, iacc_row_min_1_23), _mm512_add_epi32(iacc_row_min_1_45,iacc_row_min_1_67));
+                        __m512i iacc_row_min_2 = _mm512_add_epi32(_mm512_add_epi32(iacc_row_min_2_01, iacc_row_min_2_23), _mm512_add_epi32(iacc_row_min_2_45,iacc_row_min_2_67));
+                        __m512i iacc_row_min_3 = _mm512_add_epi32(_mm512_add_epi32(iacc_row_min_3_01, iacc_row_min_3_23), _mm512_add_epi32(iacc_row_min_3_45,iacc_row_min_3_67));
+
+                        acc_min_rows[rp * 4] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_0), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_min_rows[rp * 4]);
+                        acc_min_rows[rp * 4 + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_1), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_min_rows[rp * 4 + 1]);
+                        acc_min_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_2), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_min_rows[rp * 4 + 2]);
+                        acc_min_rows[rp * 4 + 3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_3), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_min_rows[rp * 4 + 3]);
+                    }
+                }
+            }
+            // Store the accumulated values
+            for (int i = 0; i < 16; i++) {
+                _mm512_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm512_sub_ps(acc_rows[i], acc_min_rows[i]));
+            }
+        }
+    }
+
+    for (; y < nr / 4; y ++) {
+
+        const block_q8_Kx4 * a_ptr = a_ptr_start + (y * nb);
+
+        // Take group of eight block_q2_kx8 structures at each pass of the loop and perform dot product operation
+        for (int64_t x = 0; x < anc / 8; x += 2) {
+
+            const block_q2_Kx8 * b_ptr_0 = b_ptr_start + ((x) * b_nb);
+            const block_q2_Kx8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
+
+            // Master FP accumulators
+            __m512 acc_rows[4];
+            for (int i = 0; i < 4; i++) {
+                acc_rows[i] = _mm512_setzero_ps();
+            }
+
+            __m512 acc_min_rows[4];
+            for (int i = 0; i < 4; i++) {
+                acc_min_rows[i] = _mm512_setzero_ps();
+            }
+            // For super block
+            for (int64_t b = 0; b < nb; b++) {
+                // Delta values - Load the sixteen scale values from two block_q2_kx8 structures
+                const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
+
+                // dmin values - Load the sixteen dmin values from two block_q2_kx8 structures
+                const __m512 col_dmin_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].dmin, b_ptr_1[b].dmin);
+
+                // Loop to iterate over the sixteen sub blocks of a super block - eight sub blocks are processed per iteration
+                for (int sb = 0; sb < QK_K / 128; sb++) {
+
+                    // Load the eight block_q2_k for eight sub blocks quantized values interleaved with each other in chunks of eight bytes - B0,B1 ....B6,B7
+                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + sb * 256));
+                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 32 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 64 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 96 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 128 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 160 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 192 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 224 + sb * 256));
+
+                    const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 32 + sb * 256));
+                    const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 64 + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 96 + sb * 256));
+                    const __m256i rhs_raw_mat_89AB_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 128 + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 160 + sb * 256));
+                    const __m256i rhs_raw_mat_89AB_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 192 + sb * 256));
+                    const __m256i rhs_raw_mat_CDEF_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 224 + sb * 256));
+
+                    const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+                    const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+                    const __m256i rhs_raw_mat_0145_2 = _mm256_blend_epi32(rhs_raw_mat_0123_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_2, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_2, requiredOrder), rhs_raw_mat_4567_2, 240);
+                    const __m256i rhs_raw_mat_0145_3 = _mm256_blend_epi32(rhs_raw_mat_0123_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_3, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_3, requiredOrder), rhs_raw_mat_4567_3, 240);
+
+                    const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240);
+                    const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240);
+                    const __m256i rhs_raw_mat_89CD_2 = _mm256_blend_epi32(rhs_raw_mat_89AB_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_2, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_2, requiredOrder), rhs_raw_mat_CDEF_2, 240);
+                    const __m256i rhs_raw_mat_89CD_3 = _mm256_blend_epi32(rhs_raw_mat_89AB_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_3, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_3, requiredOrder), rhs_raw_mat_CDEF_3, 240);
+
+                    const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
+                    const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
+
+                    const __m512i rhs_raw_mat_014589CD_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_2), rhs_raw_mat_89CD_2, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_2), rhs_raw_mat_ABEF_2, 1);
+                    const __m512i rhs_raw_mat_014589CD_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_3), rhs_raw_mat_89CD_3, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_3), rhs_raw_mat_ABEF_3, 1);
+
+                    //2-bit -> 8-bit
+                    const __m512i rhs_mat_014589CD_00 = _mm512_and_si512(rhs_raw_mat_014589CD_0,m3bexpanded); //B00(0-7) B01(0-7) B04(0-7) B05(0-7) B08(0-7) B09(0-7) B0C(0-7) B0D(0-7)
+                    const __m512i rhs_mat_2367ABEF_00 = _mm512_and_si512(rhs_raw_mat_2367ABEF_0,m3bexpanded); //B02(0-7) B03(0-7) B06(0-7) B07(0-7) B0A(0-7) B0B(0-7) B0E(0-7) B0F(0-7)
+                    const __m512i rhs_mat_014589CD_01 = _mm512_and_si512(rhs_raw_mat_014589CD_1,m3bexpanded); //B00(8-15) B01(8-15) B04(8-15) B05(8-15) B08(8-15) B09(8-15) B0C(8-15) B0D(8-15)
+                    const __m512i rhs_mat_2367ABEF_01 = _mm512_and_si512(rhs_raw_mat_2367ABEF_1,m3bexpanded); //B02(8-15) B03(8-15) B06(8-15) B07(8-15) B0A(8-15) B0B(8-15) B0E(8-15) B0F(8-15)
+                    const __m512i rhs_mat_014589CD_10 = _mm512_and_si512(rhs_raw_mat_014589CD_2,m3bexpanded); //B10(0-7) B11(0-7) B14(0-7) B15(0-7) B18(0-7) B19(0-7) B1C(0-7) B1D(0-7)
+                    const __m512i rhs_mat_2367ABEF_10 = _mm512_and_si512(rhs_raw_mat_2367ABEF_2,m3bexpanded); //B12(0-7) B13(0-7) B16(0-7) B17(0-7) B1A(0-7) B1B(0-7) B1E(0-7) B1F(0-7)
+                    const __m512i rhs_mat_014589CD_11 = _mm512_and_si512(rhs_raw_mat_014589CD_3,m3bexpanded); //B10(8-15) B11(8-15) B14(8-15) B15(8-15) B18(8-15) B19(8-15) B1C(8-15) B1D(8-15)
+                    const __m512i rhs_mat_2367ABEF_11 = _mm512_and_si512(rhs_raw_mat_2367ABEF_3,m3bexpanded); //B12(8-15) B13(8-15) B16(8-15) B17(8-15) B1A(8-15) B1B(8-15) B1E(8-15) B1F(8-15)
+
+                    const __m512i rhs_mat_014589CD_20 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 2), m3bexpanded); //B20(0-7) B21(0-7) B24(0-7) B25(0-7) B28(0-7) B29(0-7) B2C(0-7) B2D(0-7)
+                    const __m512i rhs_mat_2367ABEF_20 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 2), m3bexpanded); //B22(0-7) B23(0-7) B26(0-7) B27(0-7) B2A(0-7) B2B(0-7) B2E(0-7) B2F(0-7)
+
+                    const __m512i rhs_mat_014589CD_21 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 2), m3bexpanded); //B20(8-15) B21(8-15) B24(8-15) B25(8-15) B28(8-15) B29(8-15) B2C(8-15) B2D(8-15)
+                    const __m512i rhs_mat_2367ABEF_21 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 2), m3bexpanded); //B22(8-15) B23(8-15) B26(8-15) B27(8-15) B2A(8-15) B2B(8-15) B2E(8-15) B2F(8-15)
+
+                    const __m512i rhs_mat_014589CD_30 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2, 2), m3bexpanded); //B30(0-7) B31(0-7) B34(0-7) B35(0-7) B38(0-7) B39(0-7) B3C(0-7) B3D(0-7)
+                    const __m512i rhs_mat_2367ABEF_30 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2, 2), m3bexpanded); //B32(0-7) B33(0-7) B36(0-7) B37(0-7) B3A(0-7) B3B(0-7) B3E(0-7) B3F(0-7)
+
+                    const __m512i rhs_mat_014589CD_31 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3, 2), m3bexpanded); //B30(8-15) B31(8-15) B34(8-15) B35(8-15) B38(8-15) B39(8-15) B3C(8-15) B3D(8-15)
+                    const __m512i rhs_mat_2367ABEF_31 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3, 2), m3bexpanded); //B32(8-15) B33(8-15) B36(8-15) B37(8-15) B3A(8-15) B3B(8-15) B3E(8-15) B3F(8-15)
+
+                    const __m512i rhs_mat_014589CD_40 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m3bexpanded); //B40(0-7) B41(0-7) B44(0-7) B45(0-7) B48(0-7) B49(0-7) B4C(0-7) B4D(0-7)
+                    const __m512i rhs_mat_2367ABEF_40 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m3bexpanded); //B42(0-7) B43(0-7) B46(0-7) B47(0-7) B4A(0-7) B4B(0-7) B4E(0-7) B4F(0-7)
+
+                    const __m512i rhs_mat_014589CD_41 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m3bexpanded); //B40(8-15) B41(8-15) B44(8-15) B45(8-15) B48(8-15) B49(8-15) B4C(8-15) B4D(8-15)
+                    const __m512i rhs_mat_2367ABEF_41 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m3bexpanded); //B42(8-15) B43(8-15) B46(8-15) B47(8-15) B4A(8-15) B4B(8-15) B4E(8-15) B4F(8-15)
+
+                    const __m512i rhs_mat_014589CD_50 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2, 4), m3bexpanded); //B50(0-7) B51(0-7) B54(0-7) B55(0-7) B58(0-7) B59(0-7) B5C(0-7) B5D(0-7)
+                    const __m512i rhs_mat_2367ABEF_50 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2, 4), m3bexpanded); //B52(0-7) B53(0-7) B56(0-7) B57(0-7) B5A(0-7) B5B(0-7) B5E(0-7) B5F(0-7)
+
+                    const __m512i rhs_mat_014589CD_51 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3, 4), m3bexpanded); //B50(8-15) B51(8-15) B54(8-15) B55(8-15) B58(8-15) B59(8-15) B5C(8-15) B5D(8-15)
+                    const __m512i rhs_mat_2367ABEF_51 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3, 4), m3bexpanded); //B52(8-15) B53(8-15) B56(8-15) B57(8-15) B5A(8-15) B5B(8-15) B5E(8-15) B5F(8-15)
+
+                    const __m512i rhs_mat_014589CD_60 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 6), m3bexpanded); //B60(0-7) B61(0-7) B64(0-7) B65(0-7) B68(0-7) B69(0-7) B6C(0-7) B6D(0-7)
+                    const __m512i rhs_mat_2367ABEF_60 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 6), m3bexpanded); //B62(0-7) B63(0-7) B66(0-7) B67(0-7) B6A(0-7) B6B(0-7) B6E(0-7) B6F(0-7)
+
+                    const __m512i rhs_mat_014589CD_61 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 6), m3bexpanded); //B60(8-15) B61(8-15) B64(8-15) B65(8-15) B68(8-15) B69(8-15) B6C(8-15) B6D(8-15)
+                    const __m512i rhs_mat_2367ABEF_61 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 6), m3bexpanded); //B62(8-15) B63(8-15) B66(8-15) B67(8-15) B6A(8-15) B6B(8-15) B6E(8-15) B6F(8-15)
+
+                    const __m512i rhs_mat_014589CD_70 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2, 6), m3bexpanded); //B70(0-7) B71(0-7) B74(0-7) B75(0-7) B78(0-7) B79(0-7) B7C(0-7) B7D(0-7)
+                    const __m512i rhs_mat_2367ABEF_70 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2, 6), m3bexpanded); //B72(0-7) B73(0-7) B76(0-7) B77(0-7) B7A(0-7) B7B(0-7) B7E(0-7) B7F(0-7)
+
+                    const __m512i rhs_mat_014589CD_71 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3, 6), m3bexpanded); //B70(8-15) B71(8-15) B74(8-15) B75(8-15) B78(8-15) B79(8-15) B7C(8-15) B7D(8-15)
+                    const __m512i rhs_mat_2367ABEF_71 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3, 6), m3bexpanded); //B72(8-15) B73(8-15) B76(8-15) B77(8-15) B7A(8-15) B7B(8-15) B7E(8-15) B7F(8-15)
+
+                    const __m512i rhs_mat_014589CD_00_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_00, (_MM_PERM_ENUM)136); //B00(0-3) B01(0-3) B00(0-3) B01(0-3) B04(0-3) B05(0-3) B04(0-3) B05(0-3) B08(0-3) B09(0-3) B08(0-3) B09(0-3) B0C(0-3) B0D(0-3) B0C(0-3) B0D(0-3)
+                    const __m512i rhs_mat_2367ABEF_00_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_00, (_MM_PERM_ENUM)136); //B02(0-3) B03(0-3) B02(0-3) B03(0-3) B06(0-3) B07(0-3) B06(0-3) B07(0-3) B0A(0-3) B0B(0-3) B0A(0-3) B0B(0-3) B0E(0-3) B0F(0-3) B0E(0-3) B0F(0-3)
+
+                    const __m512i rhs_mat_014589CD_01_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_01, (_MM_PERM_ENUM)136); //B00(8-11) B01(8-11) B00(8-11) B01(8-11) B04(8-11) B05(8-11) B04(8-11) B05(8-11) B08(8-11) B09(8-11) B08(8-11) B09(8-11) B0C(8-11) B0D(8-11) B0C(8-11) B0D(8-11)
+                    const __m512i rhs_mat_2367ABEF_01_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_01, (_MM_PERM_ENUM)136); //B02(8-11) B03(8-11) B02(8-11) B03(8-11) B06(8-11) B07(8-11) B06(8-11) B07(8-11) B0A(8-11) B0B(8-11) B0A(8-11) B0B(8-11) B0E(8-11) B0F(8-11) B0E(8-11) B0F(8-11)
+
+                    const __m512i rhs_mat_014589CD_10_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_10, (_MM_PERM_ENUM)136); //B10(0-3) B11(0-3) B10(0-3) B11(0-3) B14(0-3) B15(0-3) B14(0-3) B15(0-3) B18(0-3) B19(0-3) B18(0-3) B19(0-3) B1C(0-3) B1D(0-3) B1C(0-3) B1D(0-3)
+                    const __m512i rhs_mat_2367ABEF_10_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_10, (_MM_PERM_ENUM)136); //B12(0-3) B13(0-3) B12(0-3) B13(0-3) B16(0-3) B17(0-3) B16(0-3) B17(0-3) B1A(0-3) B1B(0-3) B1A(0-3) B1B(0-3) B1E(0-3) B1F(0-3) B1E(0-3) B1F(0-3)
+
+                    const __m512i rhs_mat_014589CD_11_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_11, (_MM_PERM_ENUM)136); //B10(8-11) B11(8-11) B10(8-11) B11(8-11) B14(8-11) B15(8-11) B14(8-11) B15(8-11) B18(8-11) B19(8-11) B18(8-11) B19(8-11) B1C(8-11) B1D(8-11) B1C(8-11) B1D(8-11)
+                    const __m512i rhs_mat_2367ABEF_11_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_11, (_MM_PERM_ENUM)136); //B12(8-11) B13(8-11) B12(8-11) B13(8-11) B16(8-11) B17(8-11) B16(8-11) B17(8-11) B1A(8-11) B1B(8-11) B1A(8-11) B1B(8-11) B1E(8-11) B1F(8-11) B1E(8-11) B1F(8-11)
+
+                    const __m512i rhs_mat_014589CD_20_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_20, (_MM_PERM_ENUM)136); //B20(0-3) B21(0-3) B20(0-3) B21(0-3) B24(0-3) B25(0-3) B24(0-3) B25(0-3) B28(0-3) B29(0-3) B28(0-3) B29(0-3) B2C(0-3) B2D(0-3) B2C(0-3) B2D(0-3)
+                    const __m512i rhs_mat_2367ABEF_20_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_20, (_MM_PERM_ENUM)136); //B22(0-3) B23(0-3) B22(0-3) B23(0-3) B26(0-3) B27(0-3) B26(0-3) B27(0-3) B2A(0-3) B2B(0-3) B2A(0-3) B2B(0-3) B2E(0-3) B2F(0-3) B2E(0-3) B2F(0-3)
+
+                    const __m512i rhs_mat_014589CD_21_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_21, (_MM_PERM_ENUM)136); //B20(8-11) B21(8-11) B20(8-11) B21(8-11) B24(8-11) B25(8-11) B24(8-11) B25(8-11) B28(8-11) B29(8-11) B28(8-11) B29(8-11) B2C(8-11) B2D(8-11) B2C(8-11) B2D(8-11)
+                    const __m512i rhs_mat_2367ABEF_21_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_21, (_MM_PERM_ENUM)136); //B22(8-11) B23(8-11) B22(8-11) B23(8-11) B26(8-11) B27(8-11) B26(8-11) B27(8-11) B2A(8-11) B2B(8-11) B2A(8-11) B2B(8-11) B2E(8-11) B2F(8-11) B2E(8-11) B2F(8-11)
+                    const __m512i rhs_mat_014589CD_30_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_30, (_MM_PERM_ENUM)136); ///B30(0-3) B31(0-3) B30(0-3) B31(0-3) B34(0-3) B35(0-3) B34(0-3) B35(0-3) B38(0-3) B39(0-3) B38(0-3) B39(0-3) B3C(0-3) B3D(0-3) B3C(0-3) B3D(0-3)
+                    const __m512i rhs_mat_2367ABEF_30_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_30, (_MM_PERM_ENUM)136); //B32(0-3) B33(0-3) B32(0-3) B33(0-3) B36(0-3) B37(0-3) B36(0-3) B37(0-3) B3A(0-3) B3B(0-3) B3A(0-3) B3B(0-3) B3E(0-3) B3F(0-3) B3E(0-3) B3F(0-3)
+
+                    const __m512i rhs_mat_014589CD_31_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_31, (_MM_PERM_ENUM)136); //B30(8-11) B31(8-11) B30(8-11) B31(8-11) B34(8-11) B35(8-11) B34(8-11) B35(8-11) B38(8-11) B39(8-11) B38(8-11) B39(8-11) B3C(8-11) B3D(8-11) B3C(8-11) B3D(8-11)
+                    const __m512i rhs_mat_2367ABEF_31_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_31, (_MM_PERM_ENUM)136); //B32(8-11) B33(8-11) B32(8-11) B33(8-11) B36(8-11) B37(8-11) B36(8-11) B37(8-11) B3A(8-11) B3B(8-11) B3A(8-11) B3B(8-11) B3E(8-11) B3F(8-11) B3E(8-11) B3F(8-11)
+
+                    const __m512i rhs_mat_014589CD_40_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_40, (_MM_PERM_ENUM)136); //B40(0-3) B41(0-3) B40(0-3) B41(0-3) B44(0-3) B45(0-3) B44(0-3) B45(0-3) B48(0-3) B49(0-3) B48(0-3) B49(0-3) B4C(0-3) B4D(0-3) B4C(0-3) B4D(0-3)
+                    const __m512i rhs_mat_2367ABEF_40_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_40, (_MM_PERM_ENUM)136); //B42(0-3) B43(0-3) B42(0-3) B43(0-3) B46(0-3) B47(0-3) B46(0-3) B47(0-3) B4A(0-3) B4B(0-3) B4A(0-3) B4B(0-3) B4E(0-3) B4F(0-3) B4E(0-3) B4F(0-3)
+
+                    const __m512i rhs_mat_014589CD_41_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_41, (_MM_PERM_ENUM)136); //B40(8-11) B41(8-11) B40(8-11) B41(8-11) B44(8-11) B45(8-11) B44(8-11) B45(8-11) B48(8-11) B49(8-11) B48(8-11) B49(8-11) B4C(8-11) B4D(8-11) B4C(8-11) B4D(8-11)
+                    const __m512i rhs_mat_2367ABEF_41_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_41, (_MM_PERM_ENUM)136); //B42(8-11) B43(8-11) B42(8-11) B43(8-11) B46(8-11) B47(8-11) B46(8-11) B47(8-11) B4A(8-11) B4B(8-11) B4A(8-11) B4B(8-11) B4E(8-11) B4F(8-11) B4E(8-11) B4F(8-11)
+
+                    const __m512i rhs_mat_014589CD_50_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_50, (_MM_PERM_ENUM)136); //B50(0-3) B51(0-3) B50(0-3) B51(0-3) B54(0-3) B55(0-3) B54(0-3) B55(0-3) B58(0-3) B59(0-3) B58(0-3) B59(0-3) B5C(0-3) B5D(0-3) B5C(0-3) B5D(0-3)
+                    const __m512i rhs_mat_2367ABEF_50_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_50, (_MM_PERM_ENUM)136); //B52(0-3) B53(0-3) B52(0-3) B53(0-3) B56(0-3) B57(0-3) B56(0-3) B57(0-3) B5A(0-3) B5B(0-3) B5A(0-3) B5B(0-3) B5E(0-3) B5F(0-3) B5E(0-3) B5F(0-3)
+
+                    const __m512i rhs_mat_014589CD_51_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_51, (_MM_PERM_ENUM)136); //B50(8-11) B51(8-11) B50(8-11) B51(8-11) B54(8-11) B55(8-11) B54(8-11) B55(8-11) B58(8-11) B59(8-11) B58(8-11) B59(8-11) B5C(8-11) B5D(8-11) B5C(8-11) B5D(8-11)
+                    const __m512i rhs_mat_2367ABEF_51_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_51, (_MM_PERM_ENUM)136); //B52(8-11) B53(8-11) B52(8-11) B53(8-11) B56(8-11) B57(8-11) B56(8-11) B57(8-11) B5A(8-11) B5B(8-11) B5A(8-11) B5B(8-11) B5E(8-11) B5F(8-11) B5E(8-11) B5F(8-11)
+
+                    const __m512i rhs_mat_014589CD_60_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_60, (_MM_PERM_ENUM)136); //B60(0-3) B61(0-3) B60(0-3) B61(0-3) B64(0-3) B65(0-3) B64(0-3) B65(0-3) B68(0-3) B69(0-3) B68(0-3) B69(0-3) B6C(0-3) B6D(0-3) B6C(0-3) B6D(0-3)
+                    const __m512i rhs_mat_2367ABEF_60_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_60, (_MM_PERM_ENUM)136); //B62(0-3) B63(0-3) B62(0-3) B63(0-3) B66(0-3) B67(0-3) B66(0-3) B67(0-3) B6A(0-3) B6B(0-3) B6A(0-3) B6B(0-3) B6E(0-3) B6F(0-3) B6E(0-3) B6F(0-3)
+
+                    const __m512i rhs_mat_014589CD_61_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_61, (_MM_PERM_ENUM)136); //B60(8-11) B61(8-11) B60(8-11) B61(8-11) B64(8-11) B65(8-11) B64(8-11) B65(8-11) B68(8-11) B69(8-11) B68(8-11) B69(8-11) B6C(8-11) B6D(8-11) B6C(8-11) B6D(8-11)
+                    const __m512i rhs_mat_2367ABEF_61_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_61, (_MM_PERM_ENUM)136); //B62(8-11) B63(8-11) B62(8-11) B63(8-11) B66(8-11) B67(8-11) B66(8-11) B67(8-11) B6A(8-11) B6B(8-11) B6A(8-11) B6B(8-11) B6E(8-11) B6F(8-11) B6E(8-11) B6F(8-11)
+
+                    const __m512i rhs_mat_014589CD_70_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_70, (_MM_PERM_ENUM)136); //B70(0-3) B71(0-3) B70(0-3) B71(0-3) B74(0-3) B75(0-3) B74(0-3) B75(0-3) B78(0-3) B79(0-3) B78(0-3) B79(0-3) B7C(0-3) B7D(0-3) B7C(0-3) B7D(0-3)
+                    const __m512i rhs_mat_2367ABEF_70_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_70, (_MM_PERM_ENUM)136); //B72(0-3) B73(0-3) B72(0-3) B73(0-3) B76(0-3) B77(0-3) B76(0-3) B77(0-3) B7A(0-3) B7B(0-3) B7A(0-3) B7B(0-3) B7E(0-3) B7F(0-3) B7E(0-3) B7F(0-3)
+
+                    const __m512i rhs_mat_014589CD_71_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_71, (_MM_PERM_ENUM)136); //B00(8-11) B01(8-11) B00(8-11) B01(8-11) B04(8-11) B05(8-11) B04(8-11) B05(8-11) B08(8-11) B09(8-11) B08(8-11) B09(8-11) B0C(8-11) B0D(8-11) B0C(8-11) B0D(8-11)
+                    const __m512i rhs_mat_2367ABEF_71_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_71, (_MM_PERM_ENUM)136); //B72(8-11) B73(8-11) B72(8-11) B73(8-11) B76(8-11) B77(8-11) B76(8-11) B77(8-11) B7A(8-11) B7B(8-11) B7A(8-11) B7B(8-11) B7E(8-11) B7F(8-11) B7E(8-11) B7F(8-11)
+
+                    const __m512i rhs_mat_014589CD_00_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_00, (_MM_PERM_ENUM)221); //B00(4-7) B01(4-7) B00(4-7) B01(4-7) B04(4-7) B05(4-7) B04(4-7) B05(4-7) B08(4-7) B09(4-7) B08(4-7) B09(4-7) B0C(4-7) B0D(4-7) B0C(4-7) B0D(4-7)
+                    const __m512i rhs_mat_2367ABEF_00_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_00, (_MM_PERM_ENUM)221); //B02(4-7) B03(4-7) B02(4-7) B03(4-7) B06(4-7) B07(4-7) B06(4-7) B07(4-7) B0A(4-7) B0B(4-7) B0A(4-7) B0B(4-7) B0E(4-7) B0F(4-7) B0E(4-7) B0F(4-7)
+
+                    const __m512i rhs_mat_014589CD_01_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_01, (_MM_PERM_ENUM)221); //B00(12-15) B01(12-15) B00(12-15) B01(12-15) B04(12-15) B05(12-15) B04(12-15) B05(12-15) B08(12-15) B09(12-15) B08(12-15) B09(12-15) B0C(12-15) B0D(12-15) B0C(12-15) B0D(12-15)
+                    const __m512i rhs_mat_2367ABEF_01_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_01, (_MM_PERM_ENUM)221); //B02(12-15) B03(12-15) B02(12-15) B03(12-15) B06(12-15) B07(12-15) B06(12-15) B07(12-15) B0A(12-15) B0B(12-15) B0A(12-15) B0B(12-15) B0E(12-15) B0F(12-15) B0E(12-15) B0F(12-15)
+
+                    const __m512i rhs_mat_014589CD_10_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_10, (_MM_PERM_ENUM)221); //B10(4-7) B11(4-7) B10(4-7) B11(4-7) B14(4-7) B15(4-7) B14(4-7) B15(4-7) B18(4-7) B19(4-7) B18(4-7) B19(4-7) B1C(4-7) B1D(4-7) B1C(4-7) B1D(4-7)
+                    const __m512i rhs_mat_2367ABEF_10_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_10, (_MM_PERM_ENUM)221); //B12(4-7) B13(4-7) B12(4-7) B13(4-7) B16(4-7) B17(4-7) B16(4-7) B17(4-7) B1A(4-7) B1B(4-7) B1A(4-7) B1B(4-7) B1E(4-7) B1F(4-7) B1E(4-7) B1F(4-7)
+
+                    const __m512i rhs_mat_014589CD_11_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_11, (_MM_PERM_ENUM)221); //B10(12-15) B11(12-15) B10(12-15) B11(12-15) B14(12-15) B15(12-15) B14(12-15) B15(12-15) B18(12-15) B19(12-15) B18(12-15) B19(12-15) B1C(12-15) B1D(12-15) B1C(12-15) B1D(12-15)
+                    const __m512i rhs_mat_2367ABEF_11_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_11, (_MM_PERM_ENUM)221); //B12(12-15) B13(12-15) B12(12-15) B13(12-15) B16(12-15) B17(12-15) B16(12-15) B17(12-15) B1A(12-15) B1B(12-15) B1A(12-15) B1B(12-15) B1E(12-15) B1F(12-15) B1E(12-15) B1F(12-15)
+
+                    const __m512i rhs_mat_014589CD_20_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_20, (_MM_PERM_ENUM)221); //B20(4-7) B21(4-7) B20(4-7) B21(4-7) B24(4-7) B25(4-7) B24(4-7) B25(4-7) B28(4-7) B29(4-7) B28(4-7) B29(4-7) B2C(4-7) B2D(4-7) B2C(4-7) B2D(4-7)
+                    const __m512i rhs_mat_2367ABEF_20_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_20, (_MM_PERM_ENUM)221); //B22(4-7) B23(4-7) B22(4-7) B23(4-7) B26(4-7) B27(4-7) B26(4-7) B27(4-7) B2A(4-7) B2B(4-7) B2A(4-7) B2B(4-7) B2E(4-7) B2F(4-7) B2E(4-7) B2F(4-7)
+
+                    const __m512i rhs_mat_014589CD_21_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_21, (_MM_PERM_ENUM)221); //B20(12-15) B21(12-15) B20(12-15) B21(12-15) B24(12-15) B25(12-15) B24(12-15) B25(12-15) B28(12-15) B29(12-15) B28(12-15) B29(12-15) B2C(12-15) B2D(12-15) B2C(12-15) B2D(12-15)
+                    const __m512i rhs_mat_2367ABEF_21_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_21, (_MM_PERM_ENUM)221); //B22(12-15) B23(12-15) B22(12-15) B23(12-15) B26(12-15) B27(12-15) B26(12-15) B27(12-15) B2A(12-15) B2B(12-15) B2A(12-15) B2B(12-15) B2E(12-15) B2F(12-15) B2E(12-15) B2F(12-15)
+
+                    const __m512i rhs_mat_014589CD_30_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_30, (_MM_PERM_ENUM)221); //B30(4-7) B31(4-7) B30(4-7) B31(4-7) B34(4-7) B35(4-7) B34(4-7) B35(4-7) B38(4-7) B39(4-7) B38(4-7) B39(4-7) B3C(4-7) B3D(4-7) B3C(4-7) B3D(4-7)
+                    const __m512i rhs_mat_2367ABEF_30_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_30, (_MM_PERM_ENUM)221); //B32(4-7) B33(4-7) B32(4-7) B33(4-7) B36(4-7) B37(4-7) B36(4-7) B37(4-7) B3A(4-7) B3B(4-7) B3A(4-7) B3B(4-7) B3E(4-7) B3F(4-7) B3E(4-7) B3F(4-7)
+
+                    const __m512i rhs_mat_014589CD_31_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_31, (_MM_PERM_ENUM)221); //B30(12-15) B31(12-15) B30(12-15) B31(12-15) B34(12-15) B35(12-15) B34(12-15) B35(12-15) B38(12-15) B39(12-15) B38(12-15) B39(12-15) B3C(12-15) B3D(12-15) B3C(12-15) B3D(12-15)
+                    const __m512i rhs_mat_2367ABEF_31_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_31, (_MM_PERM_ENUM)221); //B32(12-15) B33(12-15) B32(12-15) B33(12-15) B36(12-15) B37(12-15) B36(12-15) B37(12-15) B3A(12-15) B3B(12-15) B3A(12-15) B3B(12-15) B3E(12-15) B3F(12-15) B3E(12-15) B3F(12-15)
+
+                    const __m512i rhs_mat_014589CD_40_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_40, (_MM_PERM_ENUM)221); //B40(4-7) B41(4-7) B40(4-7) B41(4-7) B44(4-7) B45(4-7) B44(4-7) B45(4-7) B48(4-7) B49(4-7) B48(4-7) B49(4-7) B4C(4-7) B4D(4-7) B4C(4-7) B4D(4-7)
+                    const __m512i rhs_mat_2367ABEF_40_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_40, (_MM_PERM_ENUM)221); //B42(4-7) B43(4-7) B42(4-7) B43(4-7) B46(4-7) B47(4-7) B46(4-7) B47(4-7) B4A(4-7) B4B(4-7) B4A(4-7) B4B(4-7) B4E(4-7) B4F(4-7) B4E(4-7) B4F(4-7)
+
+                    const __m512i rhs_mat_014589CD_41_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_41, (_MM_PERM_ENUM)221); //B40(12-15) B41(12-15) B40(12-15) B41(12-15) B44(12-15) B45(12-15) B44(12-15) B45(12-15) B48(12-15) B49(12-15) B48(12-15) B49(12-15) B4C(12-15) B4D(12-15) B4C(12-15) B4D(12-15)
+                    const __m512i rhs_mat_2367ABEF_41_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_41, (_MM_PERM_ENUM)221); //B42(12-15) B43(12-15) B42(12-15) B43(12-15) B46(12-15) B47(12-15) B46(12-15) B47(12-15) B4A(12-15) B4B(12-15) B4A(12-15) B4B(12-15) B4E(12-15) B4F(12-15) B4E(12-15) B4F(12-15)
+
+                    const __m512i rhs_mat_014589CD_50_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_50, (_MM_PERM_ENUM)221); //B50(4-7) B51(4-7) B50(4-7) B51(4-7) B54(4-7) B55(4-7) B54(4-7) B55(4-7) B58(4-7) B59(4-7) B58(4-7) B59(4-7) B5C(4-7) B5D(4-7) B5C(4-7) B5D(4-7)
+                    const __m512i rhs_mat_2367ABEF_50_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_50, (_MM_PERM_ENUM)221); //B52(4-7) B53(4-7) B52(4-7) B53(4-7) B56(4-7) B57(4-7) B56(4-7) B57(4-7) B5A(4-7) B5B(4-7) B5A(4-7) B5B(4-7) B5E(4-7) B5F(4-7) B5E(4-7) B5F(4-7)
+
+                    const __m512i rhs_mat_014589CD_51_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_51, (_MM_PERM_ENUM)221); //B50(12-15) B51(12-15) B50(12-15) B51(12-15) B54(12-15) B55(12-15) B54(12-15) B55(12-15) B58(12-15) B59(12-15) B58(12-15) B59(12-15) B5C(12-15) B5D(12-15) B5C(12-15) B5D(12-15)
+                    const __m512i rhs_mat_2367ABEF_51_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_51, (_MM_PERM_ENUM)221); //B52(12-15) B53(12-15) B52(12-15) B53(12-15) B56(12-15) B57(12-15) B56(12-15) B57(12-15) B5A(12-15) B5B(12-15) B5A(12-15) B5B(12-15) B5E(12-15) B5F(12-15) B5E(12-15) B5F(12-15)
+
+                    const __m512i rhs_mat_014589CD_60_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_60, (_MM_PERM_ENUM)221); //B60(4-7) B61(4-7) B60(4-7) B61(4-7) B64(4-7) B65(4-7) B64(4-7) B65(4-7) B68(4-7) B69(4-7) B68(4-7) B69(4-7) B6C(4-7) B6D(4-7) B6C(4-7) B6D(4-7)
+                    const __m512i rhs_mat_2367ABEF_60_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_60, (_MM_PERM_ENUM)221); //B62(4-7) B63(4-7) B62(4-7) B63(4-7) B66(4-7) B67(4-7) B66(4-7) B67(4-7) B6A(4-7) B6B(4-7) B6A(4-7) B6B(4-7) B6E(4-7) B6F(4-7) B6E(4-7) B6F(4-7)
+
+                    const __m512i rhs_mat_014589CD_61_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_61, (_MM_PERM_ENUM)221); //B60(12-15) B61(12-15) B60(12-15) B61(12-15) B64(12-15) B65(12-15) B64(12-15) B65(12-15) B68(12-15) B69(12-15) B68(12-15) B69(12-15) B6C(12-15) B6D(12-15) B6C(12-15) B6D(12-15)
+                    const __m512i rhs_mat_2367ABEF_61_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_61, (_MM_PERM_ENUM)221); //B62(12-15) B63(12-15) B62(12-15) B63(12-15) B66(12-15) B67(12-15) B66(12-15) B67(12-15) B6A(12-15) B6B(12-15) B6A(12-15) B6B(12-15) B6E(12-15) B6F(12-15) B6E(12-15) B6F(12-15)
+
+                    const __m512i rhs_mat_014589CD_70_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_70, (_MM_PERM_ENUM)221); //B70(4-7) B71(4-7) B70(4-7) B71(4-7) B74(4-7) B75(4-7) B74(4-7) B75(4-7) B78(4-7) B79(4-7) B78(4-7) B79(4-7) B7C(4-7) B7D(4-7) B7C(4-7) B7D(4-7)
+                    const __m512i rhs_mat_2367ABEF_70_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_70, (_MM_PERM_ENUM)221); //B72(4-7) B73(4-7) B72(4-7) B73(4-7) B76(4-7) B77(4-7) B76(4-7) B77(4-7) B7A(4-7) B7B(4-7) B7A(4-7) B7B(4-7) B7E(4-7) B7F(4-7) B7E(4-7) B7F(4-7)
+
+                    const __m512i rhs_mat_014589CD_71_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_71, (_MM_PERM_ENUM)221); //B70(12-15) B71(12-15) B70(12-15) B71(12-15) B74(12-15) B75(12-15) B74(12-15) B75(12-15) B78(12-15) B79(12-15) B78(12-15) B79(12-15) B7C(12-15) B7D(12-15) B7C(12-15) B7D(12-15)
+                    const __m512i rhs_mat_2367ABEF_71_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_71, (_MM_PERM_ENUM)221); //B72(12-15) B73(12-15) B72(12-15) B73(12-15) B76(12-15) B77(12-15) B76(12-15) B77(12-15) B7A(12-15) B7B(12-15) B7A(12-15) B7B(12-15) B7E(12-15) B7F(12-15) B7E(12-15) B7F(12-15)
+
+                    //notation:superblock subblock
+                    //s00 m00  s01 m01   s10 m10  s11 m11  s20 m20  s21 m21   s30 m30  s31 m31  s40 m40  s41 m41   s50 m50  s51 m51  s60 m60  s61 m61   s70 m70  s71 m71
+
+                    const __m128i mins_and_scales_01_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + sb * 64));
+                    const __m128i mins_and_scales_23_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + 16 + sb * 64));
+                    const __m128i mins_and_scales_45_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + 32 + sb * 64));
+                    const __m128i mins_and_scales_67_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + 48 + sb * 64));
+
+                    const __m128i mins_and_scales_01_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + sb * 64));
+                    const __m128i mins_and_scales_23_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + 16 + sb * 64));
+                    const __m128i mins_and_scales_45_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + 32 + sb * 64));
+                    const __m128i mins_and_scales_67_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + 48 + sb * 64));
+
+                    // Combine mins and scales for sub-blocks: 0-1, 2-3, 4-5, 6-7 in the sb loop
+                    const __m256i mins_and_scales_01 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_01_0), mins_and_scales_01_1, 1);
+                    const __m256i mins_and_scales_23 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_23_0), mins_and_scales_23_1, 1);
+                    const __m256i mins_and_scales_45 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_45_0), mins_and_scales_45_1, 1);
+                    const __m256i mins_and_scales_67 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_67_0), mins_and_scales_67_1, 1);
+
+                    // Extract scales which is lower half from mins_and_scales
+                    const __m256i scales_01 = _mm256_and_si256(mins_and_scales_01, m4b);
+                    const __m256i scales_23 = _mm256_and_si256(mins_and_scales_23, m4b);
+                    const __m256i scales_45 = _mm256_and_si256(mins_and_scales_45, m4b);
+                    const __m256i scales_67 = _mm256_and_si256(mins_and_scales_67, m4b);
+
+                    // Extract mins which is upper half from mins_and_scales
+                    const __m512i mins_01 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_01, 4), m4b));
+                    const __m512i mins_23 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_23, 4), m4b));
+                    const __m512i mins_45 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_45, 4), m4b));
+                    const __m512i mins_67 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_67, 4), m4b));
+
+                    const __m512i scales_0 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_01, scalesmask1));
+                    const __m512i scales_1 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_01, scalesmask2));
+                    const __m512i scales_2 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_23, scalesmask1));
+                    const __m512i scales_3 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_23, scalesmask2));
+                    const __m512i scales_4 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_45, scalesmask1));
+                    const __m512i scales_5 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_45, scalesmask2));
+                    const __m512i scales_6 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_67, scalesmask1));
+                    const __m512i scales_7 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_67, scalesmask2));
+
+                    const __m512i scale_014589CD_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)238);
+
+                    const __m512i scale_014589CD_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)238);
+
+                    const __m512i scale_014589CD_2 = _mm512_shuffle_epi32(scales_2, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_2 = _mm512_shuffle_epi32(scales_2, (_MM_PERM_ENUM)238);
+
+                    const __m512i scale_014589CD_3 = _mm512_shuffle_epi32(scales_3, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_3 = _mm512_shuffle_epi32(scales_3, (_MM_PERM_ENUM)238);
+
+                    const __m512i scale_014589CD_4 = _mm512_shuffle_epi32(scales_4, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_4 = _mm512_shuffle_epi32(scales_4, (_MM_PERM_ENUM)238);
+
+                    const __m512i scale_014589CD_5 = _mm512_shuffle_epi32(scales_5, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_5 = _mm512_shuffle_epi32(scales_5, (_MM_PERM_ENUM)238);
+
+                    const __m512i scale_014589CD_6 = _mm512_shuffle_epi32(scales_6, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_6 = _mm512_shuffle_epi32(scales_6, (_MM_PERM_ENUM)238);
+
+                    const __m512i scale_014589CD_7 = _mm512_shuffle_epi32(scales_7, (_MM_PERM_ENUM)68);
+                    const __m512i scale_2367ABEF_7 = _mm512_shuffle_epi32(scales_7, (_MM_PERM_ENUM)238);
+
+                    // Load the four block_q8_k quantized values interleaved with each other in chunks of eight bytes - A0,A1,A2,A3
+                    // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
+                    __m256i lhs_mat_ymm_0123_00 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 512 * sb)));
+                    __m256i lhs_mat_ymm_01_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 0);
+                    __m256i lhs_mat_ymm_23_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 17);
+                    __m256i lhs_mat_ymm_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 32 + 512 * sb)));
+                    __m256i lhs_mat_ymm_01_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 0);
+                    __m256i lhs_mat_ymm_23_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 17);
+                    __m256i lhs_mat_ymm_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 64 + 512 * sb)));
+                    __m256i lhs_mat_ymm_01_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 0);
+                    __m256i lhs_mat_ymm_23_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 17);
+                    __m256i lhs_mat_ymm_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 96 + 512 * sb)));
+                    __m256i lhs_mat_ymm_01_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 0);
+                    __m256i lhs_mat_ymm_23_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 17);
+                    __m256i lhs_mat_ymm_0123_20 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 128 + 512 * sb)));
+                    __m256i lhs_mat_ymm_01_20 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_20, lhs_mat_ymm_0123_20, 0);
+                    __m256i lhs_mat_ymm_23_20 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_20, lhs_mat_ymm_0123_20, 17);
+                    __m256i lhs_mat_ymm_0123_21 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 160 + 512 * sb)));
+                    __m256i lhs_mat_ymm_01_21 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_21, lhs_mat_ymm_0123_21, 0);
+                    __m256i lhs_mat_ymm_23_21 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_21, lhs_mat_ymm_0123_21, 17);
+                    __m256i lhs_mat_ymm_0123_30 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 192 + 512 * sb)));
+                    __m256i lhs_mat_ymm_01_30 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_30, lhs_mat_ymm_0123_30, 0);
+                    __m256i lhs_mat_ymm_23_30 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_30, lhs_mat_ymm_0123_30, 17);
+                    __m256i lhs_mat_ymm_0123_31 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 224 + 512 * sb)));
+                    __m256i lhs_mat_ymm_01_31 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_31, lhs_mat_ymm_0123_31, 0);
+                    __m256i lhs_mat_ymm_23_31 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_31, lhs_mat_ymm_0123_31, 17);
+
+                    __m256i lhs_mat_ymm_0123_40 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 256 + 512 * sb)));
+                    __m256i lhs_mat_ymm_01_40 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_40, lhs_mat_ymm_0123_40, 0);
+                    __m256i lhs_mat_ymm_23_40 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_40, lhs_mat_ymm_0123_40, 17);
+                    __m256i lhs_mat_ymm_0123_41 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 288 + 512 * sb)));
+                    __m256i lhs_mat_ymm_01_41 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_41, lhs_mat_ymm_0123_41, 0);
+                    __m256i lhs_mat_ymm_23_41 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_41, lhs_mat_ymm_0123_41, 17);
+                    __m256i lhs_mat_ymm_0123_50 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 320 + 512 * sb)));
+                    __m256i lhs_mat_ymm_01_50 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_50, lhs_mat_ymm_0123_50, 0);
+                    __m256i lhs_mat_ymm_23_50 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_50, lhs_mat_ymm_0123_50, 17);
+                    __m256i lhs_mat_ymm_0123_51 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 352 + 512 * sb)));
+                    __m256i lhs_mat_ymm_01_51 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_51, lhs_mat_ymm_0123_51, 0);
+                    __m256i lhs_mat_ymm_23_51 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_51, lhs_mat_ymm_0123_51, 17);
+                    __m256i lhs_mat_ymm_0123_60 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 384 + 512 * sb)));
+                    __m256i lhs_mat_ymm_01_60 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_60, lhs_mat_ymm_0123_60, 0);
+                    __m256i lhs_mat_ymm_23_60 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_60, lhs_mat_ymm_0123_60, 17);
+                    __m256i lhs_mat_ymm_0123_61 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 416 + 512 * sb)));
+                    __m256i lhs_mat_ymm_01_61 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_61, lhs_mat_ymm_0123_61, 0);
+                    __m256i lhs_mat_ymm_23_61 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_61, lhs_mat_ymm_0123_61, 17);
+                    __m256i lhs_mat_ymm_0123_70 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 448 + 512 * sb)));
+                    __m256i lhs_mat_ymm_01_70 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_70, lhs_mat_ymm_0123_70, 0);
+                    __m256i lhs_mat_ymm_23_70 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_70, lhs_mat_ymm_0123_70, 17);
+                    __m256i lhs_mat_ymm_0123_71 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 480 + 512 * sb)));
+                    __m256i lhs_mat_ymm_01_71 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_71, lhs_mat_ymm_0123_71, 0);
+                    __m256i lhs_mat_ymm_23_71 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_71, lhs_mat_ymm_0123_71, 17);
+
+                    __m512i lhs_mat_01_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_00), lhs_mat_ymm_01_00, 1);
+                    __m512i lhs_mat_23_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_00), lhs_mat_ymm_23_00, 1);
+                    __m512i lhs_mat_01_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_01), lhs_mat_ymm_01_01, 1);
+                    __m512i lhs_mat_23_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_01), lhs_mat_ymm_23_01, 1);
+
+                    __m512i lhs_mat_01_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_10), lhs_mat_ymm_01_10, 1);
+                    __m512i lhs_mat_23_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_10), lhs_mat_ymm_23_10, 1);
+                    __m512i lhs_mat_01_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_11), lhs_mat_ymm_01_11, 1);
+                    __m512i lhs_mat_23_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_11), lhs_mat_ymm_23_11, 1);
+
+                    __m512i lhs_mat_01_20 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_20), lhs_mat_ymm_01_20, 1);
+                    __m512i lhs_mat_23_20 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_20), lhs_mat_ymm_23_20, 1);
+                    __m512i lhs_mat_01_21 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_21), lhs_mat_ymm_01_21, 1);
+                    __m512i lhs_mat_23_21 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_21), lhs_mat_ymm_23_21, 1);
+
+                    __m512i lhs_mat_01_30 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_30), lhs_mat_ymm_01_30, 1);
+                    __m512i lhs_mat_23_30 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_30), lhs_mat_ymm_23_30, 1);
+                    __m512i lhs_mat_01_31 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_31), lhs_mat_ymm_01_31, 1);
+                    __m512i lhs_mat_23_31 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_31), lhs_mat_ymm_23_31, 1);
+
+                    __m512i lhs_mat_01_40 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_40), lhs_mat_ymm_01_40, 1);
+                    __m512i lhs_mat_23_40 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_40), lhs_mat_ymm_23_40, 1);
+                    __m512i lhs_mat_01_41 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_41), lhs_mat_ymm_01_41, 1);
+                    __m512i lhs_mat_23_41 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_41), lhs_mat_ymm_23_41, 1);
+
+                    __m512i lhs_mat_01_50 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_50), lhs_mat_ymm_01_50, 1);
+                    __m512i lhs_mat_23_50 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_50), lhs_mat_ymm_23_50, 1);
+                    __m512i lhs_mat_01_51 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_51), lhs_mat_ymm_01_51, 1);
+                    __m512i lhs_mat_23_51 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_51), lhs_mat_ymm_23_51, 1);
+
+                    __m512i lhs_mat_01_60 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_60), lhs_mat_ymm_01_60, 1);
+                    __m512i lhs_mat_23_60 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_60), lhs_mat_ymm_23_60, 1);
+                    __m512i lhs_mat_01_61 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_61), lhs_mat_ymm_01_61, 1);
+                    __m512i lhs_mat_23_61 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_61), lhs_mat_ymm_23_61, 1);
+
+                    __m512i lhs_mat_01_70 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_70), lhs_mat_ymm_01_70, 1);
+                    __m512i lhs_mat_23_70 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_70), lhs_mat_ymm_23_70, 1);
+                    __m512i lhs_mat_01_71 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_71), lhs_mat_ymm_01_71, 1);
+                    __m512i lhs_mat_23_71 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_71), lhs_mat_ymm_23_71, 1);
+
+                    // Bsums are loaded for the different Q8_K blocks
+                    __m128i lhs_raw_bsums_01_0123 = _mm_loadu_si128((const __m128i *)((a_ptr[b].bsums + 32 * sb)));
+                    __m128i lhs_raw_bsums_23_0123 = _mm_loadu_si128((const __m128i *)(a_ptr[b].bsums + 8 + 32 * sb));
+                    __m128i lhs_raw_bsums_01_4567 = _mm_loadu_si128((const __m128i *)((a_ptr[b].bsums + 16 + 32 * sb)));
+                    __m128i lhs_raw_bsums_23_4567 = _mm_loadu_si128((const __m128i *)(a_ptr[b].bsums + 24 + 32 * sb));
+
+                    __m256i lhs_bsums_ymm_01_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_01_0123), lhs_raw_bsums_01_0123, 1);
+                    __m512i lhs_bsums_01_0123 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_ymm_01_0123), lhs_bsums_ymm_01_0123, 1);
+                    __m256i lhs_bsums_ymm_23_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_0123), lhs_raw_bsums_23_0123, 1);
+                    __m512i lhs_bsums_23_0123 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_ymm_23_0123), lhs_bsums_ymm_23_0123, 1);
+                    __m256i lhs_bsums_ymm_01_4567 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_01_4567), lhs_raw_bsums_01_4567, 1);
+                    __m512i lhs_bsums_01_4567 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_ymm_01_4567), lhs_bsums_ymm_01_4567, 1);
+                    __m256i lhs_bsums_ymm_23_4567 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_4567), lhs_raw_bsums_23_4567, 1);
+                    __m512i lhs_bsums_23_4567 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_ymm_23_4567), lhs_bsums_ymm_23_4567, 1);
+
+                    // Shuffle pattern one - left side input
+                    const __m512i lhs_mat_01_00_sp1 = _mm512_shuffle_epi32(lhs_mat_01_00, (_MM_PERM_ENUM)160); //A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3)
+                    const __m512i lhs_mat_23_00_sp1 = _mm512_shuffle_epi32(lhs_mat_23_00, (_MM_PERM_ENUM)160); //A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3)
+
+                    const __m512i lhs_mat_01_01_sp1 = _mm512_shuffle_epi32(lhs_mat_01_01, (_MM_PERM_ENUM)160); //A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11)
+                    const __m512i lhs_mat_23_01_sp1 = _mm512_shuffle_epi32(lhs_mat_23_01, (_MM_PERM_ENUM)160); //A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11)
+
+                    const __m512i lhs_mat_01_10_sp1 = _mm512_shuffle_epi32(lhs_mat_01_10, (_MM_PERM_ENUM)160); //A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3)
+                    const __m512i lhs_mat_23_10_sp1 = _mm512_shuffle_epi32(lhs_mat_23_10, (_MM_PERM_ENUM)160); //A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3)
+
+                    const __m512i lhs_mat_01_11_sp1 = _mm512_shuffle_epi32(lhs_mat_01_11, (_MM_PERM_ENUM)160); //A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11)
+                    const __m512i lhs_mat_23_11_sp1 = _mm512_shuffle_epi32(lhs_mat_23_11, (_MM_PERM_ENUM)160); //A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11)
+
+                    const __m512i lhs_mat_01_20_sp1 = _mm512_shuffle_epi32(lhs_mat_01_20, (_MM_PERM_ENUM)160); //A20(0-3) A20(0-3) A21(0-3) A21(0-3) A20(0-3) A20(0-3) A21(0-3) A21(0-3) A20(0-3) A20(0-3) A21(0-3) A21(0-3) A20(0-3) A20(0-3) A21(0-3) A21(0-3)
+                    const __m512i lhs_mat_23_20_sp1 = _mm512_shuffle_epi32(lhs_mat_23_20, (_MM_PERM_ENUM)160); //A22(0-3) A22(0-3) A23(0-3) A23(0-3) A22(0-3) A22(0-3) A23(0-3) A23(0-3) A22(0-3) A22(0-3) A23(0-3) A23(0-3) A22(0-3) A22(0-3) A23(0-3) A23(0-3)
+
+                    const __m512i lhs_mat_01_21_sp1 = _mm512_shuffle_epi32(lhs_mat_01_21, (_MM_PERM_ENUM)160); //A20(8-11) A20(8-11) A21(8-11) A21(8-11) A20(8-11) A20(8-11) A21(8-11) A21(8-11) A20(8-11) A20(8-11) A21(8-11) A21(8-11) A20(8-11) A20(8-11) A21(8-11) A21(8-11)
+                    const __m512i lhs_mat_23_21_sp1 = _mm512_shuffle_epi32(lhs_mat_23_21, (_MM_PERM_ENUM)160); //A22(8-11) A22(8-11) A23(8-11) A23(8-11) A22(8-11) A22(8-11) A23(8-11) A23(8-11) A22(8-11) A22(8-11) A23(8-11) A23(8-11) A22(8-11) A22(8-11) A23(8-11) A23(8-11)
+
+                    const __m512i lhs_mat_01_30_sp1 = _mm512_shuffle_epi32(lhs_mat_01_30, (_MM_PERM_ENUM)160); //A30(0-3) A30(0-3) A31(0-3) A31(0-3) A30(0-3) A30(0-3) A31(0-3) A31(0-3) A30(0-3) A30(0-3) A31(0-3) A31(0-3) A30(0-3) A30(0-3) A31(0-3) A31(0-3)
+                    const __m512i lhs_mat_23_30_sp1 = _mm512_shuffle_epi32(lhs_mat_23_30, (_MM_PERM_ENUM)160); //A32(0-3) A32(0-3) A33(0-3) A33(0-3) A32(0-3) A32(0-3) A33(0-3) A33(0-3) A32(0-3) A32(0-3) A33(0-3) A33(0-3) A32(0-3) A32(0-3) A33(0-3) A33(0-3)
+
+                    const __m512i lhs_mat_01_31_sp1 = _mm512_shuffle_epi32(lhs_mat_01_31, (_MM_PERM_ENUM)160); //A30(8-11) A30(8-11) A31(8-11) A31(8-11) A30(8-11) A30(8-11) A31(8-11) A31(8-11) A30(8-11) A30(8-11) A31(8-11) A31(8-11) A30(8-11) A30(8-11) A31(8-11) A31(8-11)
+                    const __m512i lhs_mat_23_31_sp1 = _mm512_shuffle_epi32(lhs_mat_23_31, (_MM_PERM_ENUM)160); //A32(8-11) A32(8-11) A33(8-11) A33(8-11) A32(8-11) A32(8-11) A33(8-11) A33(8-11) A32(8-11) A32(8-11) A33(8-11) A33(8-11) A32(8-11) A32(8-11) A33(8-11) A33(8-11)
+
+                    const __m512i lhs_mat_01_40_sp1 = _mm512_shuffle_epi32(lhs_mat_01_40, (_MM_PERM_ENUM)160); //A40(0-3) A40(0-3) A41(0-3) A41(0-3) A40(0-3) A40(0-3) A41(0-3) A41(0-3) A40(0-3) A40(0-3) A41(0-3) A41(0-3) A40(0-3) A40(0-3) A41(0-3) A41(0-3)
+                    const __m512i lhs_mat_23_40_sp1 = _mm512_shuffle_epi32(lhs_mat_23_40, (_MM_PERM_ENUM)160); //A42(0-3) A42(0-3) A43(0-3) A43(0-3) A42(0-3) A42(0-3) A43(0-3) A43(0-3) A42(0-3) A42(0-3) A43(0-3) A43(0-3) A42(0-3) A42(0-3) A43(0-3) A43(0-3)
+
+                    const __m512i lhs_mat_01_41_sp1 = _mm512_shuffle_epi32(lhs_mat_01_41, (_MM_PERM_ENUM)160); //A40(8-11) A40(8-11) A41(8-11) A41(8-11) A40(8-11) A40(8-11) A41(8-11) A41(8-11) A40(8-11) A40(8-11) A41(8-11) A41(8-11) A40(8-11) A40(8-11) A41(8-11) A41(8-11)
+                    const __m512i lhs_mat_23_41_sp1 = _mm512_shuffle_epi32(lhs_mat_23_41, (_MM_PERM_ENUM)160); //A42(8-11) A42(8-11) A43(8-11) A43(8-11) A42(8-11) A42(8-11) A43(8-11) A43(8-11) A42(8-11) A42(8-11) A43(8-11) A43(8-11) A42(8-11) A42(8-11) A43(8-11) A43(8-11)
+
+                    const __m512i lhs_mat_01_50_sp1 = _mm512_shuffle_epi32(lhs_mat_01_50, (_MM_PERM_ENUM)160); //A50(0-3) A50(0-3) A51(0-3) A51(0-3) A50(0-3) A50(0-3) A51(0-3) A51(0-3) A50(0-3) A50(0-3) A51(0-3) A51(0-3) A50(0-3) A50(0-3) A51(0-3) A51(0-3)
+                    const __m512i lhs_mat_23_50_sp1 = _mm512_shuffle_epi32(lhs_mat_23_50, (_MM_PERM_ENUM)160); //A52(0-3) A52(0-3) A53(0-3) A53(0-3) A52(0-3) A52(0-3) A53(0-3) A53(0-3) A52(0-3) A52(0-3) A53(0-3) A53(0-3) A52(0-3) A52(0-3) A53(0-3) A53(0-3)
+
+                    const __m512i lhs_mat_01_51_sp1 = _mm512_shuffle_epi32(lhs_mat_01_51, (_MM_PERM_ENUM)160); //A50(8-11) A50(8-11) A51(8-11) A51(8-11) A50(8-11) A50(8-11) A51(8-11) A51(8-11) A50(8-11) A50(8-11) A51(8-11) A51(8-11) A50(8-11) A50(8-11) A51(8-11) A51(8-11)
+                    const __m512i lhs_mat_23_51_sp1 = _mm512_shuffle_epi32(lhs_mat_23_51, (_MM_PERM_ENUM)160); //A52(8-11) A52(8-11) A53(8-11) A53(8-11) A52(8-11) A52(8-11) A53(8-11) A53(8-11) A52(8-11) A52(8-11) A53(8-11) A53(8-11) A52(8-11) A52(8-11) A53(8-11) A53(8-11)
+
+                    const __m512i lhs_mat_01_60_sp1 = _mm512_shuffle_epi32(lhs_mat_01_60, (_MM_PERM_ENUM)160); //A60(0-3) A60(0-3) A61(0-3) A61(0-3) A60(0-3) A60(0-3) A61(0-3) A61(0-3) A60(0-3) A60(0-3) A61(0-3) A61(0-3) A60(0-3) A60(0-3) A61(0-3) A61(0-3)
+                    const __m512i lhs_mat_23_60_sp1 = _mm512_shuffle_epi32(lhs_mat_23_60, (_MM_PERM_ENUM)160); //A62(0-3) A62(0-3) A63(0-3) A63(0-3) A62(0-3) A62(0-3) A63(0-3) A63(0-3) A62(0-3) A62(0-3) A63(0-3) A63(0-3) A62(0-3) A62(0-3) A63(0-3) A63(0-3)
+
+                    const __m512i lhs_mat_01_61_sp1 = _mm512_shuffle_epi32(lhs_mat_01_61, (_MM_PERM_ENUM)160); //A60(8-11) A60(8-11) A61(8-11) A61(8-11) A60(8-11) A60(8-11) A61(8-11) A61(8-11) A60(8-11) A60(8-11) A61(8-11) A61(8-11) A60(8-11) A60(8-11) A61(8-11) A61(8-11)
+                    const __m512i lhs_mat_23_61_sp1 = _mm512_shuffle_epi32(lhs_mat_23_61, (_MM_PERM_ENUM)160); //A62(8-11) A62(8-11) A63(8-11) A63(8-11) A62(8-11) A62(8-11) A63(8-11) A63(8-11) A62(8-11) A62(8-11) A63(8-11) A63(8-11) A62(8-11) A62(8-11) A63(8-11) A63(8-11)
+
+                    const __m512i lhs_mat_01_70_sp1 = _mm512_shuffle_epi32(lhs_mat_01_70, (_MM_PERM_ENUM)160); //A70(0-3) A70(0-3) A71(0-3) A71(0-3) A70(0-3) A70(0-3) A71(0-3) A71(0-3) A70(0-3) A70(0-3) A71(0-3) A71(0-3) A70(0-3) A70(0-3) A71(0-3) A71(0-3)
+                    const __m512i lhs_mat_23_70_sp1 = _mm512_shuffle_epi32(lhs_mat_23_70, (_MM_PERM_ENUM)160); //A72(0-3) A72(0-3) A73(0-3) A73(0-3) A72(0-3) A72(0-3) A73(0-3) A73(0-3) A72(0-3) A72(0-3) A73(0-3) A73(0-3) A72(0-3) A72(0-3) A73(0-3) A73(0-3)
+
+                    const __m512i lhs_mat_01_71_sp1 = _mm512_shuffle_epi32(lhs_mat_01_71, (_MM_PERM_ENUM)160); //A70(8-11) A70(8-11) A71(8-11) A71(8-11) A70(8-11) A70(8-11) A71(8-11) A71(8-11) A70(8-11) A70(8-11) A71(8-11) A71(8-11) A70(8-11) A70(8-11) A71(8-11) A71(8-11)
+                    const __m512i lhs_mat_23_71_sp1 = _mm512_shuffle_epi32(lhs_mat_23_71, (_MM_PERM_ENUM)160); //A72(8-11) A72(8-11) A73(8-11) A73(8-11) A72(8-11) A72(8-11) A73(8-11) A73(8-11) A72(8-11) A72(8-11) A73(8-11) A73(8-11) A72(8-11) A72(8-11) A73(8-11) A73(8-11)
+
+                    const __m512i lhs_mat_01_00_sp2 = _mm512_shuffle_epi32(lhs_mat_01_00, (_MM_PERM_ENUM)245); //A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7)
+                    const __m512i lhs_mat_23_00_sp2 = _mm512_shuffle_epi32(lhs_mat_23_00, (_MM_PERM_ENUM)245); //A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7)
+
+                    const __m512i lhs_mat_01_01_sp2 = _mm512_shuffle_epi32(lhs_mat_01_01, (_MM_PERM_ENUM)245); //A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15)
+                    const __m512i lhs_mat_23_01_sp2 = _mm512_shuffle_epi32(lhs_mat_23_01, (_MM_PERM_ENUM)245); //A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15)
+
+                    const __m512i lhs_mat_01_10_sp2 = _mm512_shuffle_epi32(lhs_mat_01_10, (_MM_PERM_ENUM)245); //A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7)
+                    const __m512i lhs_mat_23_10_sp2 = _mm512_shuffle_epi32(lhs_mat_23_10, (_MM_PERM_ENUM)245); //A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7)
+
+                    const __m512i lhs_mat_01_11_sp2 = _mm512_shuffle_epi32(lhs_mat_01_11, (_MM_PERM_ENUM)245); //A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15)
+                    const __m512i lhs_mat_23_11_sp2 = _mm512_shuffle_epi32(lhs_mat_23_11, (_MM_PERM_ENUM)245); //A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15)
+
+                    const __m512i lhs_mat_01_20_sp2 = _mm512_shuffle_epi32(lhs_mat_01_20, (_MM_PERM_ENUM)245); //A20(4-7) A20(4-7) A21(4-7) A21(4-7) A20(4-7) A20(4-7) A21(4-7) A21(4-7) A20(4-7) A20(4-7) A21(4-7) A21(4-7) A20(4-7) A20(4-7) A21(4-7) A21(4-7)
+                    const __m512i lhs_mat_23_20_sp2 = _mm512_shuffle_epi32(lhs_mat_23_20, (_MM_PERM_ENUM)245); //A22(4-7) A22(4-7) A23(4-7) A23(4-7) A22(4-7) A22(4-7) A23(4-7) A23(4-7) A22(4-7) A22(4-7) A23(4-7) A23(4-7) A22(4-7) A22(4-7) A23(4-7) A23(4-7)
+
+                    const __m512i lhs_mat_01_21_sp2 = _mm512_shuffle_epi32(lhs_mat_01_21, (_MM_PERM_ENUM)245); //A20(12-15) A20(12-15) A21(12-15) A21(12-15) A20(12-15) A20(12-15) A21(12-15) A21(12-15) A20(12-15) A20(12-15) A21(12-15) A21(12-15) A20(12-15) A20(12-15) A21(12-15) A21(12-15)
+                    const __m512i lhs_mat_23_21_sp2 = _mm512_shuffle_epi32(lhs_mat_23_21, (_MM_PERM_ENUM)245); //A22(12-15) A22(12-15) A23(12-15) A23(12-15) A22(12-15) A22(12-15) A23(12-15) A23(12-15) A22(12-15) A22(12-15) A23(12-15) A23(12-15) A22(12-15) A22(12-15) A23(12-15) A23(12-15)
+
+                    const __m512i lhs_mat_01_30_sp2 = _mm512_shuffle_epi32(lhs_mat_01_30, (_MM_PERM_ENUM)245); //A30(4-7) A30(4-7) A31(4-7) A31(4-7) A30(4-7) A30(4-7) A31(4-7) A31(4-7) A30(4-7) A30(4-7) A31(4-7) A31(4-7) A30(4-7) A30(4-7) A31(4-7) A31(4-7)
+                    const __m512i lhs_mat_23_30_sp2 = _mm512_shuffle_epi32(lhs_mat_23_30, (_MM_PERM_ENUM)245); //A32(4-7) A32(4-7) A33(4-7) A33(4-7) A32(4-7) A32(4-7) A33(4-7) A33(4-7) A32(4-7) A32(4-7) A33(4-7) A33(4-7) A32(4-7) A32(4-7) A33(4-7) A33(4-7)
+
+                    const __m512i lhs_mat_01_31_sp2 = _mm512_shuffle_epi32(lhs_mat_01_31, (_MM_PERM_ENUM)245); //A30(12-15) A30(12-15) A31(12-15) A31(12-15) A30(12-15) A30(12-15) A31(12-15) A31(12-15) A30(12-15) A30(12-15) A31(12-15) A31(12-15) A30(12-15) A30(12-15) A31(12-15) A31(12-15)
+                    const __m512i lhs_mat_23_31_sp2 = _mm512_shuffle_epi32(lhs_mat_23_31, (_MM_PERM_ENUM)245); //A32(12-15) A32(12-15) A33(12-15) A33(12-15) A32(12-15) A32(12-15) A33(12-15) A33(12-15) A32(12-15) A32(12-15) A33(12-15) A33(12-15) A32(12-15) A32(12-15) A33(12-15) A33(12-15)
+
+                    const __m512i lhs_mat_01_40_sp2 = _mm512_shuffle_epi32(lhs_mat_01_40, (_MM_PERM_ENUM)245); //A40(4-7) A40(4-7) A41(4-7) A41(4-7) A40(4-7) A40(4-7) A41(4-7) A41(4-7) A40(4-7) A40(4-7) A41(4-7) A41(4-7) A40(4-7) A40(4-7) A41(4-7) A41(4-7)
+                    const __m512i lhs_mat_23_40_sp2 = _mm512_shuffle_epi32(lhs_mat_23_40, (_MM_PERM_ENUM)245); //A42(4-7) A42(4-7) A43(4-7) A43(4-7) A42(4-7) A42(4-7) A43(4-7) A43(4-7) A42(4-7) A42(4-7) A43(4-7) A43(4-7) A42(4-7) A42(4-7) A43(4-7) A43(4-7)
+
+                    const __m512i lhs_mat_01_41_sp2 = _mm512_shuffle_epi32(lhs_mat_01_41, (_MM_PERM_ENUM)245); //A40(12-15) A40(12-15) A41(12-15) A41(12-15) A40(12-15) A40(12-15) A41(12-15) A41(12-15) A40(12-15) A40(12-15) A41(12-15) A41(12-15) A40(12-15) A40(12-15) A41(12-15) A41(12-15)
+                    const __m512i lhs_mat_23_41_sp2 = _mm512_shuffle_epi32(lhs_mat_23_41, (_MM_PERM_ENUM)245); //A42(12-15) A42(12-15) A43(12-15) A43(12-15) A42(12-15) A42(12-15) A43(12-15) A43(12-15) A42(12-15) A42(12-15) A43(12-15) A43(12-15) A42(12-15) A42(12-15) A43(12-15) A43(12-15)
+
+                    const __m512i lhs_mat_01_50_sp2 = _mm512_shuffle_epi32(lhs_mat_01_50, (_MM_PERM_ENUM)245); //A50(4-7) A50(4-7) A51(4-7) A51(4-7) A50(4-7) A50(4-7) A51(4-7) A51(4-7) A50(4-7) A50(4-7) A51(4-7) A51(4-7) A50(4-7) A50(4-7) A51(4-7) A51(4-7)
+                    const __m512i lhs_mat_23_50_sp2 = _mm512_shuffle_epi32(lhs_mat_23_50, (_MM_PERM_ENUM)245); //A52(4-7) A52(4-7) A53(4-7) A53(4-7) A52(4-7) A52(4-7) A53(4-7) A53(4-7) A52(4-7) A52(4-7) A53(4-7) A53(4-7) A52(4-7) A52(4-7) A53(4-7) A53(4-7)
+
+                    const __m512i lhs_mat_01_51_sp2 = _mm512_shuffle_epi32(lhs_mat_01_51, (_MM_PERM_ENUM)245); //A50(12-15) A50(12-15) A51(12-15) A51(12-15) A50(12-15) A50(12-15) A51(12-15) A51(12-15) A50(12-15) A50(12-15) A51(12-15) A51(12-15) A50(12-15) A50(12-15) A51(12-15) A51(12-15)
+                    const __m512i lhs_mat_23_51_sp2 = _mm512_shuffle_epi32(lhs_mat_23_51, (_MM_PERM_ENUM)245); //A52(12-15) A52(12-15) A53(12-15) A53(12-15) A52(12-15) A52(12-15) A53(12-15) A53(12-15) A52(12-15) A52(12-15) A53(12-15) A53(12-15) A52(12-15) A52(12-15) A53(12-15) A53(12-15)
+
+                    const __m512i lhs_mat_01_60_sp2 = _mm512_shuffle_epi32(lhs_mat_01_60, (_MM_PERM_ENUM)245); //A60(4-7) A60(4-7) A61(4-7) A61(4-7) A60(4-7) A60(4-7) A61(4-7) A61(4-7) A60(4-7) A60(4-7) A61(4-7) A61(4-7) A60(4-7) A60(4-7) A61(4-7) A61(4-7)
+                    const __m512i lhs_mat_23_60_sp2 = _mm512_shuffle_epi32(lhs_mat_23_60, (_MM_PERM_ENUM)245); //A62(4-7) A62(4-7) A63(4-7) A63(4-7) A62(4-7) A62(4-7) A63(4-7) A63(4-7) A62(4-7) A62(4-7) A63(4-7) A63(4-7) A62(4-7) A62(4-7) A63(4-7) A63(4-7)
+
+                    const __m512i lhs_mat_01_61_sp2 = _mm512_shuffle_epi32(lhs_mat_01_61, (_MM_PERM_ENUM)245); //A60(12-15) A60(12-15) A61(12-15) A61(12-15) A60(12-15) A60(12-15) A61(12-15) A61(12-15) A60(12-15) A60(12-15) A61(12-15) A61(12-15) A60(12-15) A60(12-15) A61(12-15) A61(12-15)
+                    const __m512i lhs_mat_23_61_sp2 = _mm512_shuffle_epi32(lhs_mat_23_61, (_MM_PERM_ENUM)245); //A62(12-15) A62(12-15) A63(12-15) A63(12-15) A62(12-15) A62(12-15) A63(12-15) A63(12-15) A62(12-15) A62(12-15) A63(12-15) A63(12-15) A62(12-15) A62(12-15) A63(12-15) A63(12-15)
+
+                    const __m512i lhs_mat_01_70_sp2 = _mm512_shuffle_epi32(lhs_mat_01_70, (_MM_PERM_ENUM)245); //A70(4-7) A70(4-7) A71(4-7) A71(4-7) A70(4-7) A70(4-7) A71(4-7) A71(4-7) A70(4-7) A70(4-7) A71(4-7) A71(4-7) A70(4-7) A70(4-7) A71(4-7) A71(4-7)
+                    const __m512i lhs_mat_23_70_sp2 = _mm512_shuffle_epi32(lhs_mat_23_70, (_MM_PERM_ENUM)245); //A72(4-7) A72(4-7) A73(4-7) A73(4-7) A72(4-7) A72(4-7) A73(4-7) A73(4-7) A72(4-7) A72(4-7) A73(4-7) A73(4-7) A72(4-7) A72(4-7) A73(4-7) A73(4-7)
+
+                    const __m512i lhs_mat_01_71_sp2 = _mm512_shuffle_epi32(lhs_mat_01_71, (_MM_PERM_ENUM)245); //A70(12-15) A70(12-15) A71(12-15) A71(12-15) A70(12-15) A70(12-15) A71(12-15) A71(12-15) A70(12-15) A70(12-15) A71(12-15) A71(12-15) A70(12-15) A70(12-15) A71(12-15) A71(12-15)
+                    const __m512i lhs_mat_23_71_sp2 = _mm512_shuffle_epi32(lhs_mat_23_71, (_MM_PERM_ENUM)245); //A72(12-15) A72(12-15) A73(12-15) A73(12-15) A72(12-15) A72(12-15) A73(12-15) A73(12-15) A72(12-15) A72(12-15) A73(12-15) A73(12-15) A72(12-15) A72(12-15) A73(12-15) A73(12-15)
+
+                    // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+                    __m512i iacc_mat_00_0_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_01_00_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_01_01_sp1));
+                    __m512i iacc_mat_01_0_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_01_00_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_01_01_sp1));
+
+                    __m512i iacc_mat_10_0_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_23_00_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_23_01_sp1));
+                    __m512i iacc_mat_11_0_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_23_00_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_23_01_sp1));
+
+                    __m512i iacc_mat_00_1_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_01_10_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_01_11_sp1));
+                    __m512i iacc_mat_01_1_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_01_10_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_01_11_sp1));
+
+                    __m512i iacc_mat_10_1_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_23_10_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_23_11_sp1));
+                    __m512i iacc_mat_11_1_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_23_10_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_23_11_sp1));
+
+                    __m512i iacc_mat_00_2_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_20_sp1, lhs_mat_01_20_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_21_sp1, lhs_mat_01_21_sp1));
+                    __m512i iacc_mat_01_2_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_20_sp1, lhs_mat_01_20_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_21_sp1, lhs_mat_01_21_sp1));
+
+                    __m512i iacc_mat_10_2_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_20_sp1, lhs_mat_23_20_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_21_sp1, lhs_mat_23_21_sp1));
+                    __m512i iacc_mat_11_2_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_20_sp1, lhs_mat_23_20_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_21_sp1, lhs_mat_23_21_sp1));
+
+                    __m512i iacc_mat_00_3_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_30_sp1, lhs_mat_01_30_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_31_sp1, lhs_mat_01_31_sp1));
+                    __m512i iacc_mat_01_3_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_30_sp1, lhs_mat_01_30_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_31_sp1, lhs_mat_01_31_sp1));
+
+                    __m512i iacc_mat_10_3_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_30_sp1, lhs_mat_23_30_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_31_sp1, lhs_mat_23_31_sp1));
+                    __m512i iacc_mat_11_3_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_30_sp1, lhs_mat_23_30_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_31_sp1, lhs_mat_23_31_sp1));
+
+                    __m512i iacc_mat_00_4_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_40_sp1, lhs_mat_01_40_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_41_sp1, lhs_mat_01_41_sp1));
+                    __m512i iacc_mat_01_4_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_40_sp1, lhs_mat_01_40_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_41_sp1, lhs_mat_01_41_sp1));
+
+                    __m512i iacc_mat_10_4_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_40_sp1, lhs_mat_23_40_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_41_sp1, lhs_mat_23_41_sp1));
+                    __m512i iacc_mat_11_4_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_40_sp1, lhs_mat_23_40_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_41_sp1, lhs_mat_23_41_sp1));
+
+                    __m512i iacc_mat_00_5_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_50_sp1, lhs_mat_01_50_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_51_sp1, lhs_mat_01_51_sp1));
+                    __m512i iacc_mat_01_5_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_50_sp1, lhs_mat_01_50_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_51_sp1, lhs_mat_01_51_sp1));
+
+                    __m512i iacc_mat_10_5_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_50_sp1, lhs_mat_23_50_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_51_sp1, lhs_mat_23_51_sp1));
+                    __m512i iacc_mat_11_5_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_50_sp1, lhs_mat_23_50_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_51_sp1, lhs_mat_23_51_sp1));
+
+                    __m512i iacc_mat_00_6_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_60_sp1, lhs_mat_01_60_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_61_sp1, lhs_mat_01_61_sp1));
+                    __m512i iacc_mat_01_6_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_60_sp1, lhs_mat_01_60_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_61_sp1, lhs_mat_01_61_sp1));
+
+                    __m512i iacc_mat_10_6_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_60_sp1, lhs_mat_23_60_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_61_sp1, lhs_mat_23_61_sp1));
+                    __m512i iacc_mat_11_6_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_60_sp1, lhs_mat_23_60_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_61_sp1, lhs_mat_23_61_sp1));
+
+                    __m512i iacc_mat_00_7_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_70_sp1, lhs_mat_01_70_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_71_sp1, lhs_mat_01_71_sp1));
+                    __m512i iacc_mat_01_7_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_70_sp1, lhs_mat_01_70_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_71_sp1, lhs_mat_01_71_sp1));
+
+                    __m512i iacc_mat_10_7_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_70_sp1, lhs_mat_23_70_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_71_sp1, lhs_mat_23_71_sp1));
+                    __m512i iacc_mat_11_7_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_70_sp1, lhs_mat_23_70_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_71_sp1, lhs_mat_23_71_sp1));
+
+
+                    __m512i iacc_mat_00_0_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_01_00_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_01_01_sp2));
+                    __m512i iacc_mat_01_0_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_01_00_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_01_01_sp2));
+
+                    __m512i iacc_mat_10_0_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_23_00_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_23_01_sp2));
+                    __m512i iacc_mat_11_0_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_23_00_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_23_01_sp2));
+
+                    __m512i iacc_mat_00_1_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_01_10_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_01_11_sp2));
+                    __m512i iacc_mat_01_1_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_01_10_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_01_11_sp2));
+
+                    __m512i iacc_mat_10_1_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_23_10_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_23_11_sp2));
+                    __m512i iacc_mat_11_1_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_23_10_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_23_11_sp2));
+
+                    __m512i iacc_mat_00_2_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_20_sp2, lhs_mat_01_20_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_21_sp2, lhs_mat_01_21_sp2));
+                    __m512i iacc_mat_01_2_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_20_sp2, lhs_mat_01_20_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_21_sp2, lhs_mat_01_21_sp2));
+
+                    __m512i iacc_mat_10_2_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_20_sp2, lhs_mat_23_20_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_21_sp2, lhs_mat_23_21_sp2));
+                    __m512i iacc_mat_11_2_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_20_sp2, lhs_mat_23_20_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_21_sp2, lhs_mat_23_21_sp2));
+
+                    __m512i iacc_mat_00_3_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_30_sp2, lhs_mat_01_30_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_31_sp2, lhs_mat_01_31_sp2));
+                    __m512i iacc_mat_01_3_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_30_sp2, lhs_mat_01_30_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_31_sp2, lhs_mat_01_31_sp2));
+
+                    __m512i iacc_mat_10_3_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_30_sp2, lhs_mat_23_30_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_31_sp2, lhs_mat_23_31_sp2));
+                    __m512i iacc_mat_11_3_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_30_sp2, lhs_mat_23_30_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_31_sp2, lhs_mat_23_31_sp2));
+
+                    __m512i iacc_mat_00_4_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_40_sp2, lhs_mat_01_40_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_41_sp2, lhs_mat_01_41_sp2));
+                    __m512i iacc_mat_01_4_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_40_sp2, lhs_mat_01_40_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_41_sp2, lhs_mat_01_41_sp2));
+
+                    __m512i iacc_mat_10_4_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_40_sp2, lhs_mat_23_40_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_41_sp2, lhs_mat_23_41_sp2));
+                    __m512i iacc_mat_11_4_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_40_sp2, lhs_mat_23_40_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_41_sp2, lhs_mat_23_41_sp2));
+
+                    __m512i iacc_mat_00_5_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_50_sp2, lhs_mat_01_50_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_51_sp2, lhs_mat_01_51_sp2));
+                    __m512i iacc_mat_01_5_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_50_sp2, lhs_mat_01_50_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_51_sp2, lhs_mat_01_51_sp2));
+
+                    __m512i iacc_mat_10_5_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_50_sp2, lhs_mat_23_50_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_51_sp2, lhs_mat_23_51_sp2));
+                    __m512i iacc_mat_11_5_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_50_sp2, lhs_mat_23_50_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_51_sp2, lhs_mat_23_51_sp2));
+
+                    __m512i iacc_mat_00_6_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_60_sp2, lhs_mat_01_60_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_61_sp2, lhs_mat_01_61_sp2));
+                    __m512i iacc_mat_01_6_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_60_sp2, lhs_mat_01_60_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_61_sp2, lhs_mat_01_61_sp2));
+
+                    __m512i iacc_mat_10_6_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_60_sp2, lhs_mat_23_60_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_61_sp2, lhs_mat_23_61_sp2));
+                    __m512i iacc_mat_11_6_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_60_sp2, lhs_mat_23_60_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_61_sp2, lhs_mat_23_61_sp2));
+
+                    __m512i iacc_mat_00_7_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_70_sp2, lhs_mat_01_70_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_71_sp2, lhs_mat_01_71_sp2));
+                    __m512i iacc_mat_01_7_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_70_sp2, lhs_mat_01_70_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_71_sp2, lhs_mat_01_71_sp2));
+
+                    __m512i iacc_mat_10_7_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_70_sp2, lhs_mat_23_70_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_71_sp2, lhs_mat_23_71_sp2));
+                    __m512i iacc_mat_11_7_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_70_sp2, lhs_mat_23_70_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_71_sp2, lhs_mat_23_71_sp2));
+
+                    // Combine results from both shuffle patterns for each output block
+                    __m512i iacc_mat_00_0 = _mm512_add_epi16(iacc_mat_00_0_sp1, iacc_mat_00_0_sp2);
+                    __m512i iacc_mat_01_0 = _mm512_add_epi16(iacc_mat_01_0_sp1, iacc_mat_01_0_sp2);
+                    __m512i iacc_mat_10_0 = _mm512_add_epi16(iacc_mat_10_0_sp1, iacc_mat_10_0_sp2);
+                    __m512i iacc_mat_11_0 = _mm512_add_epi16(iacc_mat_11_0_sp1, iacc_mat_11_0_sp2);
+
+                    __m512i iacc_mat_00_1 = _mm512_add_epi16(iacc_mat_00_1_sp1, iacc_mat_00_1_sp2);
+                    __m512i iacc_mat_01_1 = _mm512_add_epi16(iacc_mat_01_1_sp1, iacc_mat_01_1_sp2);
+                    __m512i iacc_mat_10_1 = _mm512_add_epi16(iacc_mat_10_1_sp1, iacc_mat_10_1_sp2);
+                    __m512i iacc_mat_11_1 = _mm512_add_epi16(iacc_mat_11_1_sp1, iacc_mat_11_1_sp2);
+
+                    __m512i iacc_mat_00_2 = _mm512_add_epi16(iacc_mat_00_2_sp1, iacc_mat_00_2_sp2);
+                    __m512i iacc_mat_01_2 = _mm512_add_epi16(iacc_mat_01_2_sp1, iacc_mat_01_2_sp2);
+                    __m512i iacc_mat_10_2 = _mm512_add_epi16(iacc_mat_10_2_sp1, iacc_mat_10_2_sp2);
+                    __m512i iacc_mat_11_2 = _mm512_add_epi16(iacc_mat_11_2_sp1, iacc_mat_11_2_sp2);
+
+                    __m512i iacc_mat_00_3 = _mm512_add_epi16(iacc_mat_00_3_sp1, iacc_mat_00_3_sp2);
+                    __m512i iacc_mat_01_3 = _mm512_add_epi16(iacc_mat_01_3_sp1, iacc_mat_01_3_sp2);
+                    __m512i iacc_mat_10_3 = _mm512_add_epi16(iacc_mat_10_3_sp1, iacc_mat_10_3_sp2);
+                    __m512i iacc_mat_11_3 = _mm512_add_epi16(iacc_mat_11_3_sp1, iacc_mat_11_3_sp2);
+
+                    __m512i iacc_mat_00_4 = _mm512_add_epi16(iacc_mat_00_4_sp1, iacc_mat_00_4_sp2);
+                    __m512i iacc_mat_01_4 = _mm512_add_epi16(iacc_mat_01_4_sp1, iacc_mat_01_4_sp2);
+                    __m512i iacc_mat_10_4 = _mm512_add_epi16(iacc_mat_10_4_sp1, iacc_mat_10_4_sp2);
+                    __m512i iacc_mat_11_4 = _mm512_add_epi16(iacc_mat_11_4_sp1, iacc_mat_11_4_sp2);
+
+                    __m512i iacc_mat_00_5 = _mm512_add_epi16(iacc_mat_00_5_sp1, iacc_mat_00_5_sp2);
+                    __m512i iacc_mat_01_5 = _mm512_add_epi16(iacc_mat_01_5_sp1, iacc_mat_01_5_sp2);
+                    __m512i iacc_mat_10_5 = _mm512_add_epi16(iacc_mat_10_5_sp1, iacc_mat_10_5_sp2);
+                    __m512i iacc_mat_11_5 = _mm512_add_epi16(iacc_mat_11_5_sp1, iacc_mat_11_5_sp2);
+
+                    __m512i iacc_mat_00_6 = _mm512_add_epi16(iacc_mat_00_6_sp1, iacc_mat_00_6_sp2);
+                    __m512i iacc_mat_01_6 = _mm512_add_epi16(iacc_mat_01_6_sp1, iacc_mat_01_6_sp2);
+                    __m512i iacc_mat_10_6 = _mm512_add_epi16(iacc_mat_10_6_sp1, iacc_mat_10_6_sp2);
+                    __m512i iacc_mat_11_6 = _mm512_add_epi16(iacc_mat_11_6_sp1, iacc_mat_11_6_sp2);
+
+                    __m512i iacc_mat_00_7 = _mm512_add_epi16(iacc_mat_00_7_sp1, iacc_mat_00_7_sp2);
+                    __m512i iacc_mat_01_7 = _mm512_add_epi16(iacc_mat_01_7_sp1, iacc_mat_01_7_sp2);
+                    __m512i iacc_mat_10_7 = _mm512_add_epi16(iacc_mat_10_7_sp1, iacc_mat_10_7_sp2);
+                    __m512i iacc_mat_11_7 = _mm512_add_epi16(iacc_mat_11_7_sp1, iacc_mat_11_7_sp2);
+
+                    // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+                    iacc_mat_00_0 = _mm512_madd_epi16(iacc_mat_00_0, scale_014589CD_0);
+                    iacc_mat_01_0 = _mm512_madd_epi16(iacc_mat_01_0, scale_2367ABEF_0);
+                    iacc_mat_10_0 = _mm512_madd_epi16(iacc_mat_10_0, scale_014589CD_0);
+                    iacc_mat_11_0 = _mm512_madd_epi16(iacc_mat_11_0, scale_2367ABEF_0);
+
+                    iacc_mat_00_1 = _mm512_madd_epi16(iacc_mat_00_1, scale_014589CD_1);
+                    iacc_mat_01_1 = _mm512_madd_epi16(iacc_mat_01_1, scale_2367ABEF_1);
+                    iacc_mat_10_1 = _mm512_madd_epi16(iacc_mat_10_1, scale_014589CD_1);
+                    iacc_mat_11_1 = _mm512_madd_epi16(iacc_mat_11_1, scale_2367ABEF_1);
+
+                    iacc_mat_00_2 = _mm512_madd_epi16(iacc_mat_00_2, scale_014589CD_2);
+                    iacc_mat_01_2 = _mm512_madd_epi16(iacc_mat_01_2, scale_2367ABEF_2);
+                    iacc_mat_10_2 = _mm512_madd_epi16(iacc_mat_10_2, scale_014589CD_2);
+                    iacc_mat_11_2 = _mm512_madd_epi16(iacc_mat_11_2, scale_2367ABEF_2);
+
+                    iacc_mat_00_3 = _mm512_madd_epi16(iacc_mat_00_3, scale_014589CD_3);
+                    iacc_mat_01_3 = _mm512_madd_epi16(iacc_mat_01_3, scale_2367ABEF_3);
+                    iacc_mat_10_3 = _mm512_madd_epi16(iacc_mat_10_3, scale_014589CD_3);
+                    iacc_mat_11_3 = _mm512_madd_epi16(iacc_mat_11_3, scale_2367ABEF_3);
+
+                    iacc_mat_00_4 = _mm512_madd_epi16(iacc_mat_00_4, scale_014589CD_4);
+                    iacc_mat_01_4 = _mm512_madd_epi16(iacc_mat_01_4, scale_2367ABEF_4);
+                    iacc_mat_10_4 = _mm512_madd_epi16(iacc_mat_10_4, scale_014589CD_4);
+                    iacc_mat_11_4 = _mm512_madd_epi16(iacc_mat_11_4, scale_2367ABEF_4);
+
+                    iacc_mat_00_5 = _mm512_madd_epi16(iacc_mat_00_5, scale_014589CD_5);
+                    iacc_mat_01_5 = _mm512_madd_epi16(iacc_mat_01_5, scale_2367ABEF_5);
+                    iacc_mat_10_5 = _mm512_madd_epi16(iacc_mat_10_5, scale_014589CD_5);
+                    iacc_mat_11_5 = _mm512_madd_epi16(iacc_mat_11_5, scale_2367ABEF_5);
+
+                    iacc_mat_00_6 = _mm512_madd_epi16(iacc_mat_00_6, scale_014589CD_6);
+                    iacc_mat_01_6 = _mm512_madd_epi16(iacc_mat_01_6, scale_2367ABEF_6);
+                    iacc_mat_10_6 = _mm512_madd_epi16(iacc_mat_10_6, scale_014589CD_6);
+                    iacc_mat_11_6 = _mm512_madd_epi16(iacc_mat_11_6, scale_2367ABEF_6);
+
+                    iacc_mat_00_7 = _mm512_madd_epi16(iacc_mat_00_7, scale_014589CD_7);
+                    iacc_mat_01_7 = _mm512_madd_epi16(iacc_mat_01_7, scale_2367ABEF_7);
+                    iacc_mat_10_7 = _mm512_madd_epi16(iacc_mat_10_7, scale_014589CD_7);
+                    iacc_mat_11_7 = _mm512_madd_epi16(iacc_mat_11_7, scale_2367ABEF_7);
+
+                    __m512i iacc_mat_00 = _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(iacc_mat_00_0, iacc_mat_00_1), _mm512_add_epi32(iacc_mat_00_2, iacc_mat_00_3)), _mm512_add_epi32(_mm512_add_epi32(iacc_mat_00_4, iacc_mat_00_5), _mm512_add_epi32(iacc_mat_00_6, iacc_mat_00_7)));
+                    __m512i iacc_mat_01 = _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(iacc_mat_01_0, iacc_mat_01_1), _mm512_add_epi32(iacc_mat_01_2, iacc_mat_01_3)), _mm512_add_epi32(_mm512_add_epi32(iacc_mat_01_4, iacc_mat_01_5), _mm512_add_epi32(iacc_mat_01_6, iacc_mat_01_7)));
+                    __m512i iacc_mat_10 = _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(iacc_mat_10_0, iacc_mat_10_1), _mm512_add_epi32(iacc_mat_10_2, iacc_mat_10_3)), _mm512_add_epi32(_mm512_add_epi32(iacc_mat_10_4, iacc_mat_10_5), _mm512_add_epi32(iacc_mat_10_6, iacc_mat_10_7)));
+                    __m512i iacc_mat_11 = _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(iacc_mat_11_0, iacc_mat_11_1), _mm512_add_epi32(iacc_mat_11_2, iacc_mat_11_3)), _mm512_add_epi32(_mm512_add_epi32(iacc_mat_11_4, iacc_mat_11_5), _mm512_add_epi32(iacc_mat_11_6, iacc_mat_11_7)));
+
+                    // Straighten out to make 4 row vectors
+                    __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
+                    __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
+                    __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
+                    __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
+
+                    // Load the scale(d) values for all the 4 Q8_k blocks and repeat it across lanes
+                    const __m128 row_scale_f32_sse = _mm_load_ps(a_ptr[b].d);
+                    const __m256 row_scale_f32_ymm = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
+                    const __m512 row_scale_f32 = _mm512_insertf32x8(_mm512_castps256_ps512(row_scale_f32_ymm), row_scale_f32_ymm, 1);
+
+                    // Multiply with appropiate scales and accumulate (for both d and dmin) below
+                    acc_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
+                    acc_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
+                    acc_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
+                    acc_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]);
+
+                    // Take two bsums from two Q8_Ks at a time and multiply with corresponding mins values from each Q2_K
+                    __m512i iacc_row_min_0_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)0), mins_01);
+                    __m512i iacc_row_min_1_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)170), mins_01);
+                    __m512i iacc_row_min_2_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)0), mins_01);
+                    __m512i iacc_row_min_3_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)170), mins_01);
+
+                    __m512i iacc_row_min_0_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)85), mins_23);
+                    __m512i iacc_row_min_1_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)255), mins_23);
+                    __m512i iacc_row_min_2_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)85), mins_23);
+                    __m512i iacc_row_min_3_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)255), mins_23);
+
+                    __m512i iacc_row_min_0_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)0), mins_45);
+                    __m512i iacc_row_min_1_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)170), mins_45);
+                    __m512i iacc_row_min_2_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)0), mins_45);
+                    __m512i iacc_row_min_3_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)170), mins_45);
+
+                    __m512i iacc_row_min_0_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)85), mins_67);
+                    __m512i iacc_row_min_1_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)255), mins_67);
+                    __m512i iacc_row_min_2_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)85), mins_67);
+                    __m512i iacc_row_min_3_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)255), mins_67);
+
+                    __m512i iacc_row_min_0 = _mm512_add_epi32(_mm512_add_epi32(iacc_row_min_0_01, iacc_row_min_0_23), _mm512_add_epi32(iacc_row_min_0_45,iacc_row_min_0_67));
+                    __m512i iacc_row_min_1 = _mm512_add_epi32(_mm512_add_epi32(iacc_row_min_1_01, iacc_row_min_1_23), _mm512_add_epi32(iacc_row_min_1_45,iacc_row_min_1_67));
+                    __m512i iacc_row_min_2 = _mm512_add_epi32(_mm512_add_epi32(iacc_row_min_2_01, iacc_row_min_2_23), _mm512_add_epi32(iacc_row_min_2_45,iacc_row_min_2_67));
+                    __m512i iacc_row_min_3 = _mm512_add_epi32(_mm512_add_epi32(iacc_row_min_3_01, iacc_row_min_3_23), _mm512_add_epi32(iacc_row_min_3_45,iacc_row_min_3_67));
+
+                    acc_min_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_0), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_min_rows[0]);
+                    acc_min_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_1), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_min_rows[1]);
+                    acc_min_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_2), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_min_rows[2]);
+                    acc_min_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_3), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_min_rows[3]);
+                }
+            }
+            // Store accumlated values
+            for (int i = 0; i < 4; i++) {
+                _mm512_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm512_sub_ps(acc_rows[i], acc_min_rows[i]));
+            }
+        }
+    }
+
+    if (anc != nc) {
+        xstart = anc/8;
+        y = 0;
+    }
+
+#endif //AVX512F
+
+    // Take group of four block_q8_Kx4 structures at each pass of the loop and perform dot product operation
+    for (; y < anr / 4; y += 4) {
+
+        const block_q8_Kx4 * a_ptrs[4];
+
+        a_ptrs[0] = a_ptr_start + (y * nb);
+        for (int i = 0; i < 3; ++i) {
+            a_ptrs[i + 1] = a_ptrs[i] + nb;
+        }
+
+        // Take group of eight block_q2_kx8 structures at each pass of the loop and perform dot product operation
+        for (int64_t x = xstart; x < nc / 8; x++) {
+
+            const block_q2_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
+
+            // Master FP accumulators
+            __m256 acc_rows[16];
+            for (int i = 0; i < 16; i++) {
+                acc_rows[i] = _mm256_setzero_ps();
+            }
+
+            __m256 acc_min_rows[16];
+            for (int i = 0; i < 16; i++) {
+                acc_min_rows[i] = _mm256_setzero_ps();
+            }
+
+            // For super block
+            for (int64_t b = 0; b < nb; b++) {
+                // Delta values - Load the eight scale values of block_q2_kx8
+                const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
+
+                // dmin values - Load the eight dmin values of block_q2_kx8
+                const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin);
+
+                // Loop to iterate over the sixteen sub blocks of a super block - eight sub blocks are processed per iteration
+                for (int sb = 0; sb < QK_K / 128; sb++) {
+
+                    // Load the eight block_q2_K for eight sub blocks quantized values interleaved with each other in chunks of eight bytes - B0,B1 ....B6,B7
+                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + sb * 256));
+                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 128 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 160 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 192 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 224 + sb * 256));
+
+                    // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
+                    //superblock    sub block   which part of sub block
+                    const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+
+                    const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+
+                    const __m256i rhs_raw_mat_0145_2 = _mm256_blend_epi32(rhs_raw_mat_0123_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_2, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_2, requiredOrder), rhs_raw_mat_4567_2, 240);
+
+                    const __m256i rhs_raw_mat_0145_3 = _mm256_blend_epi32(rhs_raw_mat_0123_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_3, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_3, requiredOrder), rhs_raw_mat_4567_3, 240);
+
+                    // 2-bit -> 8-bit
+                    // First sub block of the eight sub blocks processed in the iteration
+                    const __m256i rhs_mat_0145_00 = _mm256_and_si256(rhs_raw_mat_0145_0, m3b); //B00(0-7) B01(0-7) B04(0-7) B05(0-7)
+                    const __m256i rhs_mat_2367_00 = _mm256_and_si256(rhs_raw_mat_2367_0, m3b); //B02(0-7) B03(0-7) B06(0-7) B07(0-7)
+
+                    const __m256i rhs_mat_0145_01 = _mm256_and_si256(rhs_raw_mat_0145_1, m3b); //B00(8-15) B01(8-15) B04(8-15) B05(8-15)
+                    const __m256i rhs_mat_2367_01 = _mm256_and_si256(rhs_raw_mat_2367_1, m3b); //B02(8-15) B03(8-15) B06(8-15) B07(8-15)
+
+                    // Second sub block of the eight sub blocks processed in the iteration
+                    const __m256i rhs_mat_0145_10 = _mm256_and_si256(rhs_raw_mat_0145_2, m3b); //B10(0-7) B11(0-7) B14(0-7) B15(0-7)
+                    const __m256i rhs_mat_2367_10 = _mm256_and_si256(rhs_raw_mat_2367_2, m3b); //B12(0-7) B13(0-7) B16(0-7) B17(0-7)
+
+                    const __m256i rhs_mat_0145_11 = _mm256_and_si256(rhs_raw_mat_0145_3, m3b); //B10(8-15) B11(8-15) B14(8-15) B15(8-15)
+                    const __m256i rhs_mat_2367_11 = _mm256_and_si256(rhs_raw_mat_2367_3, m3b); //B12(8-15) B13(8-15) B16(8-15) B17(8-15)
+
+                    // Third sub block of the eight sub blocks processed in the iteration
+                    const __m256i rhs_mat_0145_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 2), m3b); //B20(0-7) B21(0-7) B24(0-7) B25(0-7)
+                    const __m256i rhs_mat_2367_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 2), m3b); //B22(0-7) B23(0-7) B26(0-7) B27(0-7)
+
+                    const __m256i rhs_mat_0145_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 2), m3b); //B20(8-15) B21(8-15) B24(8-15) B25(8-15)
+                    const __m256i rhs_mat_2367_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 2), m3b); //B22(8-15) B23(8-15) B26(8-15) B27(8-15)
+
+                    // Fourth sub block of the eight sub blocks processed in the iteration
+                    const __m256i rhs_mat_0145_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_2, 2), m3b); //B30(0-7) B31(0-7) B34(0-7) B35(0-7)
+                    const __m256i rhs_mat_2367_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_2, 2), m3b); //B32(0-7) B33(0-7) B36(0-7) B37(0-7)
+
+                    const __m256i rhs_mat_0145_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_3, 2), m3b); //B30(8-15) B31(8-15) B34(8-15) B35(8-15)
+                    const __m256i rhs_mat_2367_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_3, 2), m3b); //B32(8-15) B33(8-15) B36(8-15) B37(8-15)
+
+                    // Fifth sub block of the eight sub blocks processed in the iteration
+                    const __m256i rhs_mat_0145_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m3b); //B40(0-7) B41(0-7) B44(0-7) B45(0-7)
+                    const __m256i rhs_mat_2367_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m3b); //B42(0-7) B43(0-7) B46(0-7) B47(0-7)
+
+                    const __m256i rhs_mat_0145_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m3b); //B40(8-15) B41(8-15) B44(8-15) B45(8-15)
+                    const __m256i rhs_mat_2367_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m3b); //B42(8-15) B43(8-15) B46(8-15) B47(8-15)
+
+                    // Sixth sub block of the eight sub blocks processed in the iteration
+                    const __m256i rhs_mat_0145_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_2, 4), m3b); //B50(0-7) B51(0-7) B54(0-7) B55(0-7)
+                    const __m256i rhs_mat_2367_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_2, 4), m3b); //B52(0-7) B53(0-7) B56(0-7) B57(0-7)
+
+                    const __m256i rhs_mat_0145_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_3, 4), m3b); //B50(8-15) B51(8-15) B54(8-15) B55(8-15)
+                    const __m256i rhs_mat_2367_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_3, 4), m3b); //B52(8-15) B53(8-15) B56(8-15) B57(8-15)
+
+                    // Seventh sub block of the eight sub blocks processed in the iteration
+                    const __m256i rhs_mat_0145_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 6), m3b); //B60(0-7) B61(0-7) B64(0-7) B65(0-7)
+                    const __m256i rhs_mat_2367_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 6), m3b); //B62(0-7) B63(0-7) B66(0-7) B67(0-7)
+
+                    const __m256i rhs_mat_0145_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 6), m3b); //B60(8-15) B61(8-15) B64(8-15) B65(8-15)
+                    const __m256i rhs_mat_2367_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 6), m3b); //B62(8-15) B63(8-15) B66(8-15) B67(8-15)
+
+                    // Eighth sub block of the eight sub blocks processed in the iteration
+                    const __m256i rhs_mat_0145_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_2, 6), m3b); //B70(0-7) B71(0-7) B74(0-7) B75(0-7)
+                    const __m256i rhs_mat_2367_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_2, 6), m3b); //B72(0-7) B73(0-7) B76(0-7) B77(0-7)
+
+                    const __m256i rhs_mat_0145_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_3, 6), m3b); //B70(8-15) B71(8-15) B74(8-15) B75(8-15)
+                    const __m256i rhs_mat_2367_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_3, 6), m3b); //B72(8-15) B73(8-15) B76(8-15) B77(8-15)
+
+                    // Shuffle pattern one - right side input
+                    const __m256i rhs_mat_0145_00_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_00, 136); //B00(0-3) B01(0-3) B00(0-3) B01(0-3) B04(0-3) B05(0-3) B04(0-3) B05(0-3)
+                    const __m256i rhs_mat_2367_00_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_00, 136); //B02(0-3) B03(0-3) B02(0-3) B03(0-3) B06(0-3) B07(0-3) B06(0-3) B07(0-3)
+
+                    const __m256i rhs_mat_0145_01_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_01, 136); //B00(8-11) B01(8-11) B00(8-11) B01(8-11) B04(8-11) B05(8-11) B04(8-11) B05(8-11)
+                    const __m256i rhs_mat_2367_01_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_01, 136); //B02(8-11) B03(8-11) B02(8-11) B03(8-11) B06(8-11) B07(8-11) B06(8-11) B07(8-11)
+
+                    const __m256i rhs_mat_0145_10_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_10, 136); //B10(0-3) B11(0-3) B10(0-3) B11(0-3) B14(0-3) B15(0-3) B14(0-3) B15(0-3)
+                    const __m256i rhs_mat_2367_10_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_10, 136); //B12(0-3) B13(0-3) B12(0-3) B13(0-3) B16(0-3) B17(0-3) B16(0-3) B17(0-3)
+
+                    const __m256i rhs_mat_0145_11_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_11, 136); //B10(8-11) B11(8-11) B10(8-11) B11(8-11) B14(8-11) B15(8-11) B14(8-11) B15(8-11)
+                    const __m256i rhs_mat_2367_11_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_11, 136); //B12(8-11) B13(8-11) B12(8-11) B13(8-11) B16(8-11) B17(8-11) B16(8-11) B17(8-11)
+
+                    const __m256i rhs_mat_0145_20_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_20, 136); //B20(0-3) B21(0-3) B20(0-3) B21(0-3) B24(0-3) B25(0-3) B24(0-3) B25(0-3)
+                    const __m256i rhs_mat_2367_20_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_20, 136); //B22(0-3) B23(0-3) B22(0-3) B23(0-3) B26(0-3) B27(0-3) B26(0-3) B27(0-3)
+
+                    const __m256i rhs_mat_0145_21_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_21, 136); //B20(8-11) B21(8-11) B20(8-11) B21(8-11) B24(8-11) B25(8-11) B24(8-11) B25(8-11)
+                    const __m256i rhs_mat_2367_21_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_21, 136); //B22(8-11) B23(8-11) B22(8-11) B23(8-11) B26(8-11) B27(8-11) B26(8-11) B27(8-11)
+
+                    const __m256i rhs_mat_0145_30_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_30, 136); //B30(0-3) B31(0-3) B30(0-3) B31(0-3) B34(0-3) B35(0-3) B34(0-3) B35(0-3)
+                    const __m256i rhs_mat_2367_30_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_30, 136); //B32(0-3) B33(0-3) B32(0-3) B33(0-3) B36(0-3) B37(0-3) B36(0-3) B37(0-3)
+
+                    const __m256i rhs_mat_0145_31_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_31, 136); //B30(8-11) B31(8-11) B30(8-11) B31(8-11) B34(8-11) B35(8-11) B34(8-11) B35(8-11
+                    const __m256i rhs_mat_2367_31_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_31, 136); //B32(8-11) B33(8-11) B32(8-11) B33(8-11) B36(8-11) B37(8-11) B36(8-11) B37(8-11)
+
+                    const __m256i rhs_mat_0145_40_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_40, 136); //B40(0-3) B41(0-3) B40(0-3) B41(0-3) B44(0-3) B45(0-3) B44(0-3) B45(0-3)
+                    const __m256i rhs_mat_2367_40_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_40, 136); //B42(0-3) B43(0-3) B42(0-3) B43(0-3) B46(0-3) B47(0-3) B46(0-3) B47(0-3)
+
+                    const __m256i rhs_mat_0145_41_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_41, 136); //B40(8-11) B41(8-11) B40(8-11) B41(8-11) B44(8-11) B45(8-11) B44(8-11) B45(8-11)
+                    const __m256i rhs_mat_2367_41_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_41, 136); //B42(8-11) B43(8-11) B42(8-11) B43(8-11) B46(8-11) B47(8-11) B46(8-11) B47(8-11)
+
+                    const __m256i rhs_mat_0145_50_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_50, 136); //B50(0-3) B51(0-3) B50(0-3) B51(0-3) B54(0-3) B55(0-3) B54(0-3) B55(0-3)
+                    const __m256i rhs_mat_2367_50_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_50, 136); //B52(0-3) B53(0-3) B52(0-3) B53(0-3) B56(0-3) B57(0-3) B56(0-3) B57(0-3)
+
+                    const __m256i rhs_mat_0145_51_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_51, 136); //B50(8-11) B51(8-11) B50(8-11) B51(8-11) B54(8-11) B55(8-11) B54(8-11) B55(8-11)
+                    const __m256i rhs_mat_2367_51_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_51, 136); //B52(8-11) B53(8-11) B52(8-11) B53(8-11) B56(8-11) B57(8-11) B56(8-11) B57(8-11)
+
+                    const __m256i rhs_mat_0145_60_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_60, 136); //B60(0-3) B61(0-3) B60(0-3) B61(0-3) B64(0-3) B65(0-3) B64(0-3) B65(0-3)
+                    const __m256i rhs_mat_2367_60_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_60, 136); //B62(0-3) B63(0-3) B62(0-3) B63(0-3) B66(0-3) B67(0-3) B66(0-3) B67(0-3)
+
+                    const __m256i rhs_mat_0145_61_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_61, 136); //B60(8-11) B61(8-11) B60(8-11) B61(8-11) B64(8-11) B65(8-11) B64(8-11) B65(8-11)
+                    const __m256i rhs_mat_2367_61_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_61, 136); //B62(8-11) B63(8-11) B62(8-11) B63(8-11) B66(8-11) B67(8-11) B66(8-11) B67(8-11)
+
+                    const __m256i rhs_mat_0145_70_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_70, 136); //B70(0-3) B71(0-3) B70(0-3) B71(0-3) B74(0-3) B75(0-3) B74(0-3) B75(0-3)
+                    const __m256i rhs_mat_2367_70_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_70, 136); //B72(0-3) B73(0-3) B72(0-3) B73(0-3) B76(0-3) B77(0-3) B76(0-3) B77(0-3)
+
+                    const __m256i rhs_mat_0145_71_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_71, 136); //B70(8-11) B71(8-11) B70(8-11) B71(8-11) B74(8-11) B75(8-11) B74(8-11) B75(8-11)
+                    const __m256i rhs_mat_2367_71_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_71, 136); //B72(8-11) B73(8-11) B72(8-11) B73(8-11) B76(8-11) B77(8-11) B76(8-11) B77(8-11)
+
+
+                    // Shuffle pattern two - right side input
+                    const __m256i rhs_mat_0145_00_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_00, 221); //B00(4-7) B01(4-7) B00(4-7) B01(4-7) B04(4-7) B05(4-7) B04(4-7) B05(4-7)
+                    const __m256i rhs_mat_2367_00_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_00, 221); //B02(4-7) B03(4-7) B02(4-7) B03(4-7) B06(4-7) B07(4-7) B06(4-7) B07(4-7)
+
+                    const __m256i rhs_mat_0145_01_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_01, 221); //B00(12-15) B01(12-15) B00(12-15) B01(12-15) B04(12-15) B05(12-15) B04(12-15) B05(12-15)
+                    const __m256i rhs_mat_2367_01_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_01, 221); //B02(12-15) B03(12-15) B02(12-15) B03(12-15) B06(12-15) B07(12-15) B06(12-15) B07(12-15)
+
+                    const __m256i rhs_mat_0145_10_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_10, 221); //B10(4-7) B11(4-7) B10(4-7) B11(4-7) B14(4-7) B15(4-7) B14(4-7) B15(4-7)
+                    const __m256i rhs_mat_2367_10_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_10, 221); //B12(4-7) B13(4-7) B12(4-7) B13(4-7) B16(4-7) B17(4-7) B16(4-7) B17(4-7)
+
+                    const __m256i rhs_mat_0145_11_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_11, 221); //B10(12-15) B11(12-15) B10(12-15) B11(12-15) B14(12-15) B15(12-15) B14(12-15) B15(12-15)
+                    const __m256i rhs_mat_2367_11_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_11, 221); //B12(12-15) B13(12-15) B12(12-15) B13(12-15) B16(12-15) B17(12-15) B16(12-15) B17(12-15)
+
+                    const __m256i rhs_mat_0145_20_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_20, 221); //B20(4-7) B21(4-7) B20(4-7) B21(4-7) B24(4-7) B25(4-7) B24(4-7) B25(4-7)
+                    const __m256i rhs_mat_2367_20_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_20, 221); //B22(4-7) B23(4-7) B22(4-7) B23(4-7) B26(4-7) B27(4-7) B26(4-7) B27(4-7)
+
+                    const __m256i rhs_mat_0145_21_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_21, 221); //B20(12-15) B21(12-15) B20(12-15) B21(12-15) B24(12-15) B25(12-15) B24(12-15) B25(12-15)
+                    const __m256i rhs_mat_2367_21_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_21, 221); //B22(12-15) B23(12-15) B22(12-15) B23(12-15) B26(12-15) B27(12-15) B26(12-15) B27(12-15)
+
+                    const __m256i rhs_mat_0145_30_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_30, 221); //B30(4-7) B31(4-7) B30(4-7) B31(4-7) B34(4-7) B35(4-7) B34(4-7) B35(4-7)
+                    const __m256i rhs_mat_2367_30_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_30, 221); //B32(4-7) B33(4-7) B32(4-7) B33(4-7) B36(4-7) B37(4-7) B36(4-7) B37(4-7)
+
+                    const __m256i rhs_mat_0145_31_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_31, 221); //B30(12-15) B31(12-15) B30(12-15) B31(12-15) B34(12-15) B35(12-15) B34(12-15) B35(12-15)
+                    const __m256i rhs_mat_2367_31_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_31, 221); //B32(12-15) B33(12-15) B32(12-15) B33(12-15) B36(12-15) B37(12-15) B36(12-15) B37(12-15)
+
+                    const __m256i rhs_mat_0145_40_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_40, 221); //B40(4-7) B41(4-7) B40(4-7) B41(4-7) B44(4-7) B45(4-7) B44(4-7) B45(4-7)
+                    const __m256i rhs_mat_2367_40_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_40, 221); //B42(4-7) B43(4-7) B42(4-7) B43(4-7) B46(4-7) B47(4-7) B46(4-7) B47(4-7)
+
+                    const __m256i rhs_mat_0145_41_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_41, 221); //B40(12-15) B41(12-15) B40(12-15) B41(12-15) B44(12-15) B45(12-15) B44(12-15) B45(12-15)
+                    const __m256i rhs_mat_2367_41_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_41, 221); //B42(12-15) B43(12-15) B42(12-15) B43(12-15) B46(12-15) B47(12-15) B46(12-15) B47(12-15)
+
+                    const __m256i rhs_mat_0145_50_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_50, 221); //B50(4-7) B51(4-7) B50(4-7) B51(4-7) B54(4-7) B55(4-7) B54(4-7) B55(4-7)
+                    const __m256i rhs_mat_2367_50_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_50, 221); //B52(4-7) B53(4-7) B52(4-7) B53(4-7) B56(4-7) B57(4-7) B56(4-7) B57(4-7)
+
+                    const __m256i rhs_mat_0145_51_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_51, 221); //B50(12-15) B51(12-15) B50(12-15) B51(12-15) B54(12-15) B55(12-15) B54(12-15) B55(12-15)
+                    const __m256i rhs_mat_2367_51_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_51, 221); //B52(12-15) B53(12-15) B52(12-15) B53(12-15) B56(12-15) B57(12-15) B56(12-15) B57(12-15)
+
+                    const __m256i rhs_mat_0145_60_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_60, 221); //B60(4-7) B61(4-7) B60(4-7) B61(4-7) B64(4-7) B65(4-7) B64(4-7) B65(4-7)
+                    const __m256i rhs_mat_2367_60_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_60, 221); //B62(4-7) B63(4-7) B62(4-7) B63(4-7) B66(4-7) B67(4-7) B66(4-7) B67(4-7)
+
+                    const __m256i rhs_mat_0145_61_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_61, 221); //B60(12-15) B61(12-15) B60(12-15) B61(12-15) B64(12-15) B65(12-15) B64(12-15) B65(12-15)
+                    const __m256i rhs_mat_2367_61_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_61, 221); //B62(12-15) B63(12-15) B62(12-15) B63(12-15) B66(12-15) B67(12-15) B66(12-15) B67(12-15)
+
+                    const __m256i rhs_mat_0145_70_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_70, 221); //B70(4-7) B71(4-7) B70(4-7) B71(4-7) B74(4-7) B75(4-7) B74(4-7) B75(4-7)
+                    const __m256i rhs_mat_2367_70_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_70, 221); //B72(4-7) B73(4-7) B72(4-7) B73(4-7) B76(4-7) B77(4-7) B76(4-7) B77(4-7)
+
+                    const __m256i rhs_mat_0145_71_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_71, 221); //B70(12-15) B71(12-15) B70(12-15) B71(12-15) B74(12-15) B75(12-15) B74(12-15) B75(12-15)
+                    const __m256i rhs_mat_2367_71_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_71, 221); //B72(12-15) B73(12-15) B72(12-15) B73(12-15) B76(12-15) B77(12-15) B76(12-15) B77(12-15)
+
+                    //Scales and Mins of corresponding sub blocks from different Q2_K structures are stored together
+                    //s00 m00  s01 m01   s10 m10  s11 m11  s20 m20  s21 m21   s30 m30  s31 m31  s40 m40  s41 m41   s50 m50  s51 m51  s60 m60  s61 m61   s70 m70  s71 m71
+
+                    // Combine mins and scales for sub-blocks: 0-1, 2-3, 4-5, 6-7 in the sb loop
+                    const __m128i mins_and_scales_01 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + sb * 64));
+                    const __m128i mins_and_scales_23 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 16 + sb * 64));
+                    const __m128i mins_and_scales_45 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 32 + sb * 64));
+                    const __m128i mins_and_scales_67 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 48 + sb * 64));
+
+                    // Extract scales which is lower half from mins_and_scales
+                    const __m128i scales_01 = _mm_and_si128(mins_and_scales_01, m4b_sse);
+                    const __m128i scales_23 = _mm_and_si128(mins_and_scales_23, m4b_sse);
+                    const __m128i scales_45 = _mm_and_si128(mins_and_scales_45, m4b_sse);
+                    const __m128i scales_67 = _mm_and_si128(mins_and_scales_67, m4b_sse);
+
+                    // Extract mins which is upper half from mins_and_scales
+                    const __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_01, 4), m4b_sse));
+                    const __m256i mins_23 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_23, 4), m4b_sse));
+                    const __m256i mins_45 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_45, 4), m4b_sse));
+                    const __m256i mins_67 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_67, 4), m4b_sse));
+
+                    const __m256i scales_0 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_01, scalesmask1_sse));
+                    const __m256i scales_1 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_01, scalesmask2_sse));
+
+                    const __m256i scales_2 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_23, scalesmask1_sse));
+                    const __m256i scales_3 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_23, scalesmask2_sse));
+
+                    const __m256i scales_4 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_45, scalesmask1_sse));
+                    const __m256i scales_5 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_45, scalesmask2_sse));
+
+                    const __m256i scales_6 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_67, scalesmask1_sse));
+                    const __m256i scales_7 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_67, scalesmask2_sse));
+
+                    const __m256i scale_0145_0 = _mm256_shuffle_epi32(scales_0, 68);
+                    const __m256i scale_2367_0 = _mm256_shuffle_epi32(scales_0, 238);
+
+                    const __m256i scale_0145_1 = _mm256_shuffle_epi32(scales_1, 68);
+                    const __m256i scale_2367_1 = _mm256_shuffle_epi32(scales_1, 238);
+
+                    const __m256i scale_0145_2 = _mm256_shuffle_epi32(scales_2, 68);
+                    const __m256i scale_2367_2 = _mm256_shuffle_epi32(scales_2, 238);
+
+                    const __m256i scale_0145_3 = _mm256_shuffle_epi32(scales_3, 68);
+                    const __m256i scale_2367_3 = _mm256_shuffle_epi32(scales_3, 238);
+
+                    const __m256i scale_0145_4 = _mm256_shuffle_epi32(scales_4, 68);
+                    const __m256i scale_2367_4 = _mm256_shuffle_epi32(scales_4, 238);
+
+                    const __m256i scale_0145_5 = _mm256_shuffle_epi32(scales_5, 68);
+                    const __m256i scale_2367_5 = _mm256_shuffle_epi32(scales_5, 238);
+
+                    const __m256i scale_0145_6 = _mm256_shuffle_epi32(scales_6, 68);
+                    const __m256i scale_2367_6 = _mm256_shuffle_epi32(scales_6, 238);
+
+                    const __m256i scale_0145_7 = _mm256_shuffle_epi32(scales_7, 68);
+                    const __m256i scale_2367_7 = _mm256_shuffle_epi32(scales_7, 238);
+
+
+                    for (int rp = 0; rp < 4; rp++) {
+
+                        // Load the four block_q8_k quantized values interleaved with each other in chunks of eight bytes - A0,A1,A2,A3
+                        // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
+                        __m256i lhs_mat_0123_00 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 512 * sb)));
+                        __m256i lhs_mat_01_00 = _mm256_permute2f128_si256(lhs_mat_0123_00, lhs_mat_0123_00, 0);
+                        __m256i lhs_mat_23_00 = _mm256_permute2f128_si256(lhs_mat_0123_00, lhs_mat_0123_00, 17);
+                        __m256i lhs_mat_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 32 + 512 * sb)));
+                        __m256i lhs_mat_01_01 = _mm256_permute2f128_si256(lhs_mat_0123_01, lhs_mat_0123_01, 0);
+                        __m256i lhs_mat_23_01 = _mm256_permute2f128_si256(lhs_mat_0123_01, lhs_mat_0123_01, 17);
+                        __m256i lhs_mat_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 64 + 512 * sb)));
+                        __m256i lhs_mat_01_10 = _mm256_permute2f128_si256(lhs_mat_0123_10, lhs_mat_0123_10, 0);
+                        __m256i lhs_mat_23_10 = _mm256_permute2f128_si256(lhs_mat_0123_10, lhs_mat_0123_10, 17);
+                        __m256i lhs_mat_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 96 + 512 * sb)));
+                        __m256i lhs_mat_01_11 = _mm256_permute2f128_si256(lhs_mat_0123_11, lhs_mat_0123_11, 0);
+                        __m256i lhs_mat_23_11 = _mm256_permute2f128_si256(lhs_mat_0123_11, lhs_mat_0123_11, 17);
+                        __m256i lhs_mat_0123_20 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 128 + 512 * sb)));
+                        __m256i lhs_mat_01_20 = _mm256_permute2f128_si256(lhs_mat_0123_20, lhs_mat_0123_20, 0);
+                        __m256i lhs_mat_23_20 = _mm256_permute2f128_si256(lhs_mat_0123_20, lhs_mat_0123_20, 17);
+                        __m256i lhs_mat_0123_21 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 160 + 512 * sb)));
+                        __m256i lhs_mat_01_21 = _mm256_permute2f128_si256(lhs_mat_0123_21, lhs_mat_0123_21, 0);
+                        __m256i lhs_mat_23_21 = _mm256_permute2f128_si256(lhs_mat_0123_21, lhs_mat_0123_21, 17);
+                        __m256i lhs_mat_0123_30 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 192 + 512 * sb)));
+                        __m256i lhs_mat_01_30 = _mm256_permute2f128_si256(lhs_mat_0123_30, lhs_mat_0123_30, 0);
+                        __m256i lhs_mat_23_30 = _mm256_permute2f128_si256(lhs_mat_0123_30, lhs_mat_0123_30, 17);
+                        __m256i lhs_mat_0123_31 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 224 + 512 * sb)));
+                        __m256i lhs_mat_01_31 = _mm256_permute2f128_si256(lhs_mat_0123_31, lhs_mat_0123_31, 0);
+                        __m256i lhs_mat_23_31 = _mm256_permute2f128_si256(lhs_mat_0123_31, lhs_mat_0123_31, 17);
+
+                        __m256i lhs_mat_0123_40 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 256 + 512 * sb)));
+                        __m256i lhs_mat_01_40 = _mm256_permute2f128_si256(lhs_mat_0123_40, lhs_mat_0123_40, 0);
+                        __m256i lhs_mat_23_40 = _mm256_permute2f128_si256(lhs_mat_0123_40, lhs_mat_0123_40, 17);
+                        __m256i lhs_mat_0123_41 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 288 + 512 * sb)));
+                        __m256i lhs_mat_01_41 = _mm256_permute2f128_si256(lhs_mat_0123_41, lhs_mat_0123_41, 0);
+                        __m256i lhs_mat_23_41 = _mm256_permute2f128_si256(lhs_mat_0123_41, lhs_mat_0123_41, 17);
+                        __m256i lhs_mat_0123_50 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 320 + 512 * sb)));
+                        __m256i lhs_mat_01_50 = _mm256_permute2f128_si256(lhs_mat_0123_50, lhs_mat_0123_50, 0);
+                        __m256i lhs_mat_23_50 = _mm256_permute2f128_si256(lhs_mat_0123_50, lhs_mat_0123_50, 17);
+                        __m256i lhs_mat_0123_51 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 352 + 512 * sb)));
+                        __m256i lhs_mat_01_51 = _mm256_permute2f128_si256(lhs_mat_0123_51, lhs_mat_0123_51, 0);
+                        __m256i lhs_mat_23_51 = _mm256_permute2f128_si256(lhs_mat_0123_51, lhs_mat_0123_51, 17);
+                        __m256i lhs_mat_0123_60 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 384 + 512 * sb)));
+                        __m256i lhs_mat_01_60 = _mm256_permute2f128_si256(lhs_mat_0123_60, lhs_mat_0123_60, 0);
+                        __m256i lhs_mat_23_60 = _mm256_permute2f128_si256(lhs_mat_0123_60, lhs_mat_0123_60, 17);
+                        __m256i lhs_mat_0123_61 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 416 + 512 * sb)));
+                        __m256i lhs_mat_01_61 = _mm256_permute2f128_si256(lhs_mat_0123_61, lhs_mat_0123_61, 0);
+                        __m256i lhs_mat_23_61 = _mm256_permute2f128_si256(lhs_mat_0123_61, lhs_mat_0123_61, 17);
+                        __m256i lhs_mat_0123_70 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 448 + 512 * sb)));
+                        __m256i lhs_mat_01_70 = _mm256_permute2f128_si256(lhs_mat_0123_70, lhs_mat_0123_70, 0);
+                        __m256i lhs_mat_23_70 = _mm256_permute2f128_si256(lhs_mat_0123_70, lhs_mat_0123_70, 17);
+                        __m256i lhs_mat_0123_71 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 480 + 512 * sb)));
+                        __m256i lhs_mat_01_71 = _mm256_permute2f128_si256(lhs_mat_0123_71, lhs_mat_0123_71, 0);
+                        __m256i lhs_mat_23_71 = _mm256_permute2f128_si256(lhs_mat_0123_71, lhs_mat_0123_71, 17);
+
+                        // Bsums are loaded for the different Q8_K blocks
+                        __m128i lhs_raw_bsums_01_0123 = _mm_loadu_si128((const __m128i *)((a_ptrs[rp][b].bsums + 32 * sb)));
+                        __m128i lhs_raw_bsums_23_0123 = _mm_loadu_si128((const __m128i *)(a_ptrs[rp][b].bsums + 8 + 32 * sb));
+                        __m128i lhs_raw_bsums_01_4567 = _mm_loadu_si128((const __m128i *)((a_ptrs[rp][b].bsums + 16 + 32 * sb)));
+                        __m128i lhs_raw_bsums_23_4567 = _mm_loadu_si128((const __m128i *)(a_ptrs[rp][b].bsums + 24 + 32 * sb));
+
+                        // Shuffle pattern one - left side input
+                        const __m256i lhs_mat_01_00_sp1 = _mm256_shuffle_epi32(lhs_mat_01_00, 160); //A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3)
+                        const __m256i lhs_mat_23_00_sp1 = _mm256_shuffle_epi32(lhs_mat_23_00, 160); //A02(0-3) A03(0-3) A02(0-3) A03(0-3) A02(0-3) A03(0-3) A02(0-3) A03(0-3)
+
+                        const __m256i lhs_mat_01_01_sp1 = _mm256_shuffle_epi32(lhs_mat_01_01, 160); //A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11)
+                        const __m256i lhs_mat_23_01_sp1 = _mm256_shuffle_epi32(lhs_mat_23_01, 160); //A02(8-11) A03(8-11) A02(8-11) A03(8-11) A02(8-11) A03(8-11) A02(8-11) A03(8-11)
+
+                        const __m256i lhs_mat_01_10_sp1 = _mm256_shuffle_epi32(lhs_mat_01_10, 160); //A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3)
+                        const __m256i lhs_mat_23_10_sp1 = _mm256_shuffle_epi32(lhs_mat_23_10, 160); //A12(0-3) A13(0-3) A12(0-3) A13(0-3) A12(0-3) A13(0-3) A12(0-3) A13(0-3)
+
+                        const __m256i lhs_mat_01_11_sp1 = _mm256_shuffle_epi32(lhs_mat_01_11, 160); //A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11)
+                        const __m256i lhs_mat_23_11_sp1 = _mm256_shuffle_epi32(lhs_mat_23_11, 160); //A12(8-11) A13(8-11) A12(8-11) A13(8-11) A12(8-11) A13(8-11) A12(8-11) A13(8-11)
+
+                        const __m256i lhs_mat_01_20_sp1 = _mm256_shuffle_epi32(lhs_mat_01_20, 160); //A20(0-3) A20(0-3) A21(0-3) A21(0-3) A20(0-3) A20(0-3) A21(0-3) A21(0-3)
+                        const __m256i lhs_mat_23_20_sp1 = _mm256_shuffle_epi32(lhs_mat_23_20, 160); //A22(0-3) A23(0-3) A22(0-3) A23(0-3) A22(0-3) A23(0-3) A22(0-3) A23(0-3)
+
+                        const __m256i lhs_mat_01_21_sp1 = _mm256_shuffle_epi32(lhs_mat_01_21, 160); //A20(8-11) A20(8-11) A21(8-11) A21(8-11) A20(8-11) A20(8-11) A21(8-11) A21(8-11)
+                        const __m256i lhs_mat_23_21_sp1 = _mm256_shuffle_epi32(lhs_mat_23_21, 160); //A22(8-11) A23(8-11) A22(8-11) A23(8-11) A22(8-11) A23(8-11) A22(8-11) A23(8-11)
+
+                        const __m256i lhs_mat_01_30_sp1 = _mm256_shuffle_epi32(lhs_mat_01_30, 160); //A30(0-3) A30(0-3) A31(0-3) A31(0-3) A30(0-3) A30(0-3) A31(0-3) A31(0-3)
+                        const __m256i lhs_mat_23_30_sp1 = _mm256_shuffle_epi32(lhs_mat_23_30, 160); //A32(0-3) A33(0-3) A32(0-3) A33(0-3) A32(0-3) A33(0-3) A32(0-3) A33(0-3)
+
+                        const __m256i lhs_mat_01_31_sp1 = _mm256_shuffle_epi32(lhs_mat_01_31, 160); //A30(8-11) A30(8-11) A31(8-11) A31(8-11) A30(8-11) A30(8-11) A31(8-11) A31(8-11)
+                        const __m256i lhs_mat_23_31_sp1 = _mm256_shuffle_epi32(lhs_mat_23_31, 160); //A32(8-11) A33(8-11) A32(8-11) A33(8-11) A32(8-11) A33(8-11) A32(8-11) A33(8-11)
+
+                        const __m256i lhs_mat_01_40_sp1 = _mm256_shuffle_epi32(lhs_mat_01_40, 160); //A40(0-3) A40(0-3) A41(0-3) A41(0-3) A40(0-3) A40(0-3) A41(0-3) A41(0-3)
+                        const __m256i lhs_mat_23_40_sp1 = _mm256_shuffle_epi32(lhs_mat_23_40, 160); //A42(0-3) A43(0-3) A42(0-3) A43(0-3) A42(0-3) A43(0-3) A42(0-3) A43(0-3)
+
+                        const __m256i lhs_mat_01_41_sp1 = _mm256_shuffle_epi32(lhs_mat_01_41, 160); //A40(8-11) A40(8-11) A41(8-11) A41(8-11) A40(8-11) A40(8-11) A41(8-11) A41(8-11)
+                        const __m256i lhs_mat_23_41_sp1 = _mm256_shuffle_epi32(lhs_mat_23_41, 160); //A42(8-11) A43(8-11) A42(8-11) A43(8-11) A42(8-11) A43(8-11) A42(8-11) A43(8-11)
+
+                        const __m256i lhs_mat_01_50_sp1 = _mm256_shuffle_epi32(lhs_mat_01_50, 160); //A50(0-3) A50(0-3) A51(0-3) A51(0-3) A50(0-3) A50(0-3) A51(0-3) A51(0-3)
+                        const __m256i lhs_mat_23_50_sp1 = _mm256_shuffle_epi32(lhs_mat_23_50, 160); //A52(0-3) A53(0-3) A52(0-3) A53(0-3) A52(0-3) A53(0-3) A52(0-3) A53(0-3)
+
+                        const __m256i lhs_mat_01_51_sp1 = _mm256_shuffle_epi32(lhs_mat_01_51, 160); //A50(8-11) A50(8-11) A51(8-11) A51(8-11) A50(8-11) A50(8-11) A51(8-11) A51(8-11)
+                        const __m256i lhs_mat_23_51_sp1 = _mm256_shuffle_epi32(lhs_mat_23_51, 160); //A52(8-11) A53(8-11) A52(8-11) A53(8-11) A52(8-11) A53(8-11) A52(8-11) A53(8-11)
+
+                        const __m256i lhs_mat_01_60_sp1 = _mm256_shuffle_epi32(lhs_mat_01_60, 160); //A60(0-3) A60(0-3) A61(0-3) A61(0-3) A60(0-3) A60(0-3) A61(0-3) A61(0-3)
+                        const __m256i lhs_mat_23_60_sp1 = _mm256_shuffle_epi32(lhs_mat_23_60, 160); //A62(0-3) A63(0-3) A62(0-3) A63(0-3) A62(0-3) A63(0-3) A62(0-3) A63(0-3)
+
+                        const __m256i lhs_mat_01_61_sp1 = _mm256_shuffle_epi32(lhs_mat_01_61, 160); //A60(8-11) A60(8-11) A61(8-11) A61(8-11) A60(8-11) A60(8-11) A61(8-11) A61(8-11)
+                        const __m256i lhs_mat_23_61_sp1 = _mm256_shuffle_epi32(lhs_mat_23_61, 160); //A62(8-11) A63(8-11) A62(8-11) A63(8-11) A62(8-11) A63(8-11) A62(8-11) A63(8-11)
+
+                        const __m256i lhs_mat_01_70_sp1 = _mm256_shuffle_epi32(lhs_mat_01_70, 160); //A70(0-3) A70(0-3) A71(0-3) A71(0-3) A70(0-3) A70(0-3) A71(0-3) A71(0-3)
+                        const __m256i lhs_mat_23_70_sp1 = _mm256_shuffle_epi32(lhs_mat_23_70, 160); //A72(0-3) A73(0-3) A72(0-3) A73(0-3) A72(0-3) A73(0-3) A72(0-3) A73(0-3)
+
+                        const __m256i lhs_mat_01_71_sp1 = _mm256_shuffle_epi32(lhs_mat_01_71, 160); //A70(8-11) A70(8-11) A71(8-11) A71(8-11) A70(8-11) A70(8-11) A71(8-11) A71(8-11)
+                        const __m256i lhs_mat_23_71_sp1 = _mm256_shuffle_epi32(lhs_mat_23_71, 160); //A72(8-11) A73(8-11) A72(8-11) A73(8-11) A72(8-11) A73(8-11) A72(8-11) A73(8-11)
+
+                        // Shuffle pattern two- left side input
+                        const __m256i lhs_mat_01_00_sp2 = _mm256_shuffle_epi32(lhs_mat_01_00, 245); //A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7)
+                        const __m256i lhs_mat_23_00_sp2 = _mm256_shuffle_epi32(lhs_mat_23_00, 245); //A02(4-7) A03(4-7) A02(4-7) A03(4-7) A02(4-7) A03(4-7) A02(4-7) A03(4-7)
+
+                        const __m256i lhs_mat_01_01_sp2 = _mm256_shuffle_epi32(lhs_mat_01_01, 245); //A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15)
+                        const __m256i lhs_mat_23_01_sp2 = _mm256_shuffle_epi32(lhs_mat_23_01, 245); //A02(12-15) A03(12-15) A02(12-15) A03(12-15) A02(12-15) A03(12-15) A02(12-15) A03(12-15)
+
+                        const __m256i lhs_mat_01_10_sp2 = _mm256_shuffle_epi32(lhs_mat_01_10, 245); //A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7)
+                        const __m256i lhs_mat_23_10_sp2 = _mm256_shuffle_epi32(lhs_mat_23_10, 245); //A12(4-7) A13(4-7) A12(4-7) A13(4-7) A12(4-7) A13(4-7) A12(4-7) A13(4-7)
+
+                        const __m256i lhs_mat_01_11_sp2 = _mm256_shuffle_epi32(lhs_mat_01_11, 245); //A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15)
+                        const __m256i lhs_mat_23_11_sp2 = _mm256_shuffle_epi32(lhs_mat_23_11, 245); //A12(12-15) A13(12-15) A12(12-15) A13(12-15) A12(12-15) A13(12-15) A12(12-15) A13(12-15)
+
+                        const __m256i lhs_mat_01_20_sp2 = _mm256_shuffle_epi32(lhs_mat_01_20, 245); //A20(4-7) A20(4-7) A21(4-7) A21(4-7) A20(4-7) A20(4-7) A21(4-7) A21(4-7)
+                        const __m256i lhs_mat_23_20_sp2 = _mm256_shuffle_epi32(lhs_mat_23_20, 245); //A22(4-7) A23(4-7) A22(4-7) A23(4-7) A22(4-7) A23(4-7) A22(4-7) A23(4-7)
+
+                        const __m256i lhs_mat_01_21_sp2 = _mm256_shuffle_epi32(lhs_mat_01_21, 245); //A20(12-15) A20(12-15) A21(12-15) A21(12-15) A20(12-15) A20(12-15) A21(12-15) A21(12-15)
+                        const __m256i lhs_mat_23_21_sp2 = _mm256_shuffle_epi32(lhs_mat_23_21, 245); //A22(12-15) A23(12-15) A22(12-15) A23(12-15) A22(12-15) A23(12-15) A22(12-15) A23(12-15)
+
+                        const __m256i lhs_mat_01_30_sp2 = _mm256_shuffle_epi32(lhs_mat_01_30, 245); //A30(4-7) A30(4-7) A31(4-7) A31(4-7) A30(4-7) A30(4-7) A31(4-7) A31(4-7)
+                        const __m256i lhs_mat_23_30_sp2 = _mm256_shuffle_epi32(lhs_mat_23_30, 245); //A32(4-7) A33(4-7) A32(4-7) A33(4-7) A32(4-7) A33(4-7) A32(4-7) A33(4-7)
+
+                        const __m256i lhs_mat_01_31_sp2 = _mm256_shuffle_epi32(lhs_mat_01_31, 245); //A30(12-15) A30(12-15) A31(12-15) A31(12-15) A30(12-15) A30(12-15) A31(12-15) A31(12-15)
+                        const __m256i lhs_mat_23_31_sp2 = _mm256_shuffle_epi32(lhs_mat_23_31, 245); //A32(12-15) A33(12-15) A32(12-15) A33(12-15) A32(12-15) A33(12-15) A32(12-15) A33(12-15)
+
+                        const __m256i lhs_mat_01_40_sp2 = _mm256_shuffle_epi32(lhs_mat_01_40, 245); //A40(4-7) A40(4-7) A41(4-7) A41(4-7) A40(4-7) A40(4-7) A41(4-7) A41(4-7)
+                        const __m256i lhs_mat_23_40_sp2 = _mm256_shuffle_epi32(lhs_mat_23_40, 245); //A42(4-7) A43(4-7) A42(4-7) A43(4-7) A42(4-7) A43(4-7) A42(4-7) A43(4-7)
+
+                        const __m256i lhs_mat_01_41_sp2 = _mm256_shuffle_epi32(lhs_mat_01_41, 245); //A40(12-15) A40(12-15) A41(12-15) A41(12-15) A40(12-15) A40(12-15) A41(12-15) A41(12-15)
+                        const __m256i lhs_mat_23_41_sp2 = _mm256_shuffle_epi32(lhs_mat_23_41, 245); //A42(12-15) A43(12-15) A42(12-15) A43(12-15) A42(12-15) A43(12-15) A42(12-15) A43(12-15)
+
+                        const __m256i lhs_mat_01_50_sp2 = _mm256_shuffle_epi32(lhs_mat_01_50, 245); //A50(4-7) A50(4-7) A51(4-7) A51(4-7) A50(4-7) A50(4-7) A51(4-7) A51(4-7)
+                        const __m256i lhs_mat_23_50_sp2 = _mm256_shuffle_epi32(lhs_mat_23_50, 245); //A52(4-7) A53(4-7) A52(4-7) A53(4-7) A52(4-7) A53(4-7) A52(4-7) A53(4-7)
+
+                        const __m256i lhs_mat_01_51_sp2 = _mm256_shuffle_epi32(lhs_mat_01_51, 245); //A50(12-15) A50(12-15) A51(12-15) A51(12-15) A50(12-15) A50(12-15) A51(12-15) A51(12-15)
+                        const __m256i lhs_mat_23_51_sp2 = _mm256_shuffle_epi32(lhs_mat_23_51, 245); //A52(12-15) A53(12-15) A52(12-15) A53(12-15) A52(12-15) A53(12-15) A52(12-15) A53(12-15)
+
+                        const __m256i lhs_mat_01_60_sp2 = _mm256_shuffle_epi32(lhs_mat_01_60, 245); //A60(4-7) A60(4-7) A61(4-7) A61(4-7) A60(4-7) A60(4-7) A61(4-7) A61(4-7)
+                        const __m256i lhs_mat_23_60_sp2 = _mm256_shuffle_epi32(lhs_mat_23_60, 245); //A62(4-7) A63(4-7) A62(4-7) A63(4-7) A62(4-7) A63(4-7) A62(4-7) A63(4-7)
+
+                        const __m256i lhs_mat_01_61_sp2 = _mm256_shuffle_epi32(lhs_mat_01_61, 245); //A60(12-15) A60(12-15) A61(12-15) A61(12-15) A60(12-15) A60(12-15) A61(12-15) A61(12-15)
+                        const __m256i lhs_mat_23_61_sp2 = _mm256_shuffle_epi32(lhs_mat_23_61, 245); //A62(12-15) A63(12-15) A62(12-15) A63(12-15) A62(12-15) A63(12-15) A62(12-15) A63(12-15)
+
+                        const __m256i lhs_mat_01_70_sp2 = _mm256_shuffle_epi32(lhs_mat_01_70, 245); //A70(4-7) A70(4-7) A71(4-7) A71(4-7) A70(4-7) A70(4-7) A71(4-7) A71(4-7)
+                        const __m256i lhs_mat_23_70_sp2 = _mm256_shuffle_epi32(lhs_mat_23_70, 245); //A72(4-7) A73(4-7) A72(4-7) A73(4-7) A72(4-7) A73(4-7) A72(4-7) A73(4-7)
+
+                        const __m256i lhs_mat_01_71_sp2 = _mm256_shuffle_epi32(lhs_mat_01_71, 245); //A70(12-15) A70(12-15) A71(12-15) A71(12-15) A70(12-15) A70(12-15) A71(12-15) A71(12-15)
+                        const __m256i lhs_mat_23_71_sp2 = _mm256_shuffle_epi32(lhs_mat_23_71, 245); //A72(12-15) A73(12-15) A72(12-15) A73(12-15) A72(12-15) A73(12-15) A72(12-15) A73(12-15)
+
+                        // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+                        __m256i iacc_mat_00_0_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_00_sp1, lhs_mat_01_00_sp1),_mm256_maddubs_epi16(rhs_mat_0145_01_sp1, lhs_mat_01_01_sp1));
+                        __m256i iacc_mat_01_0_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_00_sp1, lhs_mat_01_00_sp1),_mm256_maddubs_epi16(rhs_mat_2367_01_sp1, lhs_mat_01_01_sp1));
+
+                        __m256i iacc_mat_10_0_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_00_sp1, lhs_mat_23_00_sp1),_mm256_maddubs_epi16(rhs_mat_0145_01_sp1, lhs_mat_23_01_sp1));
+                        __m256i iacc_mat_11_0_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_00_sp1, lhs_mat_23_00_sp1),_mm256_maddubs_epi16(rhs_mat_2367_01_sp1, lhs_mat_23_01_sp1));
+
+                        __m256i iacc_mat_00_1_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_10_sp1, lhs_mat_01_10_sp1),_mm256_maddubs_epi16(rhs_mat_0145_11_sp1, lhs_mat_01_11_sp1));
+                        __m256i iacc_mat_01_1_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_10_sp1, lhs_mat_01_10_sp1),_mm256_maddubs_epi16(rhs_mat_2367_11_sp1, lhs_mat_01_11_sp1));
+
+                        __m256i iacc_mat_10_1_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_10_sp1, lhs_mat_23_10_sp1),_mm256_maddubs_epi16(rhs_mat_0145_11_sp1, lhs_mat_23_11_sp1));
+                        __m256i iacc_mat_11_1_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_10_sp1, lhs_mat_23_10_sp1),_mm256_maddubs_epi16(rhs_mat_2367_11_sp1, lhs_mat_23_11_sp1));
+
+                        __m256i iacc_mat_00_2_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_20_sp1, lhs_mat_01_20_sp1),_mm256_maddubs_epi16(rhs_mat_0145_21_sp1, lhs_mat_01_21_sp1));
+                        __m256i iacc_mat_01_2_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_20_sp1, lhs_mat_01_20_sp1),_mm256_maddubs_epi16(rhs_mat_2367_21_sp1, lhs_mat_01_21_sp1));
+
+                        __m256i iacc_mat_10_2_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_20_sp1, lhs_mat_23_20_sp1),_mm256_maddubs_epi16(rhs_mat_0145_21_sp1, lhs_mat_23_21_sp1));
+                        __m256i iacc_mat_11_2_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_20_sp1, lhs_mat_23_20_sp1),_mm256_maddubs_epi16(rhs_mat_2367_21_sp1, lhs_mat_23_21_sp1));
+
+                        __m256i iacc_mat_00_3_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_30_sp1, lhs_mat_01_30_sp1),_mm256_maddubs_epi16(rhs_mat_0145_31_sp1, lhs_mat_01_31_sp1));
+                        __m256i iacc_mat_01_3_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_30_sp1, lhs_mat_01_30_sp1),_mm256_maddubs_epi16(rhs_mat_2367_31_sp1, lhs_mat_01_31_sp1));
+
+                        __m256i iacc_mat_10_3_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_30_sp1, lhs_mat_23_30_sp1),_mm256_maddubs_epi16(rhs_mat_0145_31_sp1, lhs_mat_23_31_sp1));
+                        __m256i iacc_mat_11_3_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_30_sp1, lhs_mat_23_30_sp1),_mm256_maddubs_epi16(rhs_mat_2367_31_sp1, lhs_mat_23_31_sp1));
+
+                        __m256i iacc_mat_00_4_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_40_sp1, lhs_mat_01_40_sp1),_mm256_maddubs_epi16(rhs_mat_0145_41_sp1, lhs_mat_01_41_sp1));
+                        __m256i iacc_mat_01_4_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_40_sp1, lhs_mat_01_40_sp1),_mm256_maddubs_epi16(rhs_mat_2367_41_sp1, lhs_mat_01_41_sp1));
+
+                        __m256i iacc_mat_10_4_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_40_sp1, lhs_mat_23_40_sp1),_mm256_maddubs_epi16(rhs_mat_0145_41_sp1, lhs_mat_23_41_sp1));
+                        __m256i iacc_mat_11_4_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_40_sp1, lhs_mat_23_40_sp1),_mm256_maddubs_epi16(rhs_mat_2367_41_sp1, lhs_mat_23_41_sp1));
+
+                        __m256i iacc_mat_00_5_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_50_sp1, lhs_mat_01_50_sp1),_mm256_maddubs_epi16(rhs_mat_0145_51_sp1, lhs_mat_01_51_sp1));
+                        __m256i iacc_mat_01_5_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_50_sp1, lhs_mat_01_50_sp1),_mm256_maddubs_epi16(rhs_mat_2367_51_sp1, lhs_mat_01_51_sp1));
+
+                        __m256i iacc_mat_10_5_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_50_sp1, lhs_mat_23_50_sp1),_mm256_maddubs_epi16(rhs_mat_0145_51_sp1, lhs_mat_23_51_sp1));
+                        __m256i iacc_mat_11_5_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_50_sp1, lhs_mat_23_50_sp1),_mm256_maddubs_epi16(rhs_mat_2367_51_sp1, lhs_mat_23_51_sp1));
+
+                        __m256i iacc_mat_00_6_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_60_sp1, lhs_mat_01_60_sp1),_mm256_maddubs_epi16(rhs_mat_0145_61_sp1, lhs_mat_01_61_sp1));
+                        __m256i iacc_mat_01_6_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_60_sp1, lhs_mat_01_60_sp1),_mm256_maddubs_epi16(rhs_mat_2367_61_sp1, lhs_mat_01_61_sp1));
+
+                        __m256i iacc_mat_10_6_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_60_sp1, lhs_mat_23_60_sp1),_mm256_maddubs_epi16(rhs_mat_0145_61_sp1, lhs_mat_23_61_sp1));
+                        __m256i iacc_mat_11_6_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_60_sp1, lhs_mat_23_60_sp1),_mm256_maddubs_epi16(rhs_mat_2367_61_sp1, lhs_mat_23_61_sp1));
+
+                        __m256i iacc_mat_00_7_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_70_sp1, lhs_mat_01_70_sp1),_mm256_maddubs_epi16(rhs_mat_0145_71_sp1, lhs_mat_01_71_sp1));
+                        __m256i iacc_mat_01_7_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_70_sp1, lhs_mat_01_70_sp1),_mm256_maddubs_epi16(rhs_mat_2367_71_sp1, lhs_mat_01_71_sp1));
+
+                        __m256i iacc_mat_10_7_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_70_sp1, lhs_mat_23_70_sp1),_mm256_maddubs_epi16(rhs_mat_0145_71_sp1, lhs_mat_23_71_sp1));
+                        __m256i iacc_mat_11_7_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_70_sp1, lhs_mat_23_70_sp1),_mm256_maddubs_epi16(rhs_mat_2367_71_sp1, lhs_mat_23_71_sp1));
+
+
+                        __m256i iacc_mat_00_0_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_00_sp2, lhs_mat_01_00_sp2),_mm256_maddubs_epi16(rhs_mat_0145_01_sp2, lhs_mat_01_01_sp2));
+                        __m256i iacc_mat_01_0_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_00_sp2, lhs_mat_01_00_sp2),_mm256_maddubs_epi16(rhs_mat_2367_01_sp2, lhs_mat_01_01_sp2));
+
+                        __m256i iacc_mat_10_0_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_00_sp2, lhs_mat_23_00_sp2),_mm256_maddubs_epi16(rhs_mat_0145_01_sp2, lhs_mat_23_01_sp2));
+                        __m256i iacc_mat_11_0_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_00_sp2, lhs_mat_23_00_sp2),_mm256_maddubs_epi16(rhs_mat_2367_01_sp2, lhs_mat_23_01_sp2));
+
+                        __m256i iacc_mat_00_1_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_10_sp2, lhs_mat_01_10_sp2),_mm256_maddubs_epi16(rhs_mat_0145_11_sp2, lhs_mat_01_11_sp2));
+                        __m256i iacc_mat_01_1_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_10_sp2, lhs_mat_01_10_sp2),_mm256_maddubs_epi16(rhs_mat_2367_11_sp2, lhs_mat_01_11_sp2));
+
+                        __m256i iacc_mat_10_1_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_10_sp2, lhs_mat_23_10_sp2),_mm256_maddubs_epi16(rhs_mat_0145_11_sp2, lhs_mat_23_11_sp2));
+                        __m256i iacc_mat_11_1_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_10_sp2, lhs_mat_23_10_sp2),_mm256_maddubs_epi16(rhs_mat_2367_11_sp2, lhs_mat_23_11_sp2));
+
+                        __m256i iacc_mat_00_2_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_20_sp2, lhs_mat_01_20_sp2),_mm256_maddubs_epi16(rhs_mat_0145_21_sp2, lhs_mat_01_21_sp2));
+                        __m256i iacc_mat_01_2_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_20_sp2, lhs_mat_01_20_sp2),_mm256_maddubs_epi16(rhs_mat_2367_21_sp2, lhs_mat_01_21_sp2));
+
+                        __m256i iacc_mat_10_2_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_20_sp2, lhs_mat_23_20_sp2),_mm256_maddubs_epi16(rhs_mat_0145_21_sp2, lhs_mat_23_21_sp2));
+                        __m256i iacc_mat_11_2_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_20_sp2, lhs_mat_23_20_sp2),_mm256_maddubs_epi16(rhs_mat_2367_21_sp2, lhs_mat_23_21_sp2));
+
+                        __m256i iacc_mat_00_3_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_30_sp2, lhs_mat_01_30_sp2),_mm256_maddubs_epi16(rhs_mat_0145_31_sp2, lhs_mat_01_31_sp2));
+                        __m256i iacc_mat_01_3_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_30_sp2, lhs_mat_01_30_sp2),_mm256_maddubs_epi16(rhs_mat_2367_31_sp2, lhs_mat_01_31_sp2));
+
+                        __m256i iacc_mat_10_3_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_30_sp2, lhs_mat_23_30_sp2),_mm256_maddubs_epi16(rhs_mat_0145_31_sp2, lhs_mat_23_31_sp2));
+                        __m256i iacc_mat_11_3_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_30_sp2, lhs_mat_23_30_sp2),_mm256_maddubs_epi16(rhs_mat_2367_31_sp2, lhs_mat_23_31_sp2));
+
+                        __m256i iacc_mat_00_4_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_40_sp2, lhs_mat_01_40_sp2),_mm256_maddubs_epi16(rhs_mat_0145_41_sp2, lhs_mat_01_41_sp2));
+                        __m256i iacc_mat_01_4_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_40_sp2, lhs_mat_01_40_sp2),_mm256_maddubs_epi16(rhs_mat_2367_41_sp2, lhs_mat_01_41_sp2));
+
+                        __m256i iacc_mat_10_4_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_40_sp2, lhs_mat_23_40_sp2),_mm256_maddubs_epi16(rhs_mat_0145_41_sp2, lhs_mat_23_41_sp2));
+                        __m256i iacc_mat_11_4_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_40_sp2, lhs_mat_23_40_sp2),_mm256_maddubs_epi16(rhs_mat_2367_41_sp2, lhs_mat_23_41_sp2));
+
+                        __m256i iacc_mat_00_5_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_50_sp2, lhs_mat_01_50_sp2),_mm256_maddubs_epi16(rhs_mat_0145_51_sp2, lhs_mat_01_51_sp2));
+                        __m256i iacc_mat_01_5_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_50_sp2, lhs_mat_01_50_sp2),_mm256_maddubs_epi16(rhs_mat_2367_51_sp2, lhs_mat_01_51_sp2));
+
+                        __m256i iacc_mat_10_5_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_50_sp2, lhs_mat_23_50_sp2),_mm256_maddubs_epi16(rhs_mat_0145_51_sp2, lhs_mat_23_51_sp2));
+                        __m256i iacc_mat_11_5_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_50_sp2, lhs_mat_23_50_sp2),_mm256_maddubs_epi16(rhs_mat_2367_51_sp2, lhs_mat_23_51_sp2));
+
+                        __m256i iacc_mat_00_6_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_60_sp2, lhs_mat_01_60_sp2),_mm256_maddubs_epi16(rhs_mat_0145_61_sp2, lhs_mat_01_61_sp2));
+                        __m256i iacc_mat_01_6_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_60_sp2, lhs_mat_01_60_sp2),_mm256_maddubs_epi16(rhs_mat_2367_61_sp2, lhs_mat_01_61_sp2));
+
+                        __m256i iacc_mat_10_6_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_60_sp2, lhs_mat_23_60_sp2),_mm256_maddubs_epi16(rhs_mat_0145_61_sp2, lhs_mat_23_61_sp2));
+                        __m256i iacc_mat_11_6_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_60_sp2, lhs_mat_23_60_sp2),_mm256_maddubs_epi16(rhs_mat_2367_61_sp2, lhs_mat_23_61_sp2));
+
+                        __m256i iacc_mat_00_7_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_70_sp2, lhs_mat_01_70_sp2),_mm256_maddubs_epi16(rhs_mat_0145_71_sp2, lhs_mat_01_71_sp2));
+                        __m256i iacc_mat_01_7_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_70_sp2, lhs_mat_01_70_sp2),_mm256_maddubs_epi16(rhs_mat_2367_71_sp2, lhs_mat_01_71_sp2));
+
+                        __m256i iacc_mat_10_7_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_70_sp2, lhs_mat_23_70_sp2),_mm256_maddubs_epi16(rhs_mat_0145_71_sp2, lhs_mat_23_71_sp2));
+                        __m256i iacc_mat_11_7_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_70_sp2, lhs_mat_23_70_sp2),_mm256_maddubs_epi16(rhs_mat_2367_71_sp2, lhs_mat_23_71_sp2));
+
+                        // Combine results from both shuffle patterns for each output block
+                        __m256i iacc_mat_00_0 = _mm256_add_epi16(iacc_mat_00_0_sp1, iacc_mat_00_0_sp2);
+                        __m256i iacc_mat_01_0 = _mm256_add_epi16(iacc_mat_01_0_sp1, iacc_mat_01_0_sp2);
+                        __m256i iacc_mat_10_0 = _mm256_add_epi16(iacc_mat_10_0_sp1, iacc_mat_10_0_sp2);
+                        __m256i iacc_mat_11_0 = _mm256_add_epi16(iacc_mat_11_0_sp1, iacc_mat_11_0_sp2);
+
+                        __m256i iacc_mat_00_1 = _mm256_add_epi16(iacc_mat_00_1_sp1, iacc_mat_00_1_sp2);
+                        __m256i iacc_mat_01_1 = _mm256_add_epi16(iacc_mat_01_1_sp1, iacc_mat_01_1_sp2);
+                        __m256i iacc_mat_10_1 = _mm256_add_epi16(iacc_mat_10_1_sp1, iacc_mat_10_1_sp2);
+                        __m256i iacc_mat_11_1 = _mm256_add_epi16(iacc_mat_11_1_sp1, iacc_mat_11_1_sp2);
+
+                        __m256i iacc_mat_00_2 = _mm256_add_epi16(iacc_mat_00_2_sp1, iacc_mat_00_2_sp2);
+                        __m256i iacc_mat_01_2 = _mm256_add_epi16(iacc_mat_01_2_sp1, iacc_mat_01_2_sp2);
+                        __m256i iacc_mat_10_2 = _mm256_add_epi16(iacc_mat_10_2_sp1, iacc_mat_10_2_sp2);
+                        __m256i iacc_mat_11_2 = _mm256_add_epi16(iacc_mat_11_2_sp1, iacc_mat_11_2_sp2);
+
+                        __m256i iacc_mat_00_3 = _mm256_add_epi16(iacc_mat_00_3_sp1, iacc_mat_00_3_sp2);
+                        __m256i iacc_mat_01_3 = _mm256_add_epi16(iacc_mat_01_3_sp1, iacc_mat_01_3_sp2);
+                        __m256i iacc_mat_10_3 = _mm256_add_epi16(iacc_mat_10_3_sp1, iacc_mat_10_3_sp2);
+                        __m256i iacc_mat_11_3 = _mm256_add_epi16(iacc_mat_11_3_sp1, iacc_mat_11_3_sp2);
+
+                        __m256i iacc_mat_00_4 = _mm256_add_epi16(iacc_mat_00_4_sp1, iacc_mat_00_4_sp2);
+                        __m256i iacc_mat_01_4 = _mm256_add_epi16(iacc_mat_01_4_sp1, iacc_mat_01_4_sp2);
+                        __m256i iacc_mat_10_4 = _mm256_add_epi16(iacc_mat_10_4_sp1, iacc_mat_10_4_sp2);
+                        __m256i iacc_mat_11_4 = _mm256_add_epi16(iacc_mat_11_4_sp1, iacc_mat_11_4_sp2);
+
+                        __m256i iacc_mat_00_5 = _mm256_add_epi16(iacc_mat_00_5_sp1, iacc_mat_00_5_sp2);
+                        __m256i iacc_mat_01_5 = _mm256_add_epi16(iacc_mat_01_5_sp1, iacc_mat_01_5_sp2);
+                        __m256i iacc_mat_10_5 = _mm256_add_epi16(iacc_mat_10_5_sp1, iacc_mat_10_5_sp2);
+                        __m256i iacc_mat_11_5 = _mm256_add_epi16(iacc_mat_11_5_sp1, iacc_mat_11_5_sp2);
+
+                        __m256i iacc_mat_00_6 = _mm256_add_epi16(iacc_mat_00_6_sp1, iacc_mat_00_6_sp2);
+                        __m256i iacc_mat_01_6 = _mm256_add_epi16(iacc_mat_01_6_sp1, iacc_mat_01_6_sp2);
+                        __m256i iacc_mat_10_6 = _mm256_add_epi16(iacc_mat_10_6_sp1, iacc_mat_10_6_sp2);
+                        __m256i iacc_mat_11_6 = _mm256_add_epi16(iacc_mat_11_6_sp1, iacc_mat_11_6_sp2);
+
+                        __m256i iacc_mat_00_7 = _mm256_add_epi16(iacc_mat_00_7_sp1, iacc_mat_00_7_sp2);
+                        __m256i iacc_mat_01_7 = _mm256_add_epi16(iacc_mat_01_7_sp1, iacc_mat_01_7_sp2);
+                        __m256i iacc_mat_10_7 = _mm256_add_epi16(iacc_mat_10_7_sp1, iacc_mat_10_7_sp2);
+                        __m256i iacc_mat_11_7 = _mm256_add_epi16(iacc_mat_11_7_sp1, iacc_mat_11_7_sp2);
+
+                        // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+                        iacc_mat_00_0 = _mm256_madd_epi16(iacc_mat_00_0, scale_0145_0);
+                        iacc_mat_01_0 = _mm256_madd_epi16(iacc_mat_01_0, scale_2367_0);
+                        iacc_mat_10_0 = _mm256_madd_epi16(iacc_mat_10_0, scale_0145_0);
+                        iacc_mat_11_0 = _mm256_madd_epi16(iacc_mat_11_0, scale_2367_0);
+
+                        iacc_mat_00_1 = _mm256_madd_epi16(iacc_mat_00_1, scale_0145_1);
+                        iacc_mat_01_1 = _mm256_madd_epi16(iacc_mat_01_1, scale_2367_1);
+                        iacc_mat_10_1 = _mm256_madd_epi16(iacc_mat_10_1, scale_0145_1);
+                        iacc_mat_11_1 = _mm256_madd_epi16(iacc_mat_11_1, scale_2367_1);
+
+                        iacc_mat_00_2 = _mm256_madd_epi16(iacc_mat_00_2, scale_0145_2);
+                        iacc_mat_01_2 = _mm256_madd_epi16(iacc_mat_01_2, scale_2367_2);
+                        iacc_mat_10_2 = _mm256_madd_epi16(iacc_mat_10_2, scale_0145_2);
+                        iacc_mat_11_2 = _mm256_madd_epi16(iacc_mat_11_2, scale_2367_2);
+
+                        iacc_mat_00_3 = _mm256_madd_epi16(iacc_mat_00_3, scale_0145_3);
+                        iacc_mat_01_3 = _mm256_madd_epi16(iacc_mat_01_3, scale_2367_3);
+                        iacc_mat_10_3 = _mm256_madd_epi16(iacc_mat_10_3, scale_0145_3);
+                        iacc_mat_11_3 = _mm256_madd_epi16(iacc_mat_11_3, scale_2367_3);
+
+                        iacc_mat_00_4 = _mm256_madd_epi16(iacc_mat_00_4, scale_0145_4);
+                        iacc_mat_01_4 = _mm256_madd_epi16(iacc_mat_01_4, scale_2367_4);
+                        iacc_mat_10_4 = _mm256_madd_epi16(iacc_mat_10_4, scale_0145_4);
+                        iacc_mat_11_4 = _mm256_madd_epi16(iacc_mat_11_4, scale_2367_4);
+
+                        iacc_mat_00_5 = _mm256_madd_epi16(iacc_mat_00_5, scale_0145_5);
+                        iacc_mat_01_5 = _mm256_madd_epi16(iacc_mat_01_5, scale_2367_5);
+                        iacc_mat_10_5 = _mm256_madd_epi16(iacc_mat_10_5, scale_0145_5);
+                        iacc_mat_11_5 = _mm256_madd_epi16(iacc_mat_11_5, scale_2367_5);
+
+                        iacc_mat_00_6 = _mm256_madd_epi16(iacc_mat_00_6, scale_0145_6);
+                        iacc_mat_01_6 = _mm256_madd_epi16(iacc_mat_01_6, scale_2367_6);
+                        iacc_mat_10_6 = _mm256_madd_epi16(iacc_mat_10_6, scale_0145_6);
+                        iacc_mat_11_6 = _mm256_madd_epi16(iacc_mat_11_6, scale_2367_6);
+
+                        iacc_mat_00_7 = _mm256_madd_epi16(iacc_mat_00_7, scale_0145_7);
+                        iacc_mat_01_7 = _mm256_madd_epi16(iacc_mat_01_7, scale_2367_7);
+                        iacc_mat_10_7 = _mm256_madd_epi16(iacc_mat_10_7, scale_0145_7);
+                        iacc_mat_11_7 = _mm256_madd_epi16(iacc_mat_11_7, scale_2367_7);
+
+                        __m256i iacc_mat_00 = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_mat_00_0, iacc_mat_00_1), _mm256_add_epi32(iacc_mat_00_2, iacc_mat_00_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_mat_00_4, iacc_mat_00_5), _mm256_add_epi32(iacc_mat_00_6, iacc_mat_00_7)));
+                        __m256i iacc_mat_01 = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_mat_01_0, iacc_mat_01_1), _mm256_add_epi32(iacc_mat_01_2, iacc_mat_01_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_mat_01_4, iacc_mat_01_5), _mm256_add_epi32(iacc_mat_01_6, iacc_mat_01_7)));
+                        __m256i iacc_mat_10 = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_mat_10_0, iacc_mat_10_1), _mm256_add_epi32(iacc_mat_10_2, iacc_mat_10_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_mat_10_4, iacc_mat_10_5), _mm256_add_epi32(iacc_mat_10_6, iacc_mat_10_7)));
+                        __m256i iacc_mat_11 = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_mat_11_0, iacc_mat_11_1), _mm256_add_epi32(iacc_mat_11_2, iacc_mat_11_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_mat_11_4, iacc_mat_11_5), _mm256_add_epi32(iacc_mat_11_6, iacc_mat_11_7)));
+
+                        // Straighten out to make 4 row vectors
+                        __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
+                        __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
+                        __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
+                        __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
+
+                        // Load the scale(d) values for all the 4 Q8_k blocks and repeat it across lanes
+                        const __m128 row_scale_f32_sse = _mm_load_ps(a_ptrs[rp][b].d);
+                        const __m256 row_scale_f32 = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
+
+                        // Multiply with appropiate scales and accumulate (for both d and dmin) below
+                        acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
+                        acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
+                        acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
+                        acc_rows[rp * 4 + 3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[rp * 4 + 3]);
+
+                        __m256i lhs_bsums_01_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_01_0123), lhs_raw_bsums_01_0123, 1);
+                        __m256i lhs_bsums_23_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_0123), lhs_raw_bsums_23_0123, 1);
+                        __m256i lhs_bsums_01_4567 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_01_4567), lhs_raw_bsums_01_4567, 1);
+                        __m256i lhs_bsums_23_4567 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_4567), lhs_raw_bsums_23_4567, 1);
+
+                       // Take two bsums from two Q8_Ks at a time and multiply with corresponding mins values from each Q2_K
+                        __m256i iacc_row_min_0_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 0), mins_01);
+                        __m256i iacc_row_min_1_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 170), mins_01);
+                        __m256i iacc_row_min_2_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 0), mins_01);
+                        __m256i iacc_row_min_3_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 170), mins_01);
+
+                        __m256i iacc_row_min_0_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 85), mins_23);
+                        __m256i iacc_row_min_1_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 255), mins_23);
+                        __m256i iacc_row_min_2_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 85), mins_23);
+                        __m256i iacc_row_min_3_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 255), mins_23);
+
+                        __m256i iacc_row_min_0_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 0), mins_45);
+                        __m256i iacc_row_min_1_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 170), mins_45);
+                        __m256i iacc_row_min_2_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 0), mins_45);
+                        __m256i iacc_row_min_3_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 170), mins_45);
+
+                        __m256i iacc_row_min_0_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 85), mins_67);
+                        __m256i iacc_row_min_1_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 255), mins_67);
+                        __m256i iacc_row_min_2_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 85), mins_67);
+                        __m256i iacc_row_min_3_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 255), mins_67);
+
+                        __m256i iacc_row_min_0 = _mm256_add_epi32(_mm256_add_epi32(iacc_row_min_0_01, iacc_row_min_0_23), _mm256_add_epi32(iacc_row_min_0_45,iacc_row_min_0_67));
+                        __m256i iacc_row_min_1 = _mm256_add_epi32(_mm256_add_epi32(iacc_row_min_1_01, iacc_row_min_1_23), _mm256_add_epi32(iacc_row_min_1_45,iacc_row_min_1_67));
+                        __m256i iacc_row_min_2 = _mm256_add_epi32(_mm256_add_epi32(iacc_row_min_2_01, iacc_row_min_2_23), _mm256_add_epi32(iacc_row_min_2_45,iacc_row_min_2_67));
+                        __m256i iacc_row_min_3 = _mm256_add_epi32(_mm256_add_epi32(iacc_row_min_3_01, iacc_row_min_3_23), _mm256_add_epi32(iacc_row_min_3_45,iacc_row_min_3_67));
+
+                        acc_min_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_0), _mm256_mul_ps(col_dmin_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_min_rows[rp * 4]);
+                        acc_min_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_1), _mm256_mul_ps(col_dmin_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_min_rows[rp * 4 + 1]);
+                        acc_min_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_2), _mm256_mul_ps(col_dmin_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_min_rows[rp * 4 + 2]);
+                        acc_min_rows[rp * 4 + 3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_3), _mm256_mul_ps(col_dmin_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_min_rows[rp * 4 + 3]);
+
+                    }
+                }
+            }
+            // Store the accumulated values
+            for (int i = 0; i < 16; i++) {
+                _mm256_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm256_sub_ps(acc_rows[i], acc_min_rows[i]));
+
+            }
+        }
+    }
+
+    for (; y < nr / 4; y ++) {
+
+        const block_q8_Kx4 * a_ptr = a_ptr_start + (y * nb);
+
+        // Take group of eight block_q2_kx8 structures at each pass of the loop and perform dot product operation
+        for (int64_t x = xstart; x < nc / 8; x++) {
+
+            const block_q2_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
+
+            // Master FP accumulators
+            __m256 acc_rows[4];
+            for (int i = 0; i < 4; i++) {
+                acc_rows[i] = _mm256_setzero_ps();
+            }
+
+            __m256 acc_min_rows[4];
+            for (int i = 0; i < 4; i++) {
+                acc_min_rows[i] = _mm256_setzero_ps();
+            }
+
+            for (int64_t b = 0; b < nb; b++) {
+                // Delta values - Load the eight scale values of block_q2_kx8
+                const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
+
+                // dmin values - Load the eight dmin values of block_q2_kx8
+                const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin);
+
+                // Loop to iterate over the sixteen sub blocks of a super block - eight sub blocks are processed per iteration
+                for (int sb = 0; sb < QK_K / 128; sb++) {
+
+                    // Load the eight block_q2_k for eight sub blocks quantized values interleaved with each other in chunks of eight bytes - B0,B1 ....B6,B7
+                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + sb * 256));
+                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 128 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 160 + sb * 256));
+                    const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 192 + sb * 256));
+                    const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 224 + sb * 256));
+
+                    // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
+                    //superblock    sub block   which part of sub block
+                    const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+
+                    const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+
+                    const __m256i rhs_raw_mat_0145_2 = _mm256_blend_epi32(rhs_raw_mat_0123_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_2, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_2, requiredOrder), rhs_raw_mat_4567_2, 240);
+
+                    const __m256i rhs_raw_mat_0145_3 = _mm256_blend_epi32(rhs_raw_mat_0123_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_3, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_3, requiredOrder), rhs_raw_mat_4567_3, 240);
+
+                    // 2-bit -> 8-bit
+                    // First sub block of the eight sub blocks processed in the iteration
+                    const __m256i rhs_mat_0145_00 = _mm256_and_si256(rhs_raw_mat_0145_0, m3b); //B00(0-7) B01(0-7) B04(0-7) B05(0-7)
+                    const __m256i rhs_mat_2367_00 = _mm256_and_si256(rhs_raw_mat_2367_0, m3b); //B02(0-7) B03(0-7) B06(0-7) B07(0-7)
+
+                    const __m256i rhs_mat_0145_01 = _mm256_and_si256(rhs_raw_mat_0145_1, m3b); //B00(8-15) B01(8-15) B04(8-15) B05(8-15)
+                    const __m256i rhs_mat_2367_01 = _mm256_and_si256(rhs_raw_mat_2367_1, m3b); //B02(8-15) B03(8-15) B06(8-15) B07(8-15)
+
+                    // Second sub block of the eight sub blocks processed in the iteration
+                    const __m256i rhs_mat_0145_10 = _mm256_and_si256(rhs_raw_mat_0145_2, m3b); //B10(0-7) B11(0-7) B14(0-7) B15(0-7)
+                    const __m256i rhs_mat_2367_10 = _mm256_and_si256(rhs_raw_mat_2367_2, m3b); //B12(0-7) B13(0-7) B16(0-7) B17(0-7)
+
+                    const __m256i rhs_mat_0145_11 = _mm256_and_si256(rhs_raw_mat_0145_3, m3b); //B10(8-15) B11(8-15) B14(8-15) B15(8-15)
+                    const __m256i rhs_mat_2367_11 = _mm256_and_si256(rhs_raw_mat_2367_3, m3b); //B12(8-15) B13(8-15) B16(8-15) B17(8-15)
+
+                    // Third sub block of the eight sub blocks processed in the iteration
+                    const __m256i rhs_mat_0145_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 2), m3b); //B20(0-7) B21(0-7) B24(0-7) B25(0-7)
+                    const __m256i rhs_mat_2367_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 2), m3b); //B22(0-7) B23(0-7) B26(0-7) B27(0-7)
+
+                    const __m256i rhs_mat_0145_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 2), m3b); //B20(8-15) B21(8-15) B24(8-15) B25(8-15)
+                    const __m256i rhs_mat_2367_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 2), m3b); //B22(8-15) B23(8-15) B26(8-15) B27(8-15)
+
+                    // Fourth sub block of the eight sub blocks processed in the iteration
+                    const __m256i rhs_mat_0145_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_2, 2), m3b); //B30(0-7) B31(0-7) B34(0-7) B35(0-7)
+                    const __m256i rhs_mat_2367_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_2, 2), m3b); //B32(0-7) B33(0-7) B36(0-7) B37(0-7)
+
+                    const __m256i rhs_mat_0145_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_3, 2), m3b); //B30(8-15) B31(8-15) B34(8-15) B35(8-15)
+                    const __m256i rhs_mat_2367_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_3, 2), m3b); //B32(8-15) B33(8-15) B36(8-15) B37(8-15)
+
+                    // Fifth sub block of the eight sub blocks processed in the iteration
+                    const __m256i rhs_mat_0145_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m3b); //B40(0-7) B41(0-7) B44(0-7) B45(0-7)
+                    const __m256i rhs_mat_2367_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m3b); //B42(0-7) B43(0-7) B46(0-7) B47(0-7)
+
+                    const __m256i rhs_mat_0145_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m3b); //B40(8-15) B41(8-15) B44(8-15) B45(8-15)
+                    const __m256i rhs_mat_2367_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m3b); //B42(8-15) B43(8-15) B46(8-15) B47(8-15)
+
+                    // Sixth sub block of the eight sub blocks processed in the iteration
+                    const __m256i rhs_mat_0145_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_2, 4), m3b); //B50(0-7) B51(0-7) B54(0-7) B55(0-7)
+                    const __m256i rhs_mat_2367_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_2, 4), m3b); //B52(0-7) B53(0-7) B56(0-7) B57(0-7)
+
+                    const __m256i rhs_mat_0145_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_3, 4), m3b); //B50(8-15) B51(8-15) B54(8-15) B55(8-15)
+                    const __m256i rhs_mat_2367_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_3, 4), m3b); //B52(8-15) B53(8-15) B56(8-15) B57(8-15)
+
+                    // Seventh sub block of the eight sub blocks processed in the iteration
+                    const __m256i rhs_mat_0145_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 6), m3b); //B60(0-7) B61(0-7) B64(0-7) B65(0-7)
+                    const __m256i rhs_mat_2367_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 6), m3b); //B62(0-7) B63(0-7) B66(0-7) B67(0-7)
+
+                    const __m256i rhs_mat_0145_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 6), m3b); //B60(8-15) B61(8-15) B64(8-15) B65(8-15)
+                    const __m256i rhs_mat_2367_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 6), m3b); //B62(8-15) B63(8-15) B66(8-15) B67(8-15)
+
+                    // Eighth sub block of the eight sub blocks processed in the iteration
+                    const __m256i rhs_mat_0145_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_2, 6), m3b); //B70(0-7) B71(0-7) B74(0-7) B75(0-7)
+                    const __m256i rhs_mat_2367_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_2, 6), m3b); //B72(0-7) B73(0-7) B76(0-7) B77(0-7)
+
+                    const __m256i rhs_mat_0145_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_3, 6), m3b); //B70(8-15) B71(8-15) B74(8-15) B75(8-15)
+                    const __m256i rhs_mat_2367_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_3, 6), m3b); //B72(8-15) B73(8-15) B76(8-15) B77(8-15)
+
+                    // Shuffle pattern one - right side input
+                    const __m256i rhs_mat_0145_00_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_00, 136); //B00(0-3) B01(0-3) B00(0-3) B01(0-3) B04(0-3) B05(0-3) B04(0-3) B05(0-3)
+                    const __m256i rhs_mat_2367_00_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_00, 136); //B02(0-3) B03(0-3) B02(0-3) B03(0-3) B06(0-3) B07(0-3) B06(0-3) B07(0-3)
+
+                    const __m256i rhs_mat_0145_01_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_01, 136); //B00(8-11) B01(8-11) B00(8-11) B01(8-11) B04(8-11) B05(8-11) B04(8-11) B05(8-11)
+                    const __m256i rhs_mat_2367_01_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_01, 136); //B02(8-11) B03(8-11) B02(8-11) B03(8-11) B06(8-11) B07(8-11) B06(8-11) B07(8-11)
+
+                    const __m256i rhs_mat_0145_10_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_10, 136); //B10(0-3) B11(0-3) B10(0-3) B11(0-3) B14(0-3) B15(0-3) B14(0-3) B15(0-3)
+                    const __m256i rhs_mat_2367_10_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_10, 136); //B12(0-3) B13(0-3) B12(0-3) B13(0-3) B16(0-3) B17(0-3) B16(0-3) B17(0-3)
+
+                    const __m256i rhs_mat_0145_11_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_11, 136); //B10(8-11) B11(8-11) B10(8-11) B11(8-11) B14(8-11) B15(8-11) B14(8-11) B15(8-11)
+                    const __m256i rhs_mat_2367_11_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_11, 136); //B12(8-11) B13(8-11) B12(8-11) B13(8-11) B16(8-11) B17(8-11) B16(8-11) B17(8-11)
+
+                    const __m256i rhs_mat_0145_20_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_20, 136); //B20(0-3) B21(0-3) B20(0-3) B21(0-3) B24(0-3) B25(0-3) B24(0-3) B25(0-3)
+                    const __m256i rhs_mat_2367_20_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_20, 136); //B22(0-3) B23(0-3) B22(0-3) B23(0-3) B26(0-3) B27(0-3) B26(0-3) B27(0-3)
+
+                    const __m256i rhs_mat_0145_21_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_21, 136); //B20(8-11) B21(8-11) B20(8-11) B21(8-11) B24(8-11) B25(8-11) B24(8-11) B25(8-11)
+                    const __m256i rhs_mat_2367_21_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_21, 136); //B22(8-11) B23(8-11) B22(8-11) B23(8-11) B26(8-11) B27(8-11) B26(8-11) B27(8-11)
+
+                    const __m256i rhs_mat_0145_30_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_30, 136); //B30(0-3) B31(0-3) B30(0-3) B31(0-3) B34(0-3) B35(0-3) B34(0-3) B35(0-3)
+                    const __m256i rhs_mat_2367_30_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_30, 136); //B32(0-3) B33(0-3) B32(0-3) B33(0-3) B36(0-3) B37(0-3) B36(0-3) B37(0-3)
+
+                    const __m256i rhs_mat_0145_31_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_31, 136); //B30(8-11) B31(8-11) B30(8-11) B31(8-11) B34(8-11) B35(8-11) B34(8-11) B35(8-11
+                    const __m256i rhs_mat_2367_31_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_31, 136); //B32(8-11) B33(8-11) B32(8-11) B33(8-11) B36(8-11) B37(8-11) B36(8-11) B37(8-11)
+
+                    const __m256i rhs_mat_0145_40_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_40, 136); //B40(0-3) B41(0-3) B40(0-3) B41(0-3) B44(0-3) B45(0-3) B44(0-3) B45(0-3)
+                    const __m256i rhs_mat_2367_40_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_40, 136); //B42(0-3) B43(0-3) B42(0-3) B43(0-3) B46(0-3) B47(0-3) B46(0-3) B47(0-3)
+
+                    const __m256i rhs_mat_0145_41_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_41, 136); //B40(8-11) B41(8-11) B40(8-11) B41(8-11) B44(8-11) B45(8-11) B44(8-11) B45(8-11)
+                    const __m256i rhs_mat_2367_41_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_41, 136); //B42(8-11) B43(8-11) B42(8-11) B43(8-11) B46(8-11) B47(8-11) B46(8-11) B47(8-11)
+
+                    const __m256i rhs_mat_0145_50_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_50, 136); //B50(0-3) B51(0-3) B50(0-3) B51(0-3) B54(0-3) B55(0-3) B54(0-3) B55(0-3)
+                    const __m256i rhs_mat_2367_50_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_50, 136); //B52(0-3) B53(0-3) B52(0-3) B53(0-3) B56(0-3) B57(0-3) B56(0-3) B57(0-3)
+
+                    const __m256i rhs_mat_0145_51_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_51, 136); //B50(8-11) B51(8-11) B50(8-11) B51(8-11) B54(8-11) B55(8-11) B54(8-11) B55(8-11)
+                    const __m256i rhs_mat_2367_51_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_51, 136); //B52(8-11) B53(8-11) B52(8-11) B53(8-11) B56(8-11) B57(8-11) B56(8-11) B57(8-11)
+
+                    const __m256i rhs_mat_0145_60_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_60, 136); //B60(0-3) B61(0-3) B60(0-3) B61(0-3) B64(0-3) B65(0-3) B64(0-3) B65(0-3)
+                    const __m256i rhs_mat_2367_60_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_60, 136); //B62(0-3) B63(0-3) B62(0-3) B63(0-3) B66(0-3) B67(0-3) B66(0-3) B67(0-3)
+
+                    const __m256i rhs_mat_0145_61_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_61, 136); //B60(8-11) B61(8-11) B60(8-11) B61(8-11) B64(8-11) B65(8-11) B64(8-11) B65(8-11)
+                    const __m256i rhs_mat_2367_61_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_61, 136); //B62(8-11) B63(8-11) B62(8-11) B63(8-11) B66(8-11) B67(8-11) B66(8-11) B67(8-11)
+
+                    const __m256i rhs_mat_0145_70_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_70, 136); //B70(0-3) B71(0-3) B70(0-3) B71(0-3) B74(0-3) B75(0-3) B74(0-3) B75(0-3)
+                    const __m256i rhs_mat_2367_70_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_70, 136); //B72(0-3) B73(0-3) B72(0-3) B73(0-3) B76(0-3) B77(0-3) B76(0-3) B77(0-3)
+
+                    const __m256i rhs_mat_0145_71_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_71, 136); //B70(8-11) B71(8-11) B70(8-11) B71(8-11) B74(8-11) B75(8-11) B74(8-11) B75(8-11)
+                    const __m256i rhs_mat_2367_71_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_71, 136); //B72(8-11) B73(8-11) B72(8-11) B73(8-11) B76(8-11) B77(8-11) B76(8-11) B77(8-11)
+
+
+                    // Shuffle pattern two - right side input
+                    const __m256i rhs_mat_0145_00_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_00, 221); //B00(4-7) B01(4-7) B00(4-7) B01(4-7) B04(4-7) B05(4-7) B04(4-7) B05(4-7)
+                    const __m256i rhs_mat_2367_00_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_00, 221); //B02(4-7) B03(4-7) B02(4-7) B03(4-7) B06(4-7) B07(4-7) B06(4-7) B07(4-7)
+
+                    const __m256i rhs_mat_0145_01_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_01, 221); //B00(12-15) B01(12-15) B00(12-15) B01(12-15) B04(12-15) B05(12-15) B04(12-15) B05(12-15)
+                    const __m256i rhs_mat_2367_01_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_01, 221); //B02(12-15) B03(12-15) B02(12-15) B03(12-15) B06(12-15) B07(12-15) B06(12-15) B07(12-15)
+
+                    const __m256i rhs_mat_0145_10_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_10, 221); //B10(4-7) B11(4-7) B10(4-7) B11(4-7) B14(4-7) B15(4-7) B14(4-7) B15(4-7)
+                    const __m256i rhs_mat_2367_10_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_10, 221); //B12(4-7) B13(4-7) B12(4-7) B13(4-7) B16(4-7) B17(4-7) B16(4-7) B17(4-7)
+
+                    const __m256i rhs_mat_0145_11_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_11, 221); //B10(12-15) B11(12-15) B10(12-15) B11(12-15) B14(12-15) B15(12-15) B14(12-15) B15(12-15)
+                    const __m256i rhs_mat_2367_11_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_11, 221); //B12(12-15) B13(12-15) B12(12-15) B13(12-15) B16(12-15) B17(12-15) B16(12-15) B17(12-15)
+
+                    const __m256i rhs_mat_0145_20_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_20, 221); //B20(4-7) B21(4-7) B20(4-7) B21(4-7) B24(4-7) B25(4-7) B24(4-7) B25(4-7)
+                    const __m256i rhs_mat_2367_20_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_20, 221); //B22(4-7) B23(4-7) B22(4-7) B23(4-7) B26(4-7) B27(4-7) B26(4-7) B27(4-7)
+
+                    const __m256i rhs_mat_0145_21_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_21, 221); //B20(12-15) B21(12-15) B20(12-15) B21(12-15) B24(12-15) B25(12-15) B24(12-15) B25(12-15)
+                    const __m256i rhs_mat_2367_21_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_21, 221); //B22(12-15) B23(12-15) B22(12-15) B23(12-15) B26(12-15) B27(12-15) B26(12-15) B27(12-15)
+
+                    const __m256i rhs_mat_0145_30_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_30, 221); //B30(4-7) B31(4-7) B30(4-7) B31(4-7) B34(4-7) B35(4-7) B34(4-7) B35(4-7)
+                    const __m256i rhs_mat_2367_30_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_30, 221); //B32(4-7) B33(4-7) B32(4-7) B33(4-7) B36(4-7) B37(4-7) B36(4-7) B37(4-7)
+
+                    const __m256i rhs_mat_0145_31_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_31, 221); //B30(12-15) B31(12-15) B30(12-15) B31(12-15) B34(12-15) B35(12-15) B34(12-15) B35(12-15)
+                    const __m256i rhs_mat_2367_31_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_31, 221); //B32(12-15) B33(12-15) B32(12-15) B33(12-15) B36(12-15) B37(12-15) B36(12-15) B37(12-15)
+
+                    const __m256i rhs_mat_0145_40_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_40, 221); //B40(4-7) B41(4-7) B40(4-7) B41(4-7) B44(4-7) B45(4-7) B44(4-7) B45(4-7)
+                    const __m256i rhs_mat_2367_40_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_40, 221); //B42(4-7) B43(4-7) B42(4-7) B43(4-7) B46(4-7) B47(4-7) B46(4-7) B47(4-7)
+
+                    const __m256i rhs_mat_0145_41_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_41, 221); //B40(12-15) B41(12-15) B40(12-15) B41(12-15) B44(12-15) B45(12-15) B44(12-15) B45(12-15)
+                    const __m256i rhs_mat_2367_41_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_41, 221); //B42(12-15) B43(12-15) B42(12-15) B43(12-15) B46(12-15) B47(12-15) B46(12-15) B47(12-15)
+
+                    const __m256i rhs_mat_0145_50_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_50, 221); //B50(4-7) B51(4-7) B50(4-7) B51(4-7) B54(4-7) B55(4-7) B54(4-7) B55(4-7)
+                    const __m256i rhs_mat_2367_50_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_50, 221); //B52(4-7) B53(4-7) B52(4-7) B53(4-7) B56(4-7) B57(4-7) B56(4-7) B57(4-7)
+
+                    const __m256i rhs_mat_0145_51_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_51, 221); //B50(12-15) B51(12-15) B50(12-15) B51(12-15) B54(12-15) B55(12-15) B54(12-15) B55(12-15)
+                    const __m256i rhs_mat_2367_51_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_51, 221); //B52(12-15) B53(12-15) B52(12-15) B53(12-15) B56(12-15) B57(12-15) B56(12-15) B57(12-15)
+
+                    const __m256i rhs_mat_0145_60_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_60, 221); //B60(4-7) B61(4-7) B60(4-7) B61(4-7) B64(4-7) B65(4-7) B64(4-7) B65(4-7)
+                    const __m256i rhs_mat_2367_60_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_60, 221); //B62(4-7) B63(4-7) B62(4-7) B63(4-7) B66(4-7) B67(4-7) B66(4-7) B67(4-7)
+
+                    const __m256i rhs_mat_0145_61_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_61, 221); //B60(12-15) B61(12-15) B60(12-15) B61(12-15) B64(12-15) B65(12-15) B64(12-15) B65(12-15)
+                    const __m256i rhs_mat_2367_61_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_61, 221); //B62(12-15) B63(12-15) B62(12-15) B63(12-15) B66(12-15) B67(12-15) B66(12-15) B67(12-15)
+
+                    const __m256i rhs_mat_0145_70_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_70, 221); //B70(4-7) B71(4-7) B70(4-7) B71(4-7) B74(4-7) B75(4-7) B74(4-7) B75(4-7)
+                    const __m256i rhs_mat_2367_70_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_70, 221); //B72(4-7) B73(4-7) B72(4-7) B73(4-7) B76(4-7) B77(4-7) B76(4-7) B77(4-7)
+
+                    const __m256i rhs_mat_0145_71_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_71, 221); //B70(12-15) B71(12-15) B70(12-15) B71(12-15) B74(12-15) B75(12-15) B74(12-15) B75(12-15)
+                    const __m256i rhs_mat_2367_71_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_71, 221); //B72(12-15) B73(12-15) B72(12-15) B73(12-15) B76(12-15) B77(12-15) B76(12-15) B77(12-15)
+
+
+                    //Scales and Mins of corresponding sub blocks from different Q2_K structures are stored together
+                    //s00 m00  s01 m01   s10 m10  s11 m11  s20 m20  s21 m21   s30 m30  s31 m31  s40 m40  s41 m41   s50 m50  s51 m51  s60 m60  s61 m61   s70 m70  s71 m71
+
+                    // Combine mins and scales for sub-blocks: 0-1, 2-3, 4-5, 6-7 in the sb loop
+                    const __m128i mins_and_scales_01 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + sb * 64));
+                    const __m128i mins_and_scales_23 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 16 + sb * 64));
+                    const __m128i mins_and_scales_45 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 32 + sb * 64));
+                    const __m128i mins_and_scales_67 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 48 + sb * 64));
+
+                    // Extract scales which is lower half from mins_and_scales
+                    const __m128i scales_01 = _mm_and_si128(mins_and_scales_01, m4b_sse);
+                    const __m128i scales_23 = _mm_and_si128(mins_and_scales_23, m4b_sse);
+                    const __m128i scales_45 = _mm_and_si128(mins_and_scales_45, m4b_sse);
+                    const __m128i scales_67 = _mm_and_si128(mins_and_scales_67, m4b_sse);
+
+                    // Extract mins which is upper half from mins_and_scales
+                    const __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_01, 4), m4b_sse));
+                    const __m256i mins_23 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_23, 4), m4b_sse));
+                    const __m256i mins_45 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_45, 4), m4b_sse));
+                    const __m256i mins_67 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_67, 4), m4b_sse));
+
+                    const __m256i scales_0 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_01, scalesmask1_sse));
+                    const __m256i scales_1 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_01, scalesmask2_sse));
+
+                    const __m256i scales_2 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_23, scalesmask1_sse));
+                    const __m256i scales_3 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_23, scalesmask2_sse));
+
+                    const __m256i scales_4 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_45, scalesmask1_sse));
+                    const __m256i scales_5 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_45, scalesmask2_sse));
+
+                    const __m256i scales_6 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_67, scalesmask1_sse));
+                    const __m256i scales_7 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_67, scalesmask2_sse));
+
+                    const __m256i scale_0145_0 = _mm256_shuffle_epi32(scales_0, 68);
+                    const __m256i scale_2367_0 = _mm256_shuffle_epi32(scales_0, 238);
+
+                    const __m256i scale_0145_1 = _mm256_shuffle_epi32(scales_1, 68);
+                    const __m256i scale_2367_1 = _mm256_shuffle_epi32(scales_1, 238);
+
+                    const __m256i scale_0145_2 = _mm256_shuffle_epi32(scales_2, 68);
+                    const __m256i scale_2367_2 = _mm256_shuffle_epi32(scales_2, 238);
+
+                    const __m256i scale_0145_3 = _mm256_shuffle_epi32(scales_3, 68);
+                    const __m256i scale_2367_3 = _mm256_shuffle_epi32(scales_3, 238);
+
+                    const __m256i scale_0145_4 = _mm256_shuffle_epi32(scales_4, 68);
+                    const __m256i scale_2367_4 = _mm256_shuffle_epi32(scales_4, 238);
+
+                    const __m256i scale_0145_5 = _mm256_shuffle_epi32(scales_5, 68);
+                    const __m256i scale_2367_5 = _mm256_shuffle_epi32(scales_5, 238);
+
+                    const __m256i scale_0145_6 = _mm256_shuffle_epi32(scales_6, 68);
+                    const __m256i scale_2367_6 = _mm256_shuffle_epi32(scales_6, 238);
+
+                    const __m256i scale_0145_7 = _mm256_shuffle_epi32(scales_7, 68);
+                    const __m256i scale_2367_7 = _mm256_shuffle_epi32(scales_7, 238);
+
+                    // Load the four block_q8_k quantized values interleaved with each other in chunks of eight bytes - A0,A1,A2,A3
+                    // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
+                    __m256i lhs_mat_0123_00 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 512 * sb)));
+                    __m256i lhs_mat_01_00 = _mm256_permute2f128_si256(lhs_mat_0123_00, lhs_mat_0123_00, 0);
+                    __m256i lhs_mat_23_00 = _mm256_permute2f128_si256(lhs_mat_0123_00, lhs_mat_0123_00, 17);
+                    __m256i lhs_mat_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 32 + 512 * sb)));
+                    __m256i lhs_mat_01_01 = _mm256_permute2f128_si256(lhs_mat_0123_01, lhs_mat_0123_01, 0);
+                    __m256i lhs_mat_23_01 = _mm256_permute2f128_si256(lhs_mat_0123_01, lhs_mat_0123_01, 17);
+                    __m256i lhs_mat_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 64 + 512 * sb)));
+                    __m256i lhs_mat_01_10 = _mm256_permute2f128_si256(lhs_mat_0123_10, lhs_mat_0123_10, 0);
+                    __m256i lhs_mat_23_10 = _mm256_permute2f128_si256(lhs_mat_0123_10, lhs_mat_0123_10, 17);
+                    __m256i lhs_mat_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 96 + 512 * sb)));
+                    __m256i lhs_mat_01_11 = _mm256_permute2f128_si256(lhs_mat_0123_11, lhs_mat_0123_11, 0);
+                    __m256i lhs_mat_23_11 = _mm256_permute2f128_si256(lhs_mat_0123_11, lhs_mat_0123_11, 17);
+                    __m256i lhs_mat_0123_20 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 128 + 512 * sb)));
+                    __m256i lhs_mat_01_20 = _mm256_permute2f128_si256(lhs_mat_0123_20, lhs_mat_0123_20, 0);
+                    __m256i lhs_mat_23_20 = _mm256_permute2f128_si256(lhs_mat_0123_20, lhs_mat_0123_20, 17);
+                    __m256i lhs_mat_0123_21 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 160 + 512 * sb)));
+                    __m256i lhs_mat_01_21 = _mm256_permute2f128_si256(lhs_mat_0123_21, lhs_mat_0123_21, 0);
+                    __m256i lhs_mat_23_21 = _mm256_permute2f128_si256(lhs_mat_0123_21, lhs_mat_0123_21, 17);
+                    __m256i lhs_mat_0123_30 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 192 + 512 * sb)));
+                    __m256i lhs_mat_01_30 = _mm256_permute2f128_si256(lhs_mat_0123_30, lhs_mat_0123_30, 0);
+                    __m256i lhs_mat_23_30 = _mm256_permute2f128_si256(lhs_mat_0123_30, lhs_mat_0123_30, 17);
+                    __m256i lhs_mat_0123_31 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 224 + 512 * sb)));
+                    __m256i lhs_mat_01_31 = _mm256_permute2f128_si256(lhs_mat_0123_31, lhs_mat_0123_31, 0);
+                    __m256i lhs_mat_23_31 = _mm256_permute2f128_si256(lhs_mat_0123_31, lhs_mat_0123_31, 17);
+
+                    __m256i lhs_mat_0123_40 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 256 + 512 * sb)));
+                    __m256i lhs_mat_01_40 = _mm256_permute2f128_si256(lhs_mat_0123_40, lhs_mat_0123_40, 0);
+                    __m256i lhs_mat_23_40 = _mm256_permute2f128_si256(lhs_mat_0123_40, lhs_mat_0123_40, 17);
+                    __m256i lhs_mat_0123_41 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 288 + 512 * sb)));
+                    __m256i lhs_mat_01_41 = _mm256_permute2f128_si256(lhs_mat_0123_41, lhs_mat_0123_41, 0);
+                    __m256i lhs_mat_23_41 = _mm256_permute2f128_si256(lhs_mat_0123_41, lhs_mat_0123_41, 17);
+                    __m256i lhs_mat_0123_50 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 320 + 512 * sb)));
+                    __m256i lhs_mat_01_50 = _mm256_permute2f128_si256(lhs_mat_0123_50, lhs_mat_0123_50, 0);
+                    __m256i lhs_mat_23_50 = _mm256_permute2f128_si256(lhs_mat_0123_50, lhs_mat_0123_50, 17);
+                    __m256i lhs_mat_0123_51 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 352 + 512 * sb)));
+                    __m256i lhs_mat_01_51 = _mm256_permute2f128_si256(lhs_mat_0123_51, lhs_mat_0123_51, 0);
+                    __m256i lhs_mat_23_51 = _mm256_permute2f128_si256(lhs_mat_0123_51, lhs_mat_0123_51, 17);
+                    __m256i lhs_mat_0123_60 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 384 + 512 * sb)));
+                    __m256i lhs_mat_01_60 = _mm256_permute2f128_si256(lhs_mat_0123_60, lhs_mat_0123_60, 0);
+                    __m256i lhs_mat_23_60 = _mm256_permute2f128_si256(lhs_mat_0123_60, lhs_mat_0123_60, 17);
+                    __m256i lhs_mat_0123_61 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 416 + 512 * sb)));
+                    __m256i lhs_mat_01_61 = _mm256_permute2f128_si256(lhs_mat_0123_61, lhs_mat_0123_61, 0);
+                    __m256i lhs_mat_23_61 = _mm256_permute2f128_si256(lhs_mat_0123_61, lhs_mat_0123_61, 17);
+                    __m256i lhs_mat_0123_70 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 448 + 512 * sb)));
+                    __m256i lhs_mat_01_70 = _mm256_permute2f128_si256(lhs_mat_0123_70, lhs_mat_0123_70, 0);
+                    __m256i lhs_mat_23_70 = _mm256_permute2f128_si256(lhs_mat_0123_70, lhs_mat_0123_70, 17);
+                    __m256i lhs_mat_0123_71 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 480 + 512 * sb)));
+                    __m256i lhs_mat_01_71 = _mm256_permute2f128_si256(lhs_mat_0123_71, lhs_mat_0123_71, 0);
+                    __m256i lhs_mat_23_71 = _mm256_permute2f128_si256(lhs_mat_0123_71, lhs_mat_0123_71, 17);
+
+                    // Bsums are loaded for the different Q8_K blocks
+                    __m128i lhs_raw_bsums_01_0123 = _mm_loadu_si128((const __m128i *)((a_ptr[b].bsums + 32 * sb)));
+                    __m128i lhs_raw_bsums_23_0123 = _mm_loadu_si128((const __m128i *)(a_ptr[b].bsums + 8 + 32 * sb));
+                    __m128i lhs_raw_bsums_01_4567 = _mm_loadu_si128((const __m128i *)((a_ptr[b].bsums + 16 + 32 * sb)));
+                    __m128i lhs_raw_bsums_23_4567 = _mm_loadu_si128((const __m128i *)(a_ptr[b].bsums + 24 + 32 * sb));
+
+                    // Shuffle pattern one - left side input
+                    const __m256i lhs_mat_01_00_sp1 = _mm256_shuffle_epi32(lhs_mat_01_00, 160); //A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3)
+                    const __m256i lhs_mat_23_00_sp1 = _mm256_shuffle_epi32(lhs_mat_23_00, 160); //A02(0-3) A03(0-3) A02(0-3) A03(0-3) A02(0-3) A03(0-3) A02(0-3) A03(0-3)
+
+                    const __m256i lhs_mat_01_01_sp1 = _mm256_shuffle_epi32(lhs_mat_01_01, 160); //A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11)
+                    const __m256i lhs_mat_23_01_sp1 = _mm256_shuffle_epi32(lhs_mat_23_01, 160); //A02(8-11) A03(8-11) A02(8-11) A03(8-11) A02(8-11) A03(8-11) A02(8-11) A03(8-11)
+
+                    const __m256i lhs_mat_01_10_sp1 = _mm256_shuffle_epi32(lhs_mat_01_10, 160); //A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3)
+                    const __m256i lhs_mat_23_10_sp1 = _mm256_shuffle_epi32(lhs_mat_23_10, 160); //A12(0-3) A13(0-3) A12(0-3) A13(0-3) A12(0-3) A13(0-3) A12(0-3) A13(0-3)
+
+                    const __m256i lhs_mat_01_11_sp1 = _mm256_shuffle_epi32(lhs_mat_01_11, 160); //A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11)
+                    const __m256i lhs_mat_23_11_sp1 = _mm256_shuffle_epi32(lhs_mat_23_11, 160); //A12(8-11) A13(8-11) A12(8-11) A13(8-11) A12(8-11) A13(8-11) A12(8-11) A13(8-11)
+
+                    const __m256i lhs_mat_01_20_sp1 = _mm256_shuffle_epi32(lhs_mat_01_20, 160); //A20(0-3) A20(0-3) A21(0-3) A21(0-3) A20(0-3) A20(0-3) A21(0-3) A21(0-3)
+                    const __m256i lhs_mat_23_20_sp1 = _mm256_shuffle_epi32(lhs_mat_23_20, 160); //A22(0-3) A23(0-3) A22(0-3) A23(0-3) A22(0-3) A23(0-3) A22(0-3) A23(0-3)
+
+                    const __m256i lhs_mat_01_21_sp1 = _mm256_shuffle_epi32(lhs_mat_01_21, 160); //A20(8-11) A20(8-11) A21(8-11) A21(8-11) A20(8-11) A20(8-11) A21(8-11) A21(8-11)
+                    const __m256i lhs_mat_23_21_sp1 = _mm256_shuffle_epi32(lhs_mat_23_21, 160); //A22(8-11) A23(8-11) A22(8-11) A23(8-11) A22(8-11) A23(8-11) A22(8-11) A23(8-11)
+
+                    const __m256i lhs_mat_01_30_sp1 = _mm256_shuffle_epi32(lhs_mat_01_30, 160); //A30(0-3) A30(0-3) A31(0-3) A31(0-3) A30(0-3) A30(0-3) A31(0-3) A31(0-3)
+                    const __m256i lhs_mat_23_30_sp1 = _mm256_shuffle_epi32(lhs_mat_23_30, 160); //A32(0-3) A33(0-3) A32(0-3) A33(0-3) A32(0-3) A33(0-3) A32(0-3) A33(0-3)
+
+                    const __m256i lhs_mat_01_31_sp1 = _mm256_shuffle_epi32(lhs_mat_01_31, 160); //A30(8-11) A30(8-11) A31(8-11) A31(8-11) A30(8-11) A30(8-11) A31(8-11) A31(8-11)
+                    const __m256i lhs_mat_23_31_sp1 = _mm256_shuffle_epi32(lhs_mat_23_31, 160); //A32(8-11) A33(8-11) A32(8-11) A33(8-11) A32(8-11) A33(8-11) A32(8-11) A33(8-11)
+
+                    const __m256i lhs_mat_01_40_sp1 = _mm256_shuffle_epi32(lhs_mat_01_40, 160); //A40(0-3) A40(0-3) A41(0-3) A41(0-3) A40(0-3) A40(0-3) A41(0-3) A41(0-3)
+                    const __m256i lhs_mat_23_40_sp1 = _mm256_shuffle_epi32(lhs_mat_23_40, 160); //A42(0-3) A43(0-3) A42(0-3) A43(0-3) A42(0-3) A43(0-3) A42(0-3) A43(0-3)
+
+                    const __m256i lhs_mat_01_41_sp1 = _mm256_shuffle_epi32(lhs_mat_01_41, 160); //A40(8-11) A40(8-11) A41(8-11) A41(8-11) A40(8-11) A40(8-11) A41(8-11) A41(8-11)
+                    const __m256i lhs_mat_23_41_sp1 = _mm256_shuffle_epi32(lhs_mat_23_41, 160); //A42(8-11) A43(8-11) A42(8-11) A43(8-11) A42(8-11) A43(8-11) A42(8-11) A43(8-11)
+
+                    const __m256i lhs_mat_01_50_sp1 = _mm256_shuffle_epi32(lhs_mat_01_50, 160); //A50(0-3) A50(0-3) A51(0-3) A51(0-3) A50(0-3) A50(0-3) A51(0-3) A51(0-3)
+                    const __m256i lhs_mat_23_50_sp1 = _mm256_shuffle_epi32(lhs_mat_23_50, 160); //A52(0-3) A53(0-3) A52(0-3) A53(0-3) A52(0-3) A53(0-3) A52(0-3) A53(0-3)
+
+                    const __m256i lhs_mat_01_51_sp1 = _mm256_shuffle_epi32(lhs_mat_01_51, 160); //A50(8-11) A50(8-11) A51(8-11) A51(8-11) A50(8-11) A50(8-11) A51(8-11) A51(8-11)
+                    const __m256i lhs_mat_23_51_sp1 = _mm256_shuffle_epi32(lhs_mat_23_51, 160); //A52(8-11) A53(8-11) A52(8-11) A53(8-11) A52(8-11) A53(8-11) A52(8-11) A53(8-11)
+
+                    const __m256i lhs_mat_01_60_sp1 = _mm256_shuffle_epi32(lhs_mat_01_60, 160); //A60(0-3) A60(0-3) A61(0-3) A61(0-3) A60(0-3) A60(0-3) A61(0-3) A61(0-3)
+                    const __m256i lhs_mat_23_60_sp1 = _mm256_shuffle_epi32(lhs_mat_23_60, 160); //A62(0-3) A63(0-3) A62(0-3) A63(0-3) A62(0-3) A63(0-3) A62(0-3) A63(0-3)
+
+                    const __m256i lhs_mat_01_61_sp1 = _mm256_shuffle_epi32(lhs_mat_01_61, 160); //A60(8-11) A60(8-11) A61(8-11) A61(8-11) A60(8-11) A60(8-11) A61(8-11) A61(8-11)
+                    const __m256i lhs_mat_23_61_sp1 = _mm256_shuffle_epi32(lhs_mat_23_61, 160); //A62(8-11) A63(8-11) A62(8-11) A63(8-11) A62(8-11) A63(8-11) A62(8-11) A63(8-11)
+
+                    const __m256i lhs_mat_01_70_sp1 = _mm256_shuffle_epi32(lhs_mat_01_70, 160); //A70(0-3) A70(0-3) A71(0-3) A71(0-3) A70(0-3) A70(0-3) A71(0-3) A71(0-3)
+                    const __m256i lhs_mat_23_70_sp1 = _mm256_shuffle_epi32(lhs_mat_23_70, 160); //A72(0-3) A73(0-3) A72(0-3) A73(0-3) A72(0-3) A73(0-3) A72(0-3) A73(0-3)
+
+                    const __m256i lhs_mat_01_71_sp1 = _mm256_shuffle_epi32(lhs_mat_01_71, 160); //A70(8-11) A70(8-11) A71(8-11) A71(8-11) A70(8-11) A70(8-11) A71(8-11) A71(8-11)
+                    const __m256i lhs_mat_23_71_sp1 = _mm256_shuffle_epi32(lhs_mat_23_71, 160); //A72(8-11) A73(8-11) A72(8-11) A73(8-11) A72(8-11) A73(8-11) A72(8-11) A73(8-11)
+
+                    // Shuffle pattern two- left side input
+                    const __m256i lhs_mat_01_00_sp2 = _mm256_shuffle_epi32(lhs_mat_01_00, 245); //A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7)
+                    const __m256i lhs_mat_23_00_sp2 = _mm256_shuffle_epi32(lhs_mat_23_00, 245); //A02(4-7) A03(4-7) A02(4-7) A03(4-7) A02(4-7) A03(4-7) A02(4-7) A03(4-7)
+
+                    const __m256i lhs_mat_01_01_sp2 = _mm256_shuffle_epi32(lhs_mat_01_01, 245); //A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15)
+                    const __m256i lhs_mat_23_01_sp2 = _mm256_shuffle_epi32(lhs_mat_23_01, 245); //A02(12-15) A03(12-15) A02(12-15) A03(12-15) A02(12-15) A03(12-15) A02(12-15) A03(12-15)
+
+                    const __m256i lhs_mat_01_10_sp2 = _mm256_shuffle_epi32(lhs_mat_01_10, 245); //A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7)
+                    const __m256i lhs_mat_23_10_sp2 = _mm256_shuffle_epi32(lhs_mat_23_10, 245); //A12(4-7) A13(4-7) A12(4-7) A13(4-7) A12(4-7) A13(4-7) A12(4-7) A13(4-7)
+
+                    const __m256i lhs_mat_01_11_sp2 = _mm256_shuffle_epi32(lhs_mat_01_11, 245); //A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15)
+                    const __m256i lhs_mat_23_11_sp2 = _mm256_shuffle_epi32(lhs_mat_23_11, 245); //A12(12-15) A13(12-15) A12(12-15) A13(12-15) A12(12-15) A13(12-15) A12(12-15) A13(12-15)
+
+                    const __m256i lhs_mat_01_20_sp2 = _mm256_shuffle_epi32(lhs_mat_01_20, 245); //A20(4-7) A20(4-7) A21(4-7) A21(4-7) A20(4-7) A20(4-7) A21(4-7) A21(4-7)
+                    const __m256i lhs_mat_23_20_sp2 = _mm256_shuffle_epi32(lhs_mat_23_20, 245); //A22(4-7) A23(4-7) A22(4-7) A23(4-7) A22(4-7) A23(4-7) A22(4-7) A23(4-7)
+
+                    const __m256i lhs_mat_01_21_sp2 = _mm256_shuffle_epi32(lhs_mat_01_21, 245); //A20(12-15) A20(12-15) A21(12-15) A21(12-15) A20(12-15) A20(12-15) A21(12-15) A21(12-15)
+                    const __m256i lhs_mat_23_21_sp2 = _mm256_shuffle_epi32(lhs_mat_23_21, 245); //A22(12-15) A23(12-15) A22(12-15) A23(12-15) A22(12-15) A23(12-15) A22(12-15) A23(12-15)
+
+                    const __m256i lhs_mat_01_30_sp2 = _mm256_shuffle_epi32(lhs_mat_01_30, 245); //A30(4-7) A30(4-7) A31(4-7) A31(4-7) A30(4-7) A30(4-7) A31(4-7) A31(4-7)
+                    const __m256i lhs_mat_23_30_sp2 = _mm256_shuffle_epi32(lhs_mat_23_30, 245); //A32(4-7) A33(4-7) A32(4-7) A33(4-7) A32(4-7) A33(4-7) A32(4-7) A33(4-7)
+
+                    const __m256i lhs_mat_01_31_sp2 = _mm256_shuffle_epi32(lhs_mat_01_31, 245); //A30(12-15) A30(12-15) A31(12-15) A31(12-15) A30(12-15) A30(12-15) A31(12-15) A31(12-15)
+                    const __m256i lhs_mat_23_31_sp2 = _mm256_shuffle_epi32(lhs_mat_23_31, 245); //A32(12-15) A33(12-15) A32(12-15) A33(12-15) A32(12-15) A33(12-15) A32(12-15) A33(12-15)
+
+                    const __m256i lhs_mat_01_40_sp2 = _mm256_shuffle_epi32(lhs_mat_01_40, 245); //A40(4-7) A40(4-7) A41(4-7) A41(4-7) A40(4-7) A40(4-7) A41(4-7) A41(4-7)
+                    const __m256i lhs_mat_23_40_sp2 = _mm256_shuffle_epi32(lhs_mat_23_40, 245); //A42(4-7) A43(4-7) A42(4-7) A43(4-7) A42(4-7) A43(4-7) A42(4-7) A43(4-7)
+
+                    const __m256i lhs_mat_01_41_sp2 = _mm256_shuffle_epi32(lhs_mat_01_41, 245); //A40(12-15) A40(12-15) A41(12-15) A41(12-15) A40(12-15) A40(12-15) A41(12-15) A41(12-15)
+                    const __m256i lhs_mat_23_41_sp2 = _mm256_shuffle_epi32(lhs_mat_23_41, 245); //A42(12-15) A43(12-15) A42(12-15) A43(12-15) A42(12-15) A43(12-15) A42(12-15) A43(12-15)
+
+                    const __m256i lhs_mat_01_50_sp2 = _mm256_shuffle_epi32(lhs_mat_01_50, 245); //A50(4-7) A50(4-7) A51(4-7) A51(4-7) A50(4-7) A50(4-7) A51(4-7) A51(4-7)
+                    const __m256i lhs_mat_23_50_sp2 = _mm256_shuffle_epi32(lhs_mat_23_50, 245); //A52(4-7) A53(4-7) A52(4-7) A53(4-7) A52(4-7) A53(4-7) A52(4-7) A53(4-7)
+
+                    const __m256i lhs_mat_01_51_sp2 = _mm256_shuffle_epi32(lhs_mat_01_51, 245); //A50(12-15) A50(12-15) A51(12-15) A51(12-15) A50(12-15) A50(12-15) A51(12-15) A51(12-15)
+                    const __m256i lhs_mat_23_51_sp2 = _mm256_shuffle_epi32(lhs_mat_23_51, 245); //A52(12-15) A53(12-15) A52(12-15) A53(12-15) A52(12-15) A53(12-15) A52(12-15) A53(12-15)
+
+                    const __m256i lhs_mat_01_60_sp2 = _mm256_shuffle_epi32(lhs_mat_01_60, 245); //A60(4-7) A60(4-7) A61(4-7) A61(4-7) A60(4-7) A60(4-7) A61(4-7) A61(4-7)
+                    const __m256i lhs_mat_23_60_sp2 = _mm256_shuffle_epi32(lhs_mat_23_60, 245); //A62(4-7) A63(4-7) A62(4-7) A63(4-7) A62(4-7) A63(4-7) A62(4-7) A63(4-7)
+
+                    const __m256i lhs_mat_01_61_sp2 = _mm256_shuffle_epi32(lhs_mat_01_61, 245); //A60(12-15) A60(12-15) A61(12-15) A61(12-15) A60(12-15) A60(12-15) A61(12-15) A61(12-15)
+                    const __m256i lhs_mat_23_61_sp2 = _mm256_shuffle_epi32(lhs_mat_23_61, 245); //A62(12-15) A63(12-15) A62(12-15) A63(12-15) A62(12-15) A63(12-15) A62(12-15) A63(12-15)
+
+                    const __m256i lhs_mat_01_70_sp2 = _mm256_shuffle_epi32(lhs_mat_01_70, 245); //A70(4-7) A70(4-7) A71(4-7) A71(4-7) A70(4-7) A70(4-7) A71(4-7) A71(4-7)
+                    const __m256i lhs_mat_23_70_sp2 = _mm256_shuffle_epi32(lhs_mat_23_70, 245); //A72(4-7) A73(4-7) A72(4-7) A73(4-7) A72(4-7) A73(4-7) A72(4-7) A73(4-7)
+
+                    const __m256i lhs_mat_01_71_sp2 = _mm256_shuffle_epi32(lhs_mat_01_71, 245); //A70(12-15) A70(12-15) A71(12-15) A71(12-15) A70(12-15) A70(12-15) A71(12-15) A71(12-15)
+                    const __m256i lhs_mat_23_71_sp2 = _mm256_shuffle_epi32(lhs_mat_23_71, 245); //A72(12-15) A73(12-15) A72(12-15) A73(12-15) A72(12-15) A73(12-15) A72(12-15) A73(12-15)
+
+                    // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+                    __m256i iacc_mat_00_0_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_00_sp1, lhs_mat_01_00_sp1),_mm256_maddubs_epi16(rhs_mat_0145_01_sp1, lhs_mat_01_01_sp1));
+                    __m256i iacc_mat_01_0_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_00_sp1, lhs_mat_01_00_sp1),_mm256_maddubs_epi16(rhs_mat_2367_01_sp1, lhs_mat_01_01_sp1));
+
+                    __m256i iacc_mat_10_0_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_00_sp1, lhs_mat_23_00_sp1),_mm256_maddubs_epi16(rhs_mat_0145_01_sp1, lhs_mat_23_01_sp1));
+                    __m256i iacc_mat_11_0_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_00_sp1, lhs_mat_23_00_sp1),_mm256_maddubs_epi16(rhs_mat_2367_01_sp1, lhs_mat_23_01_sp1));
+
+                    __m256i iacc_mat_00_1_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_10_sp1, lhs_mat_01_10_sp1),_mm256_maddubs_epi16(rhs_mat_0145_11_sp1, lhs_mat_01_11_sp1));
+                    __m256i iacc_mat_01_1_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_10_sp1, lhs_mat_01_10_sp1),_mm256_maddubs_epi16(rhs_mat_2367_11_sp1, lhs_mat_01_11_sp1));
+
+                    __m256i iacc_mat_10_1_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_10_sp1, lhs_mat_23_10_sp1),_mm256_maddubs_epi16(rhs_mat_0145_11_sp1, lhs_mat_23_11_sp1));
+                    __m256i iacc_mat_11_1_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_10_sp1, lhs_mat_23_10_sp1),_mm256_maddubs_epi16(rhs_mat_2367_11_sp1, lhs_mat_23_11_sp1));
+
+                    __m256i iacc_mat_00_2_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_20_sp1, lhs_mat_01_20_sp1),_mm256_maddubs_epi16(rhs_mat_0145_21_sp1, lhs_mat_01_21_sp1));
+                    __m256i iacc_mat_01_2_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_20_sp1, lhs_mat_01_20_sp1),_mm256_maddubs_epi16(rhs_mat_2367_21_sp1, lhs_mat_01_21_sp1));
+
+                    __m256i iacc_mat_10_2_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_20_sp1, lhs_mat_23_20_sp1),_mm256_maddubs_epi16(rhs_mat_0145_21_sp1, lhs_mat_23_21_sp1));
+                    __m256i iacc_mat_11_2_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_20_sp1, lhs_mat_23_20_sp1),_mm256_maddubs_epi16(rhs_mat_2367_21_sp1, lhs_mat_23_21_sp1));
+
+                    __m256i iacc_mat_00_3_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_30_sp1, lhs_mat_01_30_sp1),_mm256_maddubs_epi16(rhs_mat_0145_31_sp1, lhs_mat_01_31_sp1));
+                    __m256i iacc_mat_01_3_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_30_sp1, lhs_mat_01_30_sp1),_mm256_maddubs_epi16(rhs_mat_2367_31_sp1, lhs_mat_01_31_sp1));
+
+                    __m256i iacc_mat_10_3_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_30_sp1, lhs_mat_23_30_sp1),_mm256_maddubs_epi16(rhs_mat_0145_31_sp1, lhs_mat_23_31_sp1));
+                    __m256i iacc_mat_11_3_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_30_sp1, lhs_mat_23_30_sp1),_mm256_maddubs_epi16(rhs_mat_2367_31_sp1, lhs_mat_23_31_sp1));
+
+                    __m256i iacc_mat_00_4_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_40_sp1, lhs_mat_01_40_sp1),_mm256_maddubs_epi16(rhs_mat_0145_41_sp1, lhs_mat_01_41_sp1));
+                    __m256i iacc_mat_01_4_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_40_sp1, lhs_mat_01_40_sp1),_mm256_maddubs_epi16(rhs_mat_2367_41_sp1, lhs_mat_01_41_sp1));
+
+                    __m256i iacc_mat_10_4_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_40_sp1, lhs_mat_23_40_sp1),_mm256_maddubs_epi16(rhs_mat_0145_41_sp1, lhs_mat_23_41_sp1));
+                    __m256i iacc_mat_11_4_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_40_sp1, lhs_mat_23_40_sp1),_mm256_maddubs_epi16(rhs_mat_2367_41_sp1, lhs_mat_23_41_sp1));
+
+                    __m256i iacc_mat_00_5_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_50_sp1, lhs_mat_01_50_sp1),_mm256_maddubs_epi16(rhs_mat_0145_51_sp1, lhs_mat_01_51_sp1));
+                    __m256i iacc_mat_01_5_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_50_sp1, lhs_mat_01_50_sp1),_mm256_maddubs_epi16(rhs_mat_2367_51_sp1, lhs_mat_01_51_sp1));
+
+                    __m256i iacc_mat_10_5_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_50_sp1, lhs_mat_23_50_sp1),_mm256_maddubs_epi16(rhs_mat_0145_51_sp1, lhs_mat_23_51_sp1));
+                    __m256i iacc_mat_11_5_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_50_sp1, lhs_mat_23_50_sp1),_mm256_maddubs_epi16(rhs_mat_2367_51_sp1, lhs_mat_23_51_sp1));
+
+                    __m256i iacc_mat_00_6_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_60_sp1, lhs_mat_01_60_sp1),_mm256_maddubs_epi16(rhs_mat_0145_61_sp1, lhs_mat_01_61_sp1));
+                    __m256i iacc_mat_01_6_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_60_sp1, lhs_mat_01_60_sp1),_mm256_maddubs_epi16(rhs_mat_2367_61_sp1, lhs_mat_01_61_sp1));
+
+                    __m256i iacc_mat_10_6_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_60_sp1, lhs_mat_23_60_sp1),_mm256_maddubs_epi16(rhs_mat_0145_61_sp1, lhs_mat_23_61_sp1));
+                    __m256i iacc_mat_11_6_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_60_sp1, lhs_mat_23_60_sp1),_mm256_maddubs_epi16(rhs_mat_2367_61_sp1, lhs_mat_23_61_sp1));
+
+                    __m256i iacc_mat_00_7_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_70_sp1, lhs_mat_01_70_sp1),_mm256_maddubs_epi16(rhs_mat_0145_71_sp1, lhs_mat_01_71_sp1));
+                    __m256i iacc_mat_01_7_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_70_sp1, lhs_mat_01_70_sp1),_mm256_maddubs_epi16(rhs_mat_2367_71_sp1, lhs_mat_01_71_sp1));
+
+                    __m256i iacc_mat_10_7_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_70_sp1, lhs_mat_23_70_sp1),_mm256_maddubs_epi16(rhs_mat_0145_71_sp1, lhs_mat_23_71_sp1));
+                    __m256i iacc_mat_11_7_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_70_sp1, lhs_mat_23_70_sp1),_mm256_maddubs_epi16(rhs_mat_2367_71_sp1, lhs_mat_23_71_sp1));
+
+
+                    __m256i iacc_mat_00_0_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_00_sp2, lhs_mat_01_00_sp2),_mm256_maddubs_epi16(rhs_mat_0145_01_sp2, lhs_mat_01_01_sp2));
+                    __m256i iacc_mat_01_0_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_00_sp2, lhs_mat_01_00_sp2),_mm256_maddubs_epi16(rhs_mat_2367_01_sp2, lhs_mat_01_01_sp2));
+
+                    __m256i iacc_mat_10_0_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_00_sp2, lhs_mat_23_00_sp2),_mm256_maddubs_epi16(rhs_mat_0145_01_sp2, lhs_mat_23_01_sp2));
+                    __m256i iacc_mat_11_0_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_00_sp2, lhs_mat_23_00_sp2),_mm256_maddubs_epi16(rhs_mat_2367_01_sp2, lhs_mat_23_01_sp2));
+
+                    __m256i iacc_mat_00_1_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_10_sp2, lhs_mat_01_10_sp2),_mm256_maddubs_epi16(rhs_mat_0145_11_sp2, lhs_mat_01_11_sp2));
+                    __m256i iacc_mat_01_1_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_10_sp2, lhs_mat_01_10_sp2),_mm256_maddubs_epi16(rhs_mat_2367_11_sp2, lhs_mat_01_11_sp2));
+
+                    __m256i iacc_mat_10_1_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_10_sp2, lhs_mat_23_10_sp2),_mm256_maddubs_epi16(rhs_mat_0145_11_sp2, lhs_mat_23_11_sp2));
+                    __m256i iacc_mat_11_1_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_10_sp2, lhs_mat_23_10_sp2),_mm256_maddubs_epi16(rhs_mat_2367_11_sp2, lhs_mat_23_11_sp2));
+
+                    __m256i iacc_mat_00_2_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_20_sp2, lhs_mat_01_20_sp2),_mm256_maddubs_epi16(rhs_mat_0145_21_sp2, lhs_mat_01_21_sp2));
+                    __m256i iacc_mat_01_2_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_20_sp2, lhs_mat_01_20_sp2),_mm256_maddubs_epi16(rhs_mat_2367_21_sp2, lhs_mat_01_21_sp2));
+
+                    __m256i iacc_mat_10_2_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_20_sp2, lhs_mat_23_20_sp2),_mm256_maddubs_epi16(rhs_mat_0145_21_sp2, lhs_mat_23_21_sp2));
+                    __m256i iacc_mat_11_2_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_20_sp2, lhs_mat_23_20_sp2),_mm256_maddubs_epi16(rhs_mat_2367_21_sp2, lhs_mat_23_21_sp2));
+
+                    __m256i iacc_mat_00_3_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_30_sp2, lhs_mat_01_30_sp2),_mm256_maddubs_epi16(rhs_mat_0145_31_sp2, lhs_mat_01_31_sp2));
+                    __m256i iacc_mat_01_3_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_30_sp2, lhs_mat_01_30_sp2),_mm256_maddubs_epi16(rhs_mat_2367_31_sp2, lhs_mat_01_31_sp2));
+
+                    __m256i iacc_mat_10_3_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_30_sp2, lhs_mat_23_30_sp2),_mm256_maddubs_epi16(rhs_mat_0145_31_sp2, lhs_mat_23_31_sp2));
+                    __m256i iacc_mat_11_3_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_30_sp2, lhs_mat_23_30_sp2),_mm256_maddubs_epi16(rhs_mat_2367_31_sp2, lhs_mat_23_31_sp2));
+
+                    __m256i iacc_mat_00_4_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_40_sp2, lhs_mat_01_40_sp2),_mm256_maddubs_epi16(rhs_mat_0145_41_sp2, lhs_mat_01_41_sp2));
+                    __m256i iacc_mat_01_4_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_40_sp2, lhs_mat_01_40_sp2),_mm256_maddubs_epi16(rhs_mat_2367_41_sp2, lhs_mat_01_41_sp2));
+
+                    __m256i iacc_mat_10_4_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_40_sp2, lhs_mat_23_40_sp2),_mm256_maddubs_epi16(rhs_mat_0145_41_sp2, lhs_mat_23_41_sp2));
+                    __m256i iacc_mat_11_4_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_40_sp2, lhs_mat_23_40_sp2),_mm256_maddubs_epi16(rhs_mat_2367_41_sp2, lhs_mat_23_41_sp2));
+
+                    __m256i iacc_mat_00_5_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_50_sp2, lhs_mat_01_50_sp2),_mm256_maddubs_epi16(rhs_mat_0145_51_sp2, lhs_mat_01_51_sp2));
+                    __m256i iacc_mat_01_5_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_50_sp2, lhs_mat_01_50_sp2),_mm256_maddubs_epi16(rhs_mat_2367_51_sp2, lhs_mat_01_51_sp2));
+
+                    __m256i iacc_mat_10_5_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_50_sp2, lhs_mat_23_50_sp2),_mm256_maddubs_epi16(rhs_mat_0145_51_sp2, lhs_mat_23_51_sp2));
+                    __m256i iacc_mat_11_5_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_50_sp2, lhs_mat_23_50_sp2),_mm256_maddubs_epi16(rhs_mat_2367_51_sp2, lhs_mat_23_51_sp2));
+
+                    __m256i iacc_mat_00_6_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_60_sp2, lhs_mat_01_60_sp2),_mm256_maddubs_epi16(rhs_mat_0145_61_sp2, lhs_mat_01_61_sp2));
+                    __m256i iacc_mat_01_6_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_60_sp2, lhs_mat_01_60_sp2),_mm256_maddubs_epi16(rhs_mat_2367_61_sp2, lhs_mat_01_61_sp2));
+
+                    __m256i iacc_mat_10_6_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_60_sp2, lhs_mat_23_60_sp2),_mm256_maddubs_epi16(rhs_mat_0145_61_sp2, lhs_mat_23_61_sp2));
+                    __m256i iacc_mat_11_6_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_60_sp2, lhs_mat_23_60_sp2),_mm256_maddubs_epi16(rhs_mat_2367_61_sp2, lhs_mat_23_61_sp2));
+
+                    __m256i iacc_mat_00_7_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_70_sp2, lhs_mat_01_70_sp2),_mm256_maddubs_epi16(rhs_mat_0145_71_sp2, lhs_mat_01_71_sp2));
+                    __m256i iacc_mat_01_7_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_70_sp2, lhs_mat_01_70_sp2),_mm256_maddubs_epi16(rhs_mat_2367_71_sp2, lhs_mat_01_71_sp2));
+
+                    __m256i iacc_mat_10_7_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_70_sp2, lhs_mat_23_70_sp2),_mm256_maddubs_epi16(rhs_mat_0145_71_sp2, lhs_mat_23_71_sp2));
+                    __m256i iacc_mat_11_7_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_70_sp2, lhs_mat_23_70_sp2),_mm256_maddubs_epi16(rhs_mat_2367_71_sp2, lhs_mat_23_71_sp2));
+
+                    // Combine results from both shuffle patterns for each output block.
+                    __m256i iacc_mat_00_0 = _mm256_add_epi16(iacc_mat_00_0_sp1, iacc_mat_00_0_sp2);
+                    __m256i iacc_mat_01_0 = _mm256_add_epi16(iacc_mat_01_0_sp1, iacc_mat_01_0_sp2);
+                    __m256i iacc_mat_10_0 = _mm256_add_epi16(iacc_mat_10_0_sp1, iacc_mat_10_0_sp2);
+                    __m256i iacc_mat_11_0 = _mm256_add_epi16(iacc_mat_11_0_sp1, iacc_mat_11_0_sp2);
+
+                    __m256i iacc_mat_00_1 = _mm256_add_epi16(iacc_mat_00_1_sp1, iacc_mat_00_1_sp2);
+                    __m256i iacc_mat_01_1 = _mm256_add_epi16(iacc_mat_01_1_sp1, iacc_mat_01_1_sp2);
+                    __m256i iacc_mat_10_1 = _mm256_add_epi16(iacc_mat_10_1_sp1, iacc_mat_10_1_sp2);
+                    __m256i iacc_mat_11_1 = _mm256_add_epi16(iacc_mat_11_1_sp1, iacc_mat_11_1_sp2);
+
+                    __m256i iacc_mat_00_2 = _mm256_add_epi16(iacc_mat_00_2_sp1, iacc_mat_00_2_sp2);
+                    __m256i iacc_mat_01_2 = _mm256_add_epi16(iacc_mat_01_2_sp1, iacc_mat_01_2_sp2);
+                    __m256i iacc_mat_10_2 = _mm256_add_epi16(iacc_mat_10_2_sp1, iacc_mat_10_2_sp2);
+                    __m256i iacc_mat_11_2 = _mm256_add_epi16(iacc_mat_11_2_sp1, iacc_mat_11_2_sp2);
+
+                    __m256i iacc_mat_00_3 = _mm256_add_epi16(iacc_mat_00_3_sp1, iacc_mat_00_3_sp2);
+                    __m256i iacc_mat_01_3 = _mm256_add_epi16(iacc_mat_01_3_sp1, iacc_mat_01_3_sp2);
+                    __m256i iacc_mat_10_3 = _mm256_add_epi16(iacc_mat_10_3_sp1, iacc_mat_10_3_sp2);
+                    __m256i iacc_mat_11_3 = _mm256_add_epi16(iacc_mat_11_3_sp1, iacc_mat_11_3_sp2);
+
+                    __m256i iacc_mat_00_4 = _mm256_add_epi16(iacc_mat_00_4_sp1, iacc_mat_00_4_sp2);
+                    __m256i iacc_mat_01_4 = _mm256_add_epi16(iacc_mat_01_4_sp1, iacc_mat_01_4_sp2);
+                    __m256i iacc_mat_10_4 = _mm256_add_epi16(iacc_mat_10_4_sp1, iacc_mat_10_4_sp2);
+                    __m256i iacc_mat_11_4 = _mm256_add_epi16(iacc_mat_11_4_sp1, iacc_mat_11_4_sp2);
+
+                    __m256i iacc_mat_00_5 = _mm256_add_epi16(iacc_mat_00_5_sp1, iacc_mat_00_5_sp2);
+                    __m256i iacc_mat_01_5 = _mm256_add_epi16(iacc_mat_01_5_sp1, iacc_mat_01_5_sp2);
+                    __m256i iacc_mat_10_5 = _mm256_add_epi16(iacc_mat_10_5_sp1, iacc_mat_10_5_sp2);
+                    __m256i iacc_mat_11_5 = _mm256_add_epi16(iacc_mat_11_5_sp1, iacc_mat_11_5_sp2);
+
+                    __m256i iacc_mat_00_6 = _mm256_add_epi16(iacc_mat_00_6_sp1, iacc_mat_00_6_sp2);
+                    __m256i iacc_mat_01_6 = _mm256_add_epi16(iacc_mat_01_6_sp1, iacc_mat_01_6_sp2);
+                    __m256i iacc_mat_10_6 = _mm256_add_epi16(iacc_mat_10_6_sp1, iacc_mat_10_6_sp2);
+                    __m256i iacc_mat_11_6 = _mm256_add_epi16(iacc_mat_11_6_sp1, iacc_mat_11_6_sp2);
+
+                    __m256i iacc_mat_00_7 = _mm256_add_epi16(iacc_mat_00_7_sp1, iacc_mat_00_7_sp2);
+                    __m256i iacc_mat_01_7 = _mm256_add_epi16(iacc_mat_01_7_sp1, iacc_mat_01_7_sp2);
+                    __m256i iacc_mat_10_7 = _mm256_add_epi16(iacc_mat_10_7_sp1, iacc_mat_10_7_sp2);
+                    __m256i iacc_mat_11_7 = _mm256_add_epi16(iacc_mat_11_7_sp1, iacc_mat_11_7_sp2);
+
+                    // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+                    iacc_mat_00_0 = _mm256_madd_epi16(iacc_mat_00_0, scale_0145_0);
+                    iacc_mat_01_0 = _mm256_madd_epi16(iacc_mat_01_0, scale_2367_0);
+                    iacc_mat_10_0 = _mm256_madd_epi16(iacc_mat_10_0, scale_0145_0);
+                    iacc_mat_11_0 = _mm256_madd_epi16(iacc_mat_11_0, scale_2367_0);
+
+                    iacc_mat_00_1 = _mm256_madd_epi16(iacc_mat_00_1, scale_0145_1);
+                    iacc_mat_01_1 = _mm256_madd_epi16(iacc_mat_01_1, scale_2367_1);
+                    iacc_mat_10_1 = _mm256_madd_epi16(iacc_mat_10_1, scale_0145_1);
+                    iacc_mat_11_1 = _mm256_madd_epi16(iacc_mat_11_1, scale_2367_1);
+
+                    iacc_mat_00_2 = _mm256_madd_epi16(iacc_mat_00_2, scale_0145_2);
+                    iacc_mat_01_2 = _mm256_madd_epi16(iacc_mat_01_2, scale_2367_2);
+                    iacc_mat_10_2 = _mm256_madd_epi16(iacc_mat_10_2, scale_0145_2);
+                    iacc_mat_11_2 = _mm256_madd_epi16(iacc_mat_11_2, scale_2367_2);
+
+                    iacc_mat_00_3 = _mm256_madd_epi16(iacc_mat_00_3, scale_0145_3);
+                    iacc_mat_01_3 = _mm256_madd_epi16(iacc_mat_01_3, scale_2367_3);
+                    iacc_mat_10_3 = _mm256_madd_epi16(iacc_mat_10_3, scale_0145_3);
+                    iacc_mat_11_3 = _mm256_madd_epi16(iacc_mat_11_3, scale_2367_3);
+
+                    iacc_mat_00_4 = _mm256_madd_epi16(iacc_mat_00_4, scale_0145_4);
+                    iacc_mat_01_4 = _mm256_madd_epi16(iacc_mat_01_4, scale_2367_4);
+                    iacc_mat_10_4 = _mm256_madd_epi16(iacc_mat_10_4, scale_0145_4);
+                    iacc_mat_11_4 = _mm256_madd_epi16(iacc_mat_11_4, scale_2367_4);
+
+                    iacc_mat_00_5 = _mm256_madd_epi16(iacc_mat_00_5, scale_0145_5);
+                    iacc_mat_01_5 = _mm256_madd_epi16(iacc_mat_01_5, scale_2367_5);
+                    iacc_mat_10_5 = _mm256_madd_epi16(iacc_mat_10_5, scale_0145_5);
+                    iacc_mat_11_5 = _mm256_madd_epi16(iacc_mat_11_5, scale_2367_5);
+
+                    iacc_mat_00_6 = _mm256_madd_epi16(iacc_mat_00_6, scale_0145_6);
+                    iacc_mat_01_6 = _mm256_madd_epi16(iacc_mat_01_6, scale_2367_6);
+                    iacc_mat_10_6 = _mm256_madd_epi16(iacc_mat_10_6, scale_0145_6);
+                    iacc_mat_11_6 = _mm256_madd_epi16(iacc_mat_11_6, scale_2367_6);
+
+                    iacc_mat_00_7 = _mm256_madd_epi16(iacc_mat_00_7, scale_0145_7);
+                    iacc_mat_01_7 = _mm256_madd_epi16(iacc_mat_01_7, scale_2367_7);
+                    iacc_mat_10_7 = _mm256_madd_epi16(iacc_mat_10_7, scale_0145_7);
+                    iacc_mat_11_7 = _mm256_madd_epi16(iacc_mat_11_7, scale_2367_7);
+
+                    __m256i iacc_mat_00 = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_mat_00_0, iacc_mat_00_1), _mm256_add_epi32(iacc_mat_00_2, iacc_mat_00_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_mat_00_4, iacc_mat_00_5), _mm256_add_epi32(iacc_mat_00_6, iacc_mat_00_7)));
+                    __m256i iacc_mat_01 = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_mat_01_0, iacc_mat_01_1), _mm256_add_epi32(iacc_mat_01_2, iacc_mat_01_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_mat_01_4, iacc_mat_01_5), _mm256_add_epi32(iacc_mat_01_6, iacc_mat_01_7)));
+                    __m256i iacc_mat_10 = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_mat_10_0, iacc_mat_10_1), _mm256_add_epi32(iacc_mat_10_2, iacc_mat_10_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_mat_10_4, iacc_mat_10_5), _mm256_add_epi32(iacc_mat_10_6, iacc_mat_10_7)));
+                    __m256i iacc_mat_11 = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_mat_11_0, iacc_mat_11_1), _mm256_add_epi32(iacc_mat_11_2, iacc_mat_11_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_mat_11_4, iacc_mat_11_5), _mm256_add_epi32(iacc_mat_11_6, iacc_mat_11_7)));
+
+                    // Straighten out to make 4 row vectors
+                    __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
+                    __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
+                    __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
+                    __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
+
+                    // Load the scale(d) values for all the 4 Q8_k blocks and repeat it across lanes
+                    const __m128 row_scale_f32_sse = _mm_load_ps(a_ptr[b].d);
+                    const __m256 row_scale_f32 = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
+
+                    // Multiply with appropiate scales and accumulate (for both d and dmin) below
+                    acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
+                    acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
+                    acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
+                    acc_rows[3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]);
+
+                    __m256i lhs_bsums_01_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_01_0123), lhs_raw_bsums_01_0123, 1);
+                    __m256i lhs_bsums_23_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_0123), lhs_raw_bsums_23_0123, 1);
+                    __m256i lhs_bsums_01_4567 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_01_4567), lhs_raw_bsums_01_4567, 1);
+                    __m256i lhs_bsums_23_4567 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_4567), lhs_raw_bsums_23_4567, 1);
+
+                    // Take two bsums from two Q8_Ks at a time and multiply with corresponding mins values from each Q2_K
+                    __m256i iacc_row_min_0_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 0), mins_01);
+                    __m256i iacc_row_min_1_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 170), mins_01);
+                    __m256i iacc_row_min_2_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 0), mins_01);
+                    __m256i iacc_row_min_3_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 170), mins_01);
+
+                    __m256i iacc_row_min_0_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 85), mins_23);
+                    __m256i iacc_row_min_1_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 255), mins_23);
+                    __m256i iacc_row_min_2_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 85), mins_23);
+                    __m256i iacc_row_min_3_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 255), mins_23);
+
+                    __m256i iacc_row_min_0_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 0), mins_45);
+                    __m256i iacc_row_min_1_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 170), mins_45);
+                    __m256i iacc_row_min_2_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 0), mins_45);
+                    __m256i iacc_row_min_3_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 170), mins_45);
+
+                    __m256i iacc_row_min_0_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 85), mins_67);
+                    __m256i iacc_row_min_1_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 255), mins_67);
+                    __m256i iacc_row_min_2_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 85), mins_67);
+                    __m256i iacc_row_min_3_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 255), mins_67);
+
+                    __m256i iacc_row_min_0 = _mm256_add_epi32(_mm256_add_epi32(iacc_row_min_0_01, iacc_row_min_0_23), _mm256_add_epi32(iacc_row_min_0_45,iacc_row_min_0_67));
+                    __m256i iacc_row_min_1 = _mm256_add_epi32(_mm256_add_epi32(iacc_row_min_1_01, iacc_row_min_1_23), _mm256_add_epi32(iacc_row_min_1_45,iacc_row_min_1_67));
+                    __m256i iacc_row_min_2 = _mm256_add_epi32(_mm256_add_epi32(iacc_row_min_2_01, iacc_row_min_2_23), _mm256_add_epi32(iacc_row_min_2_45,iacc_row_min_2_67));
+                    __m256i iacc_row_min_3 = _mm256_add_epi32(_mm256_add_epi32(iacc_row_min_3_01, iacc_row_min_3_23), _mm256_add_epi32(iacc_row_min_3_45,iacc_row_min_3_67));
+
+                    acc_min_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_0), _mm256_mul_ps(col_dmin_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_min_rows[0]);
+                    acc_min_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_1), _mm256_mul_ps(col_dmin_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_min_rows[1]);
+                    acc_min_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_2), _mm256_mul_ps(col_dmin_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_min_rows[2]);
+                    acc_min_rows[3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_3), _mm256_mul_ps(col_dmin_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_min_rows[3]);
+                }
+            }
+            // Store the accumulated values
+            for (int i = 0; i < 4; i++) {
+                _mm256_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm256_sub_ps(acc_rows[i], acc_min_rows[i]));
+            }
+        }
+    }
+#else
+
+    ggml_gemm_q2_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+
+
+#endif
+}
diff --git a/ggml/src/ggml-cpu/repack.cpp b/ggml/src/ggml-cpu/repack.cpp
index 74c1c029b94..2583aefae47 100644
--- a/ggml/src/ggml-cpu/repack.cpp
+++ b/ggml/src/ggml-cpu/repack.cpp
@@ -412,6 +412,82 @@ void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
     }
 }
 
+void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK_K;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+    float sumf[8];
+    float sum_minf[8];
+    int sumi1,sumi2,sumi3,sumi4;
+    int sumi;
+
+    const block_q8_K * a_ptr = (const block_q8_K *)vy;
+    for(int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q2_Kx8 * b_ptr = (const block_q2_Kx8 *) vx + (x * nb);
+        for (int j = 0; j < ncols_interleaved; j++) {
+            sumf[j] = 0.0;
+            sum_minf[j] = 0.0;
+        }
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (4 * blocklen)); k++) {
+                const uint8_t *scales_0 = b_ptr[l].scales + (k / 4) * 64 ;
+                const uint8_t *scales_1 = b_ptr[l].scales + (k / 4) * 64 + 16;
+                const uint8_t *scales_2 = b_ptr[l].scales + (k / 4) * 64 + 32;
+                const uint8_t *scales_3 = b_ptr[l].scales + (k / 4) * 64 + 48;
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi1 = 0;
+                    sumi2 = 0;
+                    sumi3 = 0;
+                    sumi4 = 0;
+                    sumi = 0;
+                    int offset = ((k / 2) % 2) + j * 2;
+                    for (int i = 0; i < blocklen; ++i){
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 3);
+                        const int v1 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 2 ) & 3);
+                        const int v2 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4 ) & 3);
+                        const int v3 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 6 ) & 3);
+                        sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i]);
+                        sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i + 32]);
+                        sumi3 = (v2 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i + 64]);
+                        sumi4 = (v3 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i + 96]);
+
+                        sumi1 = sumi1 * (scales_0[offset] & 0xF);
+                        sumi2 = sumi2 * (scales_1[offset] & 0xF);
+                        sumi3 = sumi3 * (scales_2[offset] & 0xF);
+                        sumi4 = sumi4 * (scales_3[offset] & 0xF);
+                        sumi += sumi1 + sumi2 + sumi3 + sumi4;
+                    }
+                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+                }
+            }
+            for(int sb = 0; sb < 8; sb++) {
+                const uint8_t *mins = b_ptr[l].scales + sb * 16;
+                for(int j = 0; j < ncols_interleaved; j++){
+                    sum_minf[j] += ((mins[j * 2] >> 4) * a_ptr[l].bsums[sb * 2] + (mins[(j * 2)+ 1] >> 4) * a_ptr[l].bsums[sb * 2 + 1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) {
+            s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j];
+        }
+    }
+}
+
 void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
@@ -711,6 +787,97 @@ void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
     }
 }
 
+void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK_K;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+    float sumf[4][8];
+    float sum_minf[4][8];
+    int sumi1, sumi2, sumi3, sumi4;
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q2_Kx8 * b_ptr = (const block_q2_Kx8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumf[m][j] = 0.0;
+                    sum_minf[m][j] = 0.0;
+                }
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (4 * blocklen)); k++) {
+
+                    const uint8_t *scales_0 = b_ptr[l].scales + (k / 4) * 64 ;
+                    const uint8_t *scales_1 = b_ptr[l].scales + (k / 4) * 64 + 16;
+                    const uint8_t *scales_2 = b_ptr[l].scales + (k / 4) * 64 + 32;
+                    const uint8_t *scales_3 = b_ptr[l].scales + (k / 4) * 64 + 48;
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi1 = 0;
+                            sumi2 = 0;
+                            sumi3 = 0;
+                            sumi4 = 0;
+                            sumi = 0;
+                            int offset = ((k / 2) % 2) + j * 2;
+                            for (int i = 0; i < blocklen; ++i){
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 3);
+                                const int v1 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 2 ) & 3);
+                                const int v2 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4 ) & 3);
+                                const int v3 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 6 ) & 3);
+                                sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 512 + (k % 4) * 4 * blocklen + m * blocklen + i]);
+                                sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 512  + (k % 4) * 4 * blocklen + m * blocklen + i + 128]);
+                                sumi3 = (v2 * a_ptr[l].qs[(k >> 2) * 512  + (k % 4) * 4 * blocklen + m * blocklen + i + 256]);
+                                sumi4 = (v3 * a_ptr[l].qs[(k >> 2) * 512  + (k % 4) * 4 * blocklen + m * blocklen + i + 384]);
+                                sumi1 = sumi1 * (scales_0[offset] & 0xF);
+                                sumi2 = sumi2 * (scales_1[offset] & 0xF);
+                                sumi3 = sumi3 * (scales_2[offset] & 0xF);
+                                sumi4 = sumi4 * (scales_3[offset] & 0xF);
+                                sumi += sumi1 + sumi2 + sumi3 + sumi4;
+                            }
+                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
+                        }
+                    }
+                }
+                for(int sb = 0; sb < 8; sb++) {
+                    const uint8_t *mins = b_ptr[l].scales + sb * 16;
+                    for(int m = 0; m < 4; m++) {
+                        const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) *  6);
+                        for(int j = 0; j < ncols_interleaved; j++) {
+                            int mins_prod = ((mins[j * 2] >> 4) * bsums[0] + (mins[(j * 2)+ 1] >> 4) * bsums[1]);
+                            sum_minf[m][j] += (mins_prod) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+                        }
+                    }
+                }
+            }
+
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
+                }
+            }
+        }
+    }
+}
+
+
 void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
@@ -914,6 +1081,50 @@ static block_q4_Kx8 make_block_q4_Kx8(block_q4_K * in, unsigned int blck_size_in
     return out;
 }
 
+static block_q2_Kx8 make_block_q2_Kx8(block_q2_K * in, unsigned int blck_size_interleave) {
+    block_q2_Kx8 out;
+
+    // Delta(scale) and dmin values of the eight Q2_K structures are copied onto the output interleaved structure
+    for (int i = 0; i < 8; i++) {
+        out.d[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d;
+    }
+
+    for (int i = 0; i < 8; i++) {
+        out.dmin[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.dmin;
+    }
+
+    const int end = QK_K * 2 / blck_size_interleave;
+
+    // Interleave Q2_K quants by taking 8 bytes at a time
+    for (int i = 0; i < end; ++i) {
+        int src_id = i % 8;
+        int src_offset = (i / 8) * blck_size_interleave;
+        int dst_offset = i * blck_size_interleave;
+
+        uint64_t elems;
+        memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
+        memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
+    }
+
+    // The below logic is designed so as to unpack and rearrange scales and mins values in Q2_K
+    // Currently the Q2_K structure has 16 scales and 16 mins packed in 16 bytes ( 4 bits for each value)
+    // The output Q2_Kx8 structure has 128 bytes for storing scales and mins
+    // Every 16 byte is packed such that it contains scales and mins for corresponding sub blocks from Q2_K structure
+    // For eg - First 16 bytes contains 16 scales and 16 mins - each of first and second sub blocks from different Q2_K structures
+
+    for(int i = 0; i < 128; i++){
+
+        // Index for selecting which q2k super block
+        int src1 = (i % 16) / 2;
+        // Index for selecting scale
+        int src2 = ((i / 16) * 2) + (i % 2);
+
+        out.scales[i] = in[src1].scales[src2];
+    }
+    return out;
+
+}
+
 static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
     GGML_ASSERT(t->type == GGML_TYPE_Q4_0);
     GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
@@ -975,6 +1186,37 @@ static int repack_q4_K_to_q4_K_8_bl(struct ggml_tensor * t, int interleave_block
     GGML_UNUSED(data_size);
 }
 
+static int repack_q2_K_to_q2_K_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_Q2_K);
+    GGML_ASSERT(interleave_block == 8);
+    constexpr int nrows_interleaved = 8;
+
+    block_q2_Kx8 * dst = (block_q2_Kx8*)t->data;
+    const block_q2_K * src = (const block_q2_K*) data;
+    block_q2_K dst_tmp[8];
+    int nrow = ggml_nrows(t);
+    int nblocks = t->ne[0] / QK_K;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q2_K));
+
+    if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i  = 0; i < nrows_interleaved; i++ ) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_q2_Kx8(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+
 static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
     GGML_ASSERT(t->type == GGML_TYPE_Q4_0);
     GGML_ASSERT(interleave_block == 8);
@@ -1095,6 +1337,10 @@ template <> int repack<block_q4_K, 8, 8>(struct ggml_tensor * t, const void * da
     return repack_q4_K_to_q4_K_8_bl(t, 8, data, data_size);
 }
 
+template <> int repack<block_q2_K, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q2_K_to_q2_K_8_bl(t, 8, data, data_size);
+}
+
 template <> int repack<block_iq4_nl, 4, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
     return repack_iq4_nl_to_iq4_nl_4_bl(t, 4, data, data_size);
 }
@@ -1124,6 +1370,10 @@ template <> void gemv<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t
     ggml_gemv_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemv<block_q2_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q2_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
 template <> void gemv<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
     ggml_gemv_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
 }
@@ -1148,6 +1398,10 @@ template <> void gemm<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t
     ggml_gemm_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemm<block_q2_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q2_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
 template <> void gemm<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
     ggml_gemm_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
 }
@@ -1421,6 +1675,9 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
     static const ggml::cpu::repack::tensor_traits<block_q4_0, 8, 8, GGML_TYPE_Q8_0> q4_0_8x8_q8_0;
     static const ggml::cpu::repack::tensor_traits<block_q4_K, 8, 8, GGML_TYPE_Q8_K> q4_K_8x8_q8_K;
 
+    // instance for Q2
+    static const ggml::cpu::repack::tensor_traits<block_q2_K, 8, 8, GGML_TYPE_Q8_K> q2_K_8x8_q8_K;
+
     // instance for IQ4
     static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0> iq4_nl_4x4_q8_0;
 
@@ -1446,6 +1703,12 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
                 return &q4_K_8x8_q8_K;
             }
         }
+    } else if (cur->type == GGML_TYPE_Q2_K) {
+        if (ggml_cpu_has_avx512()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &q2_K_8x8_q8_K;
+            }
+        }
     } else if (cur->type == GGML_TYPE_IQ4_NL) {
         if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
             if (cur->ne[1] % 4 == 0) {
diff --git a/ggml/src/ggml-cpu/repack.h b/ggml/src/ggml-cpu/repack.h
index 4421e5f8e70..cd322e74389 100644
--- a/ggml/src/ggml-cpu/repack.h
+++ b/ggml/src/ggml-cpu/repack.h
@@ -44,7 +44,14 @@ struct block_q4_Kx8 {
 };
 
 static_assert(sizeof(block_q4_Kx8) == sizeof(ggml_half) * 16 + K_SCALE_SIZE * 8 + QK_K * 4, "wrong q4_K block size/padding");
+struct block_q2_Kx8 {
+    ggml_half d[8];      // super-block scale for quantized scales
+    ggml_half dmin[8];   // super-block scale for quantized mins
+    uint8_t scales[128];  // scales and mins, quantized with 4 bits
+    uint8_t qs[512];    // 2--bit quants
+};
 
+static_assert(sizeof(block_q2_Kx8) == sizeof(ggml_half) * 16 + QK_K/2 + QK_K * 2, "wrong q2_K block size/padding");
 struct block_q8_Kx4 {
     float d[4];              // delta
     int8_t qs[QK_K * 4];     // quants
@@ -71,11 +78,13 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
 void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 
 // Native implementations
@@ -86,11 +95,13 @@ void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
 void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 
 #if defined(__cplusplus)

From a97a323023482cc44d2d44e3590f63f83e4040c8 Mon Sep 17 00:00:00 2001
From: lhez <quic_lih@quicinc.com>
Date: Fri, 1 Aug 2025 04:15:44 -0700
Subject: [PATCH 21/86] opencl: add f16 for `add`, `sub`, `mul`, `div`
 (llama/14984)

---
 ggml/src/ggml-opencl/ggml-opencl.cpp | 206 ++++++++++++++++++---------
 ggml/src/ggml-opencl/kernels/add.cl  |  73 ++++++++++
 ggml/src/ggml-opencl/kernels/div.cl  |  66 +++++++++
 ggml/src/ggml-opencl/kernels/mul.cl  |  73 ++++++++++
 ggml/src/ggml-opencl/kernels/sub.cl  |  66 +++++++++
 5 files changed, 414 insertions(+), 70 deletions(-)

diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
index 984d35a2ecf..150842f366a 100644
--- a/ggml/src/ggml-opencl/ggml-opencl.cpp
+++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
@@ -400,10 +400,10 @@ struct ggml_backend_opencl_context {
     cl_program program_mul_mm_f32_f32_l4_lm;
     cl_program program_mul_mm_f16_f32_l4_lm;
 
-    cl_kernel kernel_add, kernel_add_row;
-    cl_kernel kernel_mul, kernel_mul_row;
-    cl_kernel kernel_div, kernel_div_row;
-    cl_kernel kernel_sub, kernel_sub_row;
+    cl_kernel kernel_add, kernel_add_row, kernel_add_f16, kernel_add_row_f16;
+    cl_kernel kernel_mul, kernel_mul_row, kernel_mul_f16, kernel_mul_row_f16;
+    cl_kernel kernel_div, kernel_div_row, kernel_div_f16, kernel_div_row_f16;
+    cl_kernel kernel_sub, kernel_sub_row, kernel_sub_f16, kernel_sub_row_f16;
     cl_kernel kernel_scale;
     cl_kernel kernel_silu, kernel_silu_4;
     cl_kernel kernel_gelu, kernel_gelu_4;
@@ -674,8 +674,10 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         backend_ctx->program_add =
             build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
 
-        CL_CHECK((backend_ctx->kernel_add     = clCreateKernel(backend_ctx->program_add, "kernel_add", &err), err));
-        CL_CHECK((backend_ctx->kernel_add_row = clCreateKernel(backend_ctx->program_add, "kernel_add_row", &err), err));
+        CL_CHECK((backend_ctx->kernel_add         = clCreateKernel(backend_ctx->program_add, "kernel_add", &err), err));
+        CL_CHECK((backend_ctx->kernel_add_row     = clCreateKernel(backend_ctx->program_add, "kernel_add_row", &err), err));
+        CL_CHECK((backend_ctx->kernel_add_f16     = clCreateKernel(backend_ctx->program_add, "kernel_add_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_add_row_f16 = clCreateKernel(backend_ctx->program_add, "kernel_add_row_f16", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -1089,8 +1091,10 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         backend_ctx->program_mul =
             build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
 
-        CL_CHECK((backend_ctx->kernel_mul     = clCreateKernel(backend_ctx->program_mul, "kernel_mul", &err), err));
-        CL_CHECK((backend_ctx->kernel_mul_row = clCreateKernel(backend_ctx->program_mul, "kernel_mul_row", &err), err));
+        CL_CHECK((backend_ctx->kernel_mul         = clCreateKernel(backend_ctx->program_mul, "kernel_mul", &err), err));
+        CL_CHECK((backend_ctx->kernel_mul_row     = clCreateKernel(backend_ctx->program_mul, "kernel_mul_row", &err), err));
+        CL_CHECK((backend_ctx->kernel_mul_f16     = clCreateKernel(backend_ctx->program_mul, "kernel_mul_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_mul_row_f16 = clCreateKernel(backend_ctx->program_mul, "kernel_mul_row_f16", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -1288,11 +1292,16 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
 #else
         const std::string kernel_src = read_file("div.cl");
 #endif
+        std::string compile_opts = std::string("-cl-std=") + opencl_c_std +
+                               " -cl-mad-enable -cl-finite-math-only ";
+
         backend_ctx->program_div =
             build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
 
-        CL_CHECK((backend_ctx->kernel_div     = clCreateKernel(backend_ctx->program_div, "kernel_div", &err), err));
-        CL_CHECK((backend_ctx->kernel_div_row = clCreateKernel(backend_ctx->program_div, "kernel_div_row", &err), err));
+        CL_CHECK((backend_ctx->kernel_div         = clCreateKernel(backend_ctx->program_div, "kernel_div", &err), err));
+        CL_CHECK((backend_ctx->kernel_div_row     = clCreateKernel(backend_ctx->program_div, "kernel_div_row", &err), err));
+        CL_CHECK((backend_ctx->kernel_div_f16     = clCreateKernel(backend_ctx->program_div, "kernel_div_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_div_row_f16 = clCreateKernel(backend_ctx->program_div, "kernel_div_row_f16", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -1308,8 +1317,10 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         backend_ctx->program_sub =
             build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
 
-        CL_CHECK((backend_ctx->kernel_sub     = clCreateKernel(backend_ctx->program_sub, "kernel_sub", &err), err));
-        CL_CHECK((backend_ctx->kernel_sub_row = clCreateKernel(backend_ctx->program_sub, "kernel_sub_row", &err), err));
+        CL_CHECK((backend_ctx->kernel_sub         = clCreateKernel(backend_ctx->program_sub, "kernel_sub", &err), err));
+        CL_CHECK((backend_ctx->kernel_sub_row     = clCreateKernel(backend_ctx->program_sub, "kernel_sub_row", &err), err));
+        CL_CHECK((backend_ctx->kernel_sub_f16     = clCreateKernel(backend_ctx->program_sub, "kernel_sub_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_sub_row_f16 = clCreateKernel(backend_ctx->program_sub, "kernel_sub_row_f16", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -2447,12 +2458,15 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 default:
                     return false;
             }
-        case GGML_OP_ADD:
         case GGML_OP_SCALE:
+            return op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]);
+        case GGML_OP_ADD:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
         case GGML_OP_SUB:
-            return op->src[0]->type == GGML_TYPE_F32;
+            return (op->src[0]->type == op->src[1]->type) &&
+                   (op->src[0]->type == op->type) &&
+                   (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16);
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
                 case GGML_UNARY_OP_GELU:
@@ -3680,35 +3694,39 @@ static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const
     GGML_ASSERT(dst);
     GGML_ASSERT(dst->extra);
 
-    const int  ne00 = src0 ? src0->ne[0] : 0;
-    const int  ne01 = src0 ? src0->ne[1] : 0;
-    const int  ne02 = src0 ? src0->ne[2] : 0;
-    const int  ne03 = src0 ? src0->ne[3] : 0;
+    GGML_ASSERT(src0->type == src1->type);
+    GGML_ASSERT(src0->type == dst->type);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
 
-    const cl_ulong nb00 = src0 ? src0->nb[0] : 0;
-    const cl_ulong nb01 = src0 ? src0->nb[1] : 0;
-    const cl_ulong nb02 = src0 ? src0->nb[2] : 0;
-    const cl_ulong nb03 = src0 ? src0->nb[3] : 0;
+    const int  ne00 = src0->ne[0];
+    const int  ne01 = src0->ne[1];
+    const int  ne02 = src0->ne[2];
+    const int  ne03 = src0->ne[3];
 
-    const int  ne10 = src1 ? src1->ne[0] : 0;
-    const int  ne11 = src1 ? src1->ne[1] : 0;
-    const int  ne12 = src1 ? src1->ne[2] : 0;
-    const int  ne13 = src1 ? src1->ne[3] : 0; UNUSED(ne13);
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
 
-    const cl_ulong nb10 = src1 ? src1->nb[0] : 0;
-    const cl_ulong nb11 = src1 ? src1->nb[1] : 0;
-    const cl_ulong nb12 = src1 ? src1->nb[2] : 0;
-    const cl_ulong nb13 = src1 ? src1->nb[3] : 0; UNUSED(nb13);
+    const int  ne10 = src1->ne[0];
+    const int  ne11 = src1->ne[1];
+    const int  ne12 = src1->ne[2];
+    const int  ne13 = src1->ne[3]; UNUSED(ne13);
 
-    const int  ne0  = dst ? dst->ne[0] : 0;
-    const int  ne1  = dst ? dst->ne[1] : 0;
-    const int  ne2  = dst ? dst->ne[2] : 0;
-    const int  ne3  = dst ? dst->ne[3] : 0;
+    const cl_ulong nb10 = src1->nb[0];
+    const cl_ulong nb11 = src1->nb[1];
+    const cl_ulong nb12 = src1->nb[2];
+    const cl_ulong nb13 = src1->nb[3]; UNUSED(nb13);
 
-    const cl_ulong nb0  = dst ? dst->nb[0] : 0;
-    const cl_ulong nb1  = dst ? dst->nb[1] : 0;
-    const cl_ulong nb2  = dst ? dst->nb[2] : 0;
-    const cl_ulong nb3  = dst ? dst->nb[3] : 0;
+    const int  ne0  = dst->ne[0];
+    const int  ne1  = dst->ne[1];
+    const int  ne2  = dst->ne[2];
+    const int  ne3  = dst->ne[3];
+
+    const cl_ulong nb0  = dst->nb[0];
+    const cl_ulong nb1  = dst->nb[1];
+    const cl_ulong nb2  = dst->nb[2];
+    const cl_ulong nb3  = dst->nb[3];
 
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
@@ -3731,7 +3749,12 @@ static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const
 
         bcast_row = true;
         int ne = ne00 / 4;
-        kernel = backend_ctx->kernel_add_row;
+
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_add_row;
+        } else {
+            kernel = backend_ctx->kernel_add_row_f16;
+        }
 
         CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
         CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
@@ -3741,7 +3764,11 @@ static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const
         CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
         CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne));
     } else {
-        kernel = backend_ctx->kernel_add;
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_add;
+        } else {
+            kernel = backend_ctx->kernel_add_f16;
+        }
 
         CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
         CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
@@ -3803,35 +3830,39 @@ static void ggml_cl_mul(ggml_backend_t backend, const ggml_tensor * src0, const
     GGML_ASSERT(dst);
     GGML_ASSERT(dst->extra);
 
-    const int ne00 = src0 ? src0->ne[0] : 0;
-    const int ne01 = src0 ? src0->ne[1] : 0;
-    const int ne02 = src0 ? src0->ne[2] : 0;
-    const int ne03 = src0 ? src0->ne[3] : 0;
+    GGML_ASSERT(src0->type == src1->type);
+    GGML_ASSERT(src0->type == dst->type);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
 
-    const cl_ulong nb00 = src0 ? src0->nb[0] : 0;
-    const cl_ulong nb01 = src0 ? src0->nb[1] : 0;
-    const cl_ulong nb02 = src0 ? src0->nb[2] : 0;
-    const cl_ulong nb03 = src0 ? src0->nb[3] : 0;
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
 
-    const int ne10 = src1 ? src1->ne[0] : 0;
-    const int ne11 = src1 ? src1->ne[1] : 0;
-    const int ne12 = src1 ? src1->ne[2] : 0;
-    const int ne13 = src1 ? src1->ne[3] : 0; UNUSED(ne13);
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
 
-    const cl_ulong nb10 = src1 ? src1->nb[0] : 0;
-    const cl_ulong nb11 = src1 ? src1->nb[1] : 0;
-    const cl_ulong nb12 = src1 ? src1->nb[2] : 0;
-    const cl_ulong nb13 = src1 ? src1->nb[3] : 0; UNUSED(nb13);
+    const int ne10 = src1->ne[0];
+    const int ne11 = src1->ne[1];
+    const int ne12 = src1->ne[2];
+    const int ne13 = src1->ne[3]; UNUSED(ne13);
+
+    const cl_ulong nb10 = src1->nb[0];
+    const cl_ulong nb11 = src1->nb[1];
+    const cl_ulong nb12 = src1->nb[2];
+    const cl_ulong nb13 = src1->nb[3]; UNUSED(nb13);
 
-    const int ne0  = dst ? dst->ne[0] : 0;
-    const int ne1  = dst ? dst->ne[1] : 0;
-    const int ne2  = dst ? dst->ne[2] : 0;
-    const int ne3  = dst ? dst->ne[3] : 0;
+    const int ne0  = dst->ne[0];
+    const int ne1  = dst->ne[1];
+    const int ne2  = dst->ne[2];
+    const int ne3  = dst->ne[3];
 
-    const cl_ulong nb0  = dst ? dst->nb[0] : 0;
-    const cl_ulong nb1  = dst ? dst->nb[1] : 0;
-    const cl_ulong nb2  = dst ? dst->nb[2] : 0;
-    const cl_ulong nb3  = dst ? dst->nb[3] : 0;
+    const cl_ulong nb0  = dst->nb[0];
+    const cl_ulong nb1  = dst->nb[1];
+    const cl_ulong nb2  = dst->nb[2];
+    const cl_ulong nb3  = dst->nb[3];
 
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
@@ -3854,7 +3885,12 @@ static void ggml_cl_mul(ggml_backend_t backend, const ggml_tensor * src0, const
 
         bcast_row = true;
         int ne = ne00 / 4;
-        kernel = backend_ctx->kernel_mul_row;
+
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_mul_row;
+        } else {
+            kernel = backend_ctx->kernel_mul_row_f16;
+        }
 
         CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
         CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
@@ -3864,7 +3900,11 @@ static void ggml_cl_mul(ggml_backend_t backend, const ggml_tensor * src0, const
         CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
         CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne));
     } else {
-        kernel = backend_ctx->kernel_mul;
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_mul;
+        } else {
+            kernel = backend_ctx->kernel_mul_f16;
+        }
 
         CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
         CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
@@ -3926,6 +3966,10 @@ static void ggml_cl_div(ggml_backend_t backend, const ggml_tensor * src0, const
     GGML_ASSERT(dst);
     GGML_ASSERT(dst->extra);
 
+    GGML_ASSERT(src0->type == src1->type);
+    GGML_ASSERT(src0->type == dst->type);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
+
     const int ne00 = src0->ne[0];
     const int ne01 = src0->ne[1];
     const int ne02 = src0->ne[2];
@@ -3974,7 +4018,12 @@ static void ggml_cl_div(ggml_backend_t backend, const ggml_tensor * src0, const
 
         bcast_row = true;
         int ne = ne00 / 4;
-        kernel = backend_ctx->kernel_div_row;
+
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_div_row;
+        } else {
+            kernel = backend_ctx->kernel_div_row_f16;
+        }
 
         CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
         CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
@@ -3984,7 +4033,11 @@ static void ggml_cl_div(ggml_backend_t backend, const ggml_tensor * src0, const
         CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
         CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne));
     } else {
-        kernel = backend_ctx->kernel_div;
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_div;
+        } else {
+            kernel = backend_ctx->kernel_div_f16;
+        }
 
         CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
         CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
@@ -4034,6 +4087,10 @@ static void ggml_cl_sub(ggml_backend_t backend, const ggml_tensor * src0, const
     GGML_ASSERT(dst);
     GGML_ASSERT(dst->extra);
 
+    GGML_ASSERT(src0->type == src1->type);
+    GGML_ASSERT(src0->type == dst->type);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
+
     const int ne00 = src0->ne[0];
     const int ne01 = src0->ne[1];
     const int ne02 = src0->ne[2];
@@ -4082,7 +4139,12 @@ static void ggml_cl_sub(ggml_backend_t backend, const ggml_tensor * src0, const
 
         bcast_row = true;
         int ne = ne00 / 4;
-        kernel = backend_ctx->kernel_sub_row;
+
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_sub_row;
+        } else {
+            kernel = backend_ctx->kernel_sub_row_f16;
+        }
 
         CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
         CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
@@ -4092,7 +4154,11 @@ static void ggml_cl_sub(ggml_backend_t backend, const ggml_tensor * src0, const
         CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
         CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne));
     } else {
-        kernel = backend_ctx->kernel_sub;
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_sub;
+        } else {
+            kernel = backend_ctx->kernel_sub_f16;
+        }
 
         CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
         CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
diff --git a/ggml/src/ggml-opencl/kernels/add.cl b/ggml/src/ggml-opencl/kernels/add.cl
index f73f3c01343..8bc926c8893 100644
--- a/ggml/src/ggml-opencl/kernels/add.cl
+++ b/ggml/src/ggml-opencl/kernels/add.cl
@@ -81,3 +81,76 @@ kernel void kernel_add_row(
     uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
     dst[gid] = src0[gid] + src1[idx1];
 }
+
+kernel void kernel_add_f16(
+        global char * src0,
+        ulong  offset0,
+        global char * src1,
+        ulong  offset1,
+        global char * dst,
+        ulong  offsetd,
+        int   ne00,
+        int   ne01,
+        int   ne02,
+        int   ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int   ne10,
+        int   ne11,
+        int   ne12,
+        int   ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int   ne0,
+        int   ne1,
+        int   ne2,
+        int   ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst = dst + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03 % ne13;
+    int i12 = i02 % ne12;
+    int i11 = i01 % ne11;
+
+    global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    global char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i10 = i0 % ne10;
+        *((global half *)(dst_ptr + i0*nb0)) = *((global half *)(src0_ptr + i0*nb00)) + *((global half *)(src1_ptr + i10*nb10));
+    }
+}
+
+kernel void kernel_add_row_f16(
+        global half4 * src0,
+        ulong  offset0,
+        global half4 * src1,
+        ulong  offset1,
+        global half4 * dst,
+        ulong  offsetd,
+        int ne
+) {
+    src0 = (global half4*)((global char*)src0 + offset0);
+    src1 = (global half4*)((global char*)src1 + offset1);
+    dst = (global half4*)((global char*)dst + offsetd);
+
+    // This performs better than using %.
+    uint gid = get_global_id(0);
+    uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
+    dst[gid] = src0[gid] + src1[idx1];
+}
diff --git a/ggml/src/ggml-opencl/kernels/div.cl b/ggml/src/ggml-opencl/kernels/div.cl
index d453ad99be4..6d9b4ade9fe 100644
--- a/ggml/src/ggml-opencl/kernels/div.cl
+++ b/ggml/src/ggml-opencl/kernels/div.cl
@@ -70,3 +70,69 @@ kernel void kernel_div_row(
     uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
     dst[gid] = src0[gid] / src1[idx1];
 }
+
+kernel void kernel_div_f16(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03 % ne13;
+    int i12 = i02 % ne12;
+    int i11 = i01 % ne11;
+
+    global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    global char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i10 = i0 % ne10;
+        *((global half *)(dst_ptr + i0*nb0)) = *((global half *)(src0_ptr + i0*nb00)) / *((global half *)(src1_ptr + i10*nb10));
+    }
+}
+
+kernel void kernel_div_row_f16(
+        global half4 * src0,
+        ulong offset0,
+        global half4 * src1,
+        ulong offset1,
+        global half4 * dst,
+        ulong offsetd,
+        int ne
+) {
+    src0 = (global half4*)((global char*)src0 + offset0);
+    src1 = (global half4*)((global char*)src1 + offset1);
+    dst = (global half4*)((global char*)dst + offsetd);
+
+    // This performs better than using %.
+    uint gid = get_global_id(0);
+    uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
+    dst[gid] = src0[gid] / src1[idx1];
+}
diff --git a/ggml/src/ggml-opencl/kernels/mul.cl b/ggml/src/ggml-opencl/kernels/mul.cl
index 2a2b4eb70a1..b12a592165f 100644
--- a/ggml/src/ggml-opencl/kernels/mul.cl
+++ b/ggml/src/ggml-opencl/kernels/mul.cl
@@ -77,3 +77,76 @@ kernel void kernel_mul_row(
     uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
     dst[gid] = src0[gid] * src1[idx1];
 }
+
+kernel void kernel_mul_f16(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03 % ne13;
+    int i12 = i02 % ne12;
+    int i11 = i01 % ne11;
+
+    global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    global char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i10 = i0 % ne10;
+        *((global half *)(dst_ptr + i0*nb0)) = *((global half *)(src0_ptr + i0*nb00)) * *((global half *)(src1_ptr + i10*nb10));
+    }
+}
+
+kernel void kernel_mul_row_f16(
+        global half4 * src0,
+        ulong offset0,
+        global half4 * src1,
+        ulong offset1,
+        global half4 * dst,
+        ulong offsetd,
+        int ne
+) {
+    src0 = (global half4*)((global char*)src0 + offset0);
+    src1 = (global half4*)((global char*)src1 + offset1);
+    dst = (global half4*)((global char*)dst + offsetd);
+
+    // This performs better than using %.
+    uint gid = get_global_id(0);
+    uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
+    dst[gid] = src0[gid] * src1[idx1];
+}
diff --git a/ggml/src/ggml-opencl/kernels/sub.cl b/ggml/src/ggml-opencl/kernels/sub.cl
index 041e88ad3a0..423ed595ca8 100644
--- a/ggml/src/ggml-opencl/kernels/sub.cl
+++ b/ggml/src/ggml-opencl/kernels/sub.cl
@@ -70,3 +70,69 @@ kernel void kernel_sub_row(
     uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
     dst[gid] = src0[gid] - src1[idx1];
 }
+
+kernel void kernel_sub_f16(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03 % ne13;
+    int i12 = i02 % ne12;
+    int i11 = i01 % ne11;
+
+    global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    global char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i10 = i0 % ne10;
+        *((global half *)(dst_ptr + i0*nb0)) = *((global half *)(src0_ptr + i0*nb00)) - *((global half *)(src1_ptr + i10*nb10));
+    }
+}
+
+kernel void kernel_sub_row_f16(
+        global half4 * src0,
+        ulong offset0,
+        global half4 * src1,
+        ulong offset1,
+        global half4 * dst,
+        ulong offsetd,
+        int ne
+) {
+    src0 = (global half4*)((global char*)src0 + offset0);
+    src1 = (global half4*)((global char*)src1 + offset1);
+    dst = (global half4*)((global char*)dst + offsetd);
+
+    // This performs better than using %.
+    uint gid = get_global_id(0);
+    uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
+    dst[gid] = src0[gid] - src1[idx1];
+}

From 10f523cebfe9e96c205d48e2682f933907204c7f Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= <johannesg@5d6.de>
Date: Fri, 1 Aug 2025 20:47:32 +0200
Subject: [PATCH 22/86] CUDA: fix MMQ nwarps for AMD with warp_size==32
 (llama/15014)

---
 ggml/src/ggml-cuda/mmq.cuh | 18 +++++++-----------
 1 file changed, 7 insertions(+), 11 deletions(-)

diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh
index dd60529faa4..04a8d80e121 100644
--- a/ggml/src/ggml-cuda/mmq.cuh
+++ b/ggml/src/ggml-cuda/mmq.cuh
@@ -251,25 +251,21 @@ static constexpr __device__ int mmq_get_granularity_device(const int /*mmq_x*/)
 #endif // AMD_MFMA_AVAILABLE
 
 #if defined(GGML_USE_HIP)
-static int mmq_get_nwarps_host(const int cc) {
-    return amd_mfma_available(cc) ? 8 : 4;
+static int mmq_get_nwarps_host(const int cc, const int warp_size) {
+    return amd_mfma_available(cc) ? 8 : 256/warp_size;
 }
 #else
-static int mmq_get_nwarps_host(const int /*cc*/) {
-    return 8;
+static int mmq_get_nwarps_host(const int /*cc*/, const int warp_size) {
+    return 256/warp_size;
 }
 #endif // (GGML_USE_HIP)
 
 static constexpr __device__ int mmq_get_nwarps_device() {
-#if defined(GGML_USE_HIP)
 #if defined(AMD_MFMA_AVAILABLE)
     return 8;
 #else
-    return 4;
+    return 256/ggml_cuda_get_physical_warp_size();
 #endif // AMD_MFMA_AVAILABLE
-#else
-    return 8;
-#endif // defined(GGML_USE_HIP)
 }
 
 // ------------------------------------------------------------
@@ -3472,7 +3468,7 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
     const int cc = ggml_cuda_info().devices[id].cc;
     const int nsm = ggml_cuda_info().devices[id].nsm;
     const int warp_size = ggml_cuda_info().devices[id].warp_size;
-    const int nwarps = mmq_get_nwarps_host(cc);
+    const int nwarps = mmq_get_nwarps_host(cc, warp_size);
     const int mmq_y = get_mmq_y_host(cc);
 
     const dim3 block_dims(warp_size, nwarps, 1);
@@ -3559,7 +3555,7 @@ void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cuda
     const int    cc     = ggml_cuda_info().devices[id].cc;
     const size_t smpbo  = ggml_cuda_info().devices[id].smpbo;
     const int warp_size = ggml_cuda_info().devices[id].warp_size;
-    const int nwarps    = mmq_get_nwarps_host(cc);
+    const int nwarps    = mmq_get_nwarps_host(cc, warp_size);
 
     const int mmq_x_max = get_mmq_x_max_host(cc);
     const int mmq_y = get_mmq_y_host(cc);

From eb1dd358a1d227f8a2d7cd00d35595752db062f7 Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Sat, 2 Aug 2025 02:57:04 -0500
Subject: [PATCH 23/86] vulkan: optimizations for direct convolution
 (llama/14933)

* vulkan: optimizations for direct convolution

- Empirically choose a better tile size. Reducing BS_K/BS_NPQ helps fill
  the GPU. The new size should be amenable to using coopmat, too.
- Fix shmem bank conflicts. 16B padding should work with coopmat.
- Some explicit loop unrolling.
- Skip math/stores work for parts of the tile that are OOB.
- Apply fastdiv opt.
- Disable shuffles for NV.

* Three tiles sizes for CONV_2D, and a heuristic to choose

* reallow collectives for pre-Turing

* make SHMEM_PAD a spec constant

* fixes for intel perf - no shmem padding, placeholder shader core count

* shader variants with/without unrolling

* 0cc4m's fixes for AMD perf

Co-authored-by: 0cc4m <picard12@live.de>

---------

Co-authored-by: 0cc4m <picard12@live.de>
---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp          | 244 +++++++++++++-----
 .../ggml-vulkan/vulkan-shaders/conv2d_mm.comp |  88 ++++---
 .../vulkan-shaders/vulkan-shaders-gen.cpp     |   7 +-
 3 files changed, 233 insertions(+), 106 deletions(-)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 75b58c26fc1..2cd32fbb578 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -222,6 +222,7 @@ enum vk_device_architecture {
     AMD_RDNA2,
     AMD_RDNA3,
     INTEL_XE2,
+    NVIDIA_PRE_TURING,
 };
 
 // HSK x HSV
@@ -315,10 +316,33 @@ static vk_device_architecture get_device_architecture(const vk::PhysicalDevice&
             // https://www.intel.com/content/www/us/en/docs/oneapi/optimization-guide-gpu/2025-0/intel-xe-gpu-architecture.html
             return vk_device_architecture::INTEL_XE2;
         }
+    } else if (props.vendorID == VK_VENDOR_ID_NVIDIA) {
+        const std::vector<vk::ExtensionProperties> ext_props = device.enumerateDeviceExtensionProperties();
+
+        bool cooperative_matrix = false;
+
+        // Detect "pre-turing" based on lack of coopmat support.
+        for (const auto& properties : ext_props) {
+            if (strcmp("VK_KHR_cooperative_matrix", properties.extensionName) == 0) {
+                cooperative_matrix = true;
+                break;
+            }
+        }
+
+        if (!cooperative_matrix) {
+            return vk_device_architecture::NVIDIA_PRE_TURING;
+        }
     }
     return vk_device_architecture::OTHER;
 }
 
+enum vk_conv_shapes {
+    CONV_SHAPE_128x128,
+    CONV_SHAPE_64x32,
+    CONV_SHAPE_32x256,
+    CONV_SHAPE_COUNT,
+};
+
 struct vk_device_struct {
     std::recursive_mutex mutex;
 
@@ -483,8 +507,8 @@ struct vk_device_struct {
     vk_pipeline pipeline_rwkv_wkv6_f32;
     vk_pipeline pipeline_rwkv_wkv7_f32;
     vk_pipeline pipeline_opt_step_adamw_f32;
-    vk_pipeline pipeline_conv2d_f32;
-    vk_pipeline pipeline_conv2d_f16_f32;
+    vk_pipeline pipeline_conv2d_f32[CONV_SHAPE_COUNT];
+    vk_pipeline pipeline_conv2d_f16_f32[CONV_SHAPE_COUNT];
     vk_pipeline pipeline_conv2d_dw_whcn_f32;
     vk_pipeline pipeline_conv2d_dw_cwhn_f32;
 
@@ -908,8 +932,22 @@ struct vk_op_conv2d_push_constants {
     uint32_t nb1;
     uint32_t nb2;
     uint32_t nb3;
+
+    // init_fastdiv_values constants for dividing by KW, KW*KH, OW, OW*OH
+    uint32_t KWmp;   uint32_t KWL;
+    uint32_t KWKHmp; uint32_t KWKHL;
+    uint32_t OWmp;   uint32_t OWL;
+    uint32_t OWOHmp; uint32_t OWOHL;
 };
 
+template <> void init_pushconst_fastdiv(vk_op_conv2d_push_constants &p) {
+    // Compute magic values to divide by KW, KW*KH, OW, OW*OH
+    init_fastdiv_values(p.KW,       p.KWmp,    p.KWL);
+    init_fastdiv_values(p.KW*p.KH,  p.KWKHmp,  p.KWKHL);
+    init_fastdiv_values(p.OW,       p.OWmp,    p.OWL);
+    init_fastdiv_values(p.OW*p.OH,  p.OWOHmp,  p.OWOHL);
+}
+
 struct vk_op_conv2d_dw_push_constants {
     uint32_t ne;
     uint32_t batches;
@@ -3048,48 +3086,89 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_opt_step_adamw_f32, "opt_step_adamw_f32", opt_step_adamw_f32_len, opt_step_adamw_f32_data, "main", 5, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
 
     // conv2d
-    uint32_t conv2d_WG_SIZE  = 256;
-    uint32_t conv2d_BS_K     = 128;
-    uint32_t conv2d_BS_CRS   = 16;
-    uint32_t use_collectives = 0;  // Enables subgroup ops for preventing the re-calculation of indices.
-    if (device->subgroup_shuffle &&
-        device->vendor_id != VK_VENDOR_ID_INTEL) {  // Do not enable collectives on Intel, see PR 14316
-        use_collectives = 1;
-        conv2d_BS_CRS   = std::min(
-            device->subgroup_size,
-            conv2d_BS_CRS);  // CRS block size should be capped at sugroup size for correctness when shuffle is used.
-    }
-    uint32_t conv2d_BS_NPQ = 128;
-    uint32_t conv2d_TS_K   = 8;
-    uint32_t conv2d_shmem_req =
-        (conv2d_BS_K * (conv2d_BS_CRS + 1) + conv2d_BS_CRS * (conv2d_BS_NPQ + 1)) * sizeof(float);
-    if (device->properties.limits.maxComputeSharedMemorySize < conv2d_shmem_req) {
-        conv2d_BS_CRS = 8;
-        if (use_collectives) {
-            conv2d_BS_CRS = std::min(device->subgroup_size, conv2d_BS_CRS);
-        }
-    }
-
-    if (use_collectives) {
-        ggml_vk_create_pipeline(
-            device, device->pipeline_conv2d_f32, "conv2d_f32", conv2d_f32_len, conv2d_f32_data, "main", 3,
-            sizeof(vk_op_conv2d_push_constants), { conv2d_BS_K, conv2d_BS_NPQ, 1 },
-            { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives }, 1, true, true);
-        ggml_vk_create_pipeline(
-            device, device->pipeline_conv2d_f16_f32, "conv2d_f16_f32", conv2d_f16_f32_len, conv2d_f16_f32_data, "main", 3,
-            sizeof(vk_op_conv2d_push_constants), { conv2d_BS_K, conv2d_BS_NPQ, 1 },
-            { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives }, 1, true, true);
-    } else {
-        ggml_vk_create_pipeline(
-            device, device->pipeline_conv2d_f32, "conv2d_f32", conv2d_f32_len, conv2d_f32_data, "main", 3,
-            sizeof(vk_op_conv2d_push_constants), { conv2d_BS_K, conv2d_BS_NPQ, 1 },
-            { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives }, 1, true,
-            false);
-        ggml_vk_create_pipeline(
-            device, device->pipeline_conv2d_f16_f32, "conv2d_f16_f32", conv2d_f16_f32_len, conv2d_f16_f32_data, "main", 3,
-            sizeof(vk_op_conv2d_push_constants), { conv2d_BS_K, conv2d_BS_NPQ, 1 },
-            { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives }, 1, true,
-            false);
+    for (uint32_t s = 0; s < CONV_SHAPE_COUNT; ++s) {
+        uint32_t conv2d_WG_SIZE  = 256;
+        uint32_t conv2d_BS_K     = 128;
+        uint32_t conv2d_BS_CRS   = 16;
+        uint32_t use_collectives = 0;  // Enables subgroup ops for preventing the re-calculation of indices.
+        uint32_t conv2d_BS_NPQ = 128;
+        uint32_t conv2d_TS_K   = 8;
+        uint32_t conv2d_SHMEM_PAD = 4;
+        bool conv2d_UNROLL = true;
+
+        if (device->vendor_id == VK_VENDOR_ID_INTEL) {
+            conv2d_SHMEM_PAD = 0;
+            conv2d_UNROLL = false;
+        } else if (device->vendor_id == VK_VENDOR_ID_AMD) {
+            conv2d_SHMEM_PAD = device->architecture == vk_device_architecture::AMD_GCN ? 1 : 4;
+        }
+
+        switch (s) {
+        default:
+        case CONV_SHAPE_128x128:
+            conv2d_BS_K = 128;
+            conv2d_BS_NPQ = 128;
+            conv2d_BS_CRS = 16;
+            if (device->vendor_id == VK_VENDOR_ID_AMD && device->architecture != vk_device_architecture::AMD_GCN) {
+                conv2d_UNROLL = false;
+            }
+            break;
+        case CONV_SHAPE_64x32:
+            conv2d_BS_K = 64;
+            conv2d_BS_NPQ = 32;
+            conv2d_BS_CRS = 32;
+            conv2d_TS_K   = 4;
+            break;
+        case CONV_SHAPE_32x256:
+            conv2d_BS_K = 32;
+            conv2d_BS_NPQ = 256;
+            conv2d_BS_CRS = 16;
+            break;
+        }
+
+        // Use collectives on pre-Turing NVIDIA GPUs and GCN AMD cards, which had slower integer math.
+        bool allow_collectives_nv = device->vendor_id != VK_VENDOR_ID_NVIDIA ||
+                                    device->architecture == vk_device_architecture::NVIDIA_PRE_TURING;
+        bool allow_collectives_amd = device->vendor_id != VK_VENDOR_ID_AMD ||
+                                     device->architecture == vk_device_architecture::AMD_GCN;
+
+        if (device->subgroup_shuffle &&
+            device->vendor_id != VK_VENDOR_ID_INTEL &&   // Do not enable collectives on Intel, see PR 14316.
+            allow_collectives_nv &&
+            allow_collectives_amd) {
+            use_collectives = 1;
+            conv2d_BS_CRS   = std::min(
+                device->subgroup_size,
+                conv2d_BS_CRS);  // CRS block size should be capped at subgroup size for correctness when shuffle is used.
+        }
+
+        uint32_t conv2d_shmem_req =
+            (conv2d_BS_K * (conv2d_BS_CRS + conv2d_SHMEM_PAD) + conv2d_BS_CRS * (conv2d_BS_NPQ + conv2d_SHMEM_PAD)) * sizeof(float);
+        if (device->properties.limits.maxComputeSharedMemorySize < conv2d_shmem_req) {
+            conv2d_BS_CRS = 8;
+            if (use_collectives) {
+                conv2d_BS_CRS = std::min(device->subgroup_size, conv2d_BS_CRS);
+            }
+        }
+
+        std::array<uint32_t, 3> wg_denoms = { conv2d_BS_K, conv2d_BS_NPQ, 1 };
+        std::vector<uint32_t> spec_constants = { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives, conv2d_SHMEM_PAD };
+
+        if (conv2d_UNROLL) {
+            ggml_vk_create_pipeline(
+                device, device->pipeline_conv2d_f32[s], "conv2d_f32", conv2d_f32_unroll_len, conv2d_f32_unroll_data, "main", 3,
+                sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
+            ggml_vk_create_pipeline(
+                device, device->pipeline_conv2d_f16_f32[s], "conv2d_f16_f32", conv2d_f16_f32_unroll_len, conv2d_f16_f32_unroll_data, "main", 3,
+                sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
+        } else {
+            ggml_vk_create_pipeline(
+                device, device->pipeline_conv2d_f32[s], "conv2d_f32", conv2d_f32_len, conv2d_f32_data, "main", 3,
+                sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
+            ggml_vk_create_pipeline(
+                device, device->pipeline_conv2d_f16_f32[s], "conv2d_f16_f32", conv2d_f16_f32_len, conv2d_f16_f32_data, "main", 3,
+                sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
+        }
     }
 
     ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_whcn_f32, "conv2d_dw_whcn_f32", conv2d_dw_whcn_f32_len, conv2d_dw_whcn_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);
@@ -6641,6 +6720,34 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     }
 }
 
+static std::array<uint32_t, 3> ggml_vk_get_conv_elements(const ggml_tensor *dst) {
+    const ggml_tensor *src0 = dst->src[0];
+    const ggml_tensor *src1 = dst->src[1];
+
+    // src0 - kernel:   [KW, KH, Cin, Cout]
+    // src1 - input:    [W, H, Cin, N]
+    // dst - result:    [OW, OH, Cout, N]
+
+    // Copied from ggml.c: int64_t ggml_calc_conv_output_size(int64_t ins, int64_t ks, int s, int p, int d)
+    auto calc_conv_output_size = [](int64_t ins, int64_t ks, int s, int p, int d) -> int64_t {
+        return (ins + 2 * p - d * (ks - 1) - 1) / s + 1;
+    };
+    // parallelize in {OW/BS_K, OH/BS_NPQ, 1}
+    int64_t W    = src1->ne[0];
+    int64_t H    = src1->ne[1];
+    int64_t KW   = src0->ne[0];
+    int64_t KH   = src0->ne[1];
+    int64_t Cout = src0->ne[3];
+    int64_t N    = src1->ne[3];
+    int64_t OH   = calc_conv_output_size(H, KH, dst->op_params[1], dst->op_params[3], dst->op_params[5]);
+    int64_t OW   = calc_conv_output_size(W, KW, dst->op_params[0], dst->op_params[2], dst->op_params[4]);
+    int64_t NPQ  = N * OW * OH;
+
+    // Tile output matrix to (K/NB_K, NPQ/NB_NPQ, 1) workgroups
+    std::array<uint32_t, 3> elements = { static_cast<uint32_t>(Cout), static_cast<uint32_t>(NPQ), 1 };
+    return elements;
+}
+
 static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op) {
     switch (op) {
     case GGML_OP_GET_ROWS:
@@ -6970,10 +7077,30 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
     case GGML_OP_CONV_2D:
         if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 &&
             ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
+            auto elements = ggml_vk_get_conv_elements(dst);
+            vk_conv_shapes shape;
+
+            uint32_t tiles[CONV_SHAPE_COUNT];
+            for (uint32_t i = 0; i < CONV_SHAPE_COUNT; ++i) {
+                tiles[i] = CEIL_DIV(elements[0], ctx->device->pipeline_conv2d_f32[i]->wg_denoms[0]) * CEIL_DIV(elements[1], ctx->device->pipeline_conv2d_f32[i]->wg_denoms[1]);
+            }
+
+            // We can't query number of shader cores on Intel, use 32 as a placeholder
+            // so small convolutions will still choose a smaller tile.
+            const uint32_t shader_core_count = ctx->device->shader_core_count > 0 ? ctx->device->shader_core_count : 32;
+
+            if (elements[0] > 64 && tiles[CONV_SHAPE_128x128] >= shader_core_count * 2) {
+                shape = CONV_SHAPE_128x128;
+            } else if (elements[0] <= 32 && tiles[CONV_SHAPE_32x256] >= shader_core_count * 2) {
+                shape = CONV_SHAPE_32x256;
+            } else {
+                shape = CONV_SHAPE_64x32;
+            }
+
             if (src0->type == GGML_TYPE_F32) {
-                return ctx->device->pipeline_conv2d_f32;
+                return ctx->device->pipeline_conv2d_f32[shape];
             } else if (src0->type == GGML_TYPE_F16) {
-                return ctx->device->pipeline_conv2d_f16_f32;
+                return ctx->device->pipeline_conv2d_f16_f32[shape];
             }
         }
         return nullptr;
@@ -7301,29 +7428,8 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         } break;
     case GGML_OP_CONV_2D:
         {
-            // src0 - kernel:   [KW, KH, Cin, Cout]
-            // src1 - input:    [W, H, Cin, N]
-            // dst - result:    [OW, OH, Cout, N]
-
-            // Copied from ggml.c: int64_t ggml_calc_conv_output_size(int64_t ins, int64_t ks, int s, int p, int d)
-            auto calc_conv_output_size = [](int64_t ins, int64_t ks, int s, int p, int d) -> int64_t {
-                return (ins + 2 * p - d * (ks - 1) - 1) / s + 1;
-            };
-            // parallelize in {OW/BS_K, OH/BS_NPQ, 1}
-            int64_t W    = src1->ne[0];
-            int64_t H    = src1->ne[1];
-            int64_t KW   = src0->ne[0];
-            int64_t KH   = src0->ne[1];
-            int64_t Cout = src0->ne[3];
-            int64_t N    = src1->ne[3];
-            int64_t OH   = calc_conv_output_size(H, KH, dst->op_params[1], dst->op_params[3], dst->op_params[5]);
-            int64_t OW   = calc_conv_output_size(W, KW, dst->op_params[0], dst->op_params[2], dst->op_params[4]);
-            int64_t NPQ  = N * OW * OH;
-
-            // Tile output matrix to (K/NB_K, NPQ/NB_NPQ, 1) workgroups
-            elements = { static_cast<uint32_t>(Cout), static_cast<uint32_t>(NPQ), 1 };
-        }
-        break;
+            elements = ggml_vk_get_conv_elements(dst);
+        } break;
     case GGML_OP_ADD:
     case GGML_OP_SUB:
     case GGML_OP_DIV:
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp
index 481940a52b3..04a10c012f4 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp
@@ -1,14 +1,13 @@
 #version 450
 
+#extension GL_EXT_control_flow_attributes : enable
+
 #ifdef USE_COLLECTIVES
 #    extension GL_KHR_shader_subgroup_shuffle : enable
 #endif
 
 #include "types.comp"
 
-// Make spec constant
-#define SHMEM_PAD 0
-
 // shape notation: [dim(N), ..., dim(0)] -- stride(dim(j)) >= stride(dim(i)) if i > j
 layout(binding = 0) readonly buffer A {
     A_TYPE knl_data[];
@@ -56,6 +55,12 @@ layout(push_constant) uniform parameter {
     uint32_t nb1;
     uint32_t nb2;
     uint32_t nb3;
+
+    // fastdiv helper values
+    uint32_t KWmp;   uint32_t KWL;
+    uint32_t KWKHmp; uint32_t KWKHL;
+    uint32_t OWmp;   uint32_t OWL;
+    uint32_t OWOHmp; uint32_t OWOHL;
 }
 
 p;
@@ -68,6 +73,7 @@ layout(constant_id = 3) const uint BS_NPQ          = 128;
 // Thread-tile sizes
 layout(constant_id = 4) const uint TS_K            = 8;
 layout(constant_id = 5) const uint use_collectives = 1;
+layout(constant_id = 6) const uint SHMEM_PAD       = 4;
 
 uint32_t       tid     = gl_LocalInvocationID.x;
 const uint32_t WG_SIZE = gl_WorkGroupSize.x;
@@ -131,6 +137,14 @@ uint32_t       Br    = tid / BS_NPQ;
 uint32_t       Bc    = tid % BS_NPQ;
 const uint32_t BrpWg = WG_SIZE / BS_NPQ;
 
+// see init_fastdiv_values in ggml-vulkan.cpp
+uint fastdiv(uint n, uint mp, uint L) {
+    uint msbs, lsbs;
+    // msbs = mulhi(n, mp)
+    umulExtended(n, mp, msbs, lsbs);
+    return (msbs + n) >> L;
+}
+
 void main() {
     for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
         for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
@@ -151,9 +165,9 @@ void main() {
         uint32_t cached_KW_idx;
         if (use_collectives == 1) {
             cached_CRS_idx                = B_idx_CRS * BS_CRS + gl_SubgroupInvocationID;
-            cached_Cin_idx                = cached_CRS_idx / (p.KW * p.KH);
+            cached_Cin_idx                = fastdiv(cached_CRS_idx, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
             uint32_t cached_CRS_remainder = (cached_CRS_idx - cached_Cin_idx * p.KW * p.KH);
-            cached_KH_idx                 = cached_CRS_remainder / p.KW;
+            cached_KH_idx                 = fastdiv(cached_CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
             cached_KW_idx                 = cached_CRS_remainder - cached_KH_idx * p.KW;
 
             CRS_idx_a = subgroupShuffle(cached_CRS_idx, Ac);
@@ -162,16 +176,16 @@ void main() {
             KW_idx_a  = subgroupShuffle(cached_KW_idx, Ac);
         } else {
             CRS_idx_a              = B_idx_CRS * BS_CRS + Ac;  // Global CRS_idx_a (column index of A)
-            Cin_idx_a              = CRS_idx_a / (p.KW * p.KH);
+            Cin_idx_a              = fastdiv(CRS_idx_a, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
             uint32_t CRS_remainder = CRS_idx_a - Cin_idx_a * p.KW * p.KH;
-            KH_idx_a               = CRS_remainder / p.KW;
+            KH_idx_a               = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
             KW_idx_a               = CRS_remainder - KH_idx_a * p.KW;
         }
 #else
         CRS_idx_a     = B_idx_CRS * BS_CRS + Ac;  // Global CRS_idx_a (column index of A)
-        Cin_idx_a     = CRS_idx_a / (p.KW * p.KH);
+        Cin_idx_a     = fastdiv(CRS_idx_a, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH); / (p.KW * p.KH);
         CRS_remainder = CRS_idx_a - Cin_idx_a * p.KW * p.KH;
-        KH_idx_a      = CRS_remainder / p.KW;
+        KH_idx_a      = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
         KW_idx_a      = CRS_remainder - KH_idx_a * p.KW;
 #endif
 
@@ -188,13 +202,13 @@ void main() {
             Ash[B_ly * Ash_stride + B_lx] = val;
         }
         /* Load input to B_block: (BS_CRS x BS_NPQ) */
-        for (uint32_t r_offset = 0; r_offset < BS_CRS; r_offset += BrpWg) {
+        UNROLL for (uint32_t r_offset = 0; r_offset < BS_CRS; r_offset += BrpWg) {
             uint32_t B_ly          = r_offset + Br;             /* Row index of B block */
             uint32_t B_lx          = Bc;
             uint32_t NPQ_idx       = B_idx_NPQ * BS_NPQ + B_lx; /* Global NPQ index (column index of B) */
-            uint32_t N_idx         = NPQ_idx / (p.OH * p.OW);
+            uint32_t N_idx         = fastdiv(NPQ_idx, p.OWOHmp, p.OWOHL); // divide by p.OH * p.OW;
             uint32_t NPQ_remainder = NPQ_idx - N_idx * p.OH * p.OW;
-            uint32_t OH_idx        = NPQ_remainder / p.OW;
+            uint32_t OH_idx        = fastdiv(NPQ_remainder, p.OWmp, p.OWL); // divide by p.OW;
             uint32_t OW_idx        = NPQ_remainder - OH_idx * p.OW;
 
             uint32_t CRS_idx_b;
@@ -209,16 +223,16 @@ void main() {
                 KW_idx_b  = subgroupShuffle(cached_KW_idx, r_offset + Br);
             } else {
                 CRS_idx_b              = B_idx_CRS * BS_CRS + B_ly; /* Global CRS index (row index of B) */
-                Cin_idx_b              = CRS_idx_b / (p.KW * p.KH);
+                Cin_idx_b              = fastdiv(CRS_idx_b, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
                 uint32_t CRS_remainder = CRS_idx_b - Cin_idx_b * p.KW * p.KH;
-                KH_idx_b               = CRS_remainder / p.KW;
+                KH_idx_b               = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
                 KW_idx_b               = CRS_remainder - KH_idx_b * p.KW;
             }
 #else
             CRS_idx_b              = B_idx_CRS * BS_CRS + B_ly; /* Global CRS index (row index of B) */
-            Cin_idx_b              = CRS_idx_b / (p.KW * p.KH);
+            Cin_idx_b              = fastdiv(CRS_idx_b, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
             uint32_t CRS_remainder = CRS_idx_b - Cin_idx_b * p.KW * p.KH;
-            KH_idx_b               = CRS_remainder / p.KW;
+            KH_idx_b               = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
             KW_idx_b               = CRS_remainder - KH_idx_b * p.KW;
 #endif
 
@@ -233,32 +247,36 @@ void main() {
             Bsh[B_ly * Bsh_stride + B_lx] = val;
         }
         barrier();
-        for (uint32_t CRS_lidx = 0; CRS_lidx < BS_CRS; CRS_lidx++) {
-            for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
-                regA[T_ly] = Ash[(T_y * TS_K + T_ly) * Ash_stride + CRS_lidx];
-            }
-            for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
-                regB[T_lx] = Bsh[CRS_lidx * Bsh_stride + T_x * TS_NPQ + T_lx];
-            }
-            for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+        if (T_y * TS_K < K) {
+            UNROLL for (uint32_t CRS_lidx = 0; CRS_lidx < BS_CRS; CRS_lidx++) {
+                for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+                    regA[T_ly] = Ash[(T_y * TS_K + T_ly) * Ash_stride + CRS_lidx];
+                }
                 for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
-                    regC[T_ly][T_lx] = fma(regA[T_ly], regB[T_lx], regC[T_ly][T_lx]);
+                    regB[T_lx] = Bsh[CRS_lidx * Bsh_stride + T_x * TS_NPQ + T_lx];
+                }
+                for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+                    for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+                        regC[T_ly][T_lx] = fma(regA[T_ly], regB[T_lx], regC[T_ly][T_lx]);
+                    }
                 }
             }
         }
         barrier();
     }
     /* Save C* */
-    for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
-        for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
-            uint32_t K_idx   = B_idx_K * BS_K + T_y * TS_K + T_ly;
-            uint32_t NPQ_idx = B_idx_NPQ * BS_NPQ + T_x * TS_NPQ + T_lx;
-            uint32_t N_idx   = NPQ_idx / (p.OH * p.OW);
-            uint32_t OH_idx  = (NPQ_idx - N_idx * p.OH * p.OW) / p.OW;
-            uint32_t OW_idx  = NPQ_idx - N_idx * p.OH * p.OW - OH_idx * p.OW;
-            uint32_t dst_idx = OW_idx + OH_idx * p.nb1 + K_idx * p.nb2 + N_idx * p.nb3;
-            if (K_idx < K && NPQ_idx < NPQ) {
-                dst_data[dst_idx] = regC[T_ly][T_lx];
+    if (T_y * TS_K < K) {
+        for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+            for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+                uint32_t K_idx   = B_idx_K * BS_K + T_y * TS_K + T_ly;
+                uint32_t NPQ_idx = B_idx_NPQ * BS_NPQ + T_x * TS_NPQ + T_lx;
+                uint32_t N_idx   = fastdiv(NPQ_idx, p.OWOHmp, p.OWOHL); // divide by p.OH * p.OW;
+                uint32_t OH_idx  = fastdiv(NPQ_idx - N_idx * p.OH * p.OW, p.OWmp, p.OWL); // divide by p.OW;
+                uint32_t OW_idx  = NPQ_idx - N_idx * p.OH * p.OW - OH_idx * p.OW;
+                uint32_t dst_idx = OW_idx + OH_idx * p.nb1 + K_idx * p.nb2 + N_idx * p.nb3;
+                if (K_idx < K && NPQ_idx < NPQ) {
+                    dst_data[dst_idx] = regC[T_ly][T_lx];
+                }
             }
         }
     }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
index f9f0c95b8b2..b634e52d64d 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -655,8 +655,11 @@ void process_shaders() {
 
     string_to_spv("opt_step_adamw_f32", "opt_step_adamw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
 
-    string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}});
-    string_to_spv("conv2d_f16_f32", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}});
+    string_to_spv("conv2d_f32_unroll", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}});
+    string_to_spv("conv2d_f16_f32_unroll", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}});
+
+    string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", ""}});
+    string_to_spv("conv2d_f16_f32", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", ""}});
 
     string_to_spv("conv2d_dw_whcn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));
     string_to_spv("conv2d_dw_cwhn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}}));

From 3dd77561de1e0fea7b3d2423d5a41d9f53f32bb9 Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Sat, 2 Aug 2025 03:48:30 -0500
Subject: [PATCH 24/86] vulkan: Support ne[3]>1 in noncontig matrix-vector
 multiply (llama/15015)

---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp           | 15 ++++++++++-----
 .../vulkan-shaders/mul_mat_vec_nc.comp         | 18 +++++++++++-------
 2 files changed, 21 insertions(+), 12 deletions(-)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 2cd32fbb578..648cdd79b7d 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -2885,7 +2885,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
             ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_len,              mul_mat_vec_p021_f16_f32_data,              "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true);
         }
     }
-    ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 9 * sizeof(uint32_t), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 12 * sizeof(uint32_t), {1, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_group_norm_f32, "group_norm_f32", group_norm_f32_len, group_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
@@ -5821,7 +5821,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
     const uint64_t ne00 = src0->ne[0];
     const uint64_t ne01 = src0->ne[1];
     const uint64_t ne02 = src0->ne[2];
-    // const uint64_t ne03 = src0->ne[3];
+    const uint64_t ne03 = src0->ne[3];
 
     const uint64_t nb01 = src0->nb[1];
     const uint64_t nb02 = src0->nb[2];
@@ -5833,7 +5833,12 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
     const uint64_t ne12 = src1->ne[2];
     // const uint64_t ne13 = src1->ne[3];
 
+    const uint32_t nb03 = (uint32_t)(src0->nb[3] / sizeof(ggml_fp16_t));
+    const uint32_t nb13 = (uint32_t)(src1->nb[3] / sizeof(float));
+    const uint32_t nb23 = (uint32_t)(dst->nb[3] / sizeof(float));
+
     GGML_ASSERT(ne11 == 1);
+    GGML_ASSERT(src0->ne[3] == src1->ne[3]); // checked in supports_op
 
     ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
     ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
@@ -5849,7 +5854,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
         src1_uma = d_Qy != nullptr;
     }
 
-    const uint64_t d_ne = ne01 * ne11 * ne12;
+    const uint64_t d_ne = ne01 * ne11 * ne12 * ne03;
 
     const uint32_t row_stride_x = nb01 / sizeof(ggml_fp16_t);
     const uint32_t channel_stride_x = nb02 / sizeof(ggml_fp16_t);
@@ -5884,10 +5889,10 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
     const uint64_t d_shader_offset = d_buf_offset - d_buffer_offset;
 
     // compute
-    const std::array<uint32_t, 9> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, channel_stride_y, (uint32_t)(ne12 / ne02), (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
+    const std::array<uint32_t, 12> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, channel_stride_y, (uint32_t)(ne12 / ne02), (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)), nb03, nb13, nb23 };
     ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32,
-        { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
+        { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, pc, { (uint32_t)ne03, (uint32_t)ne01, (uint32_t)ne12 });
 }
 
 static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp
index bc633369f9b..638878d94ce 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp
@@ -26,6 +26,9 @@ layout (push_constant) uniform parameter
     uint ne12;
     uint b_offset;
     uint d_offset;
+    uint nb03;
+    uint nb13;
+    uint nb23;
 } p;
 
 shared FLOAT_TYPE tmp[BLOCK_SIZE];
@@ -34,6 +37,7 @@ void main() {
     const uint tid       = gl_LocalInvocationID.x;
     const uint row_x     = gl_GlobalInvocationID.y;
     const uint channel   = gl_GlobalInvocationID.z;
+    const uint i3        = gl_WorkGroupID.x;
     const uint channel_x = channel / p.channel_x_divisor;
     const uint channel_y = channel % p.ne12;
 
@@ -41,7 +45,7 @@ void main() {
     const uint nrows_dst = p.nrows_x;
     const uint row_dst   = row_x;
 
-    const uint idst = channel*nrows_dst + row_dst;
+    const uint idst = i3*p.nb23 + channel*nrows_dst + row_dst;
 
     FLOAT_TYPE temp = 0.0f;
 
@@ -58,8 +62,8 @@ void main() {
 
                 const uint row_y = col_x;
 
-                const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
-                const uint iy = channel_y*p.channel_stride_y + row_y;
+                const uint ix = i3*p.nb03 + channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
+                const uint iy = i3*p.nb13 + channel_y*p.channel_stride_y + row_y;
 
                 const vec4 av4 = vec4(data_a_v4[ix / 4]);
                 const vec4 bv4 = vec4(data_b_v4[iy / 4]);
@@ -74,8 +78,8 @@ void main() {
 
             const uint row_y = col_x;
 
-            const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
-            const uint iy = channel_y*p.channel_stride_y + row_y;
+            const uint ix = i3*p.nb03 + channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
+            const uint iy = i3*p.nb13 + channel_y*p.channel_stride_y + row_y;
 
             const vec4 av4 = vec4(data_a_v4[ix / 4]);
             const vec4 bv4 = vec4(data_b_v4[iy / 4]);
@@ -91,8 +95,8 @@ void main() {
 
             const uint row_y = col_x;
 
-            const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
-            const uint iy = channel_y*p.channel_stride_y + row_y;
+            const uint ix = i3*p.nb03 + channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
+            const uint iy = i3*p.nb13 + channel_y*p.channel_stride_y + row_y;
 
             const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]);
 

From d77de8405ba13cc1e44f9e969e171aa4dbce4cc8 Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Sat, 2 Aug 2025 04:21:37 -0500
Subject: [PATCH 25/86] vulkan: coopmat2 mul_mat optimizations (llama/14934)

- Increase tile size for k-quants, to match non-k-quants
- Choose more carefully between large and medium tiles, considering how it
  interacts with split_k
- Allow larger/non-power of two split_k, and make the splits a multiple of 256
- Use split_k==3 to when >1/2 and <=2/3 of the SMs would hae been used
---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp | 68 ++++++++++++++++++++--------
 1 file changed, 48 insertions(+), 20 deletions(-)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 648cdd79b7d..e095b26a484 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -2106,12 +2106,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
         s_mmq_wg_denoms = { 32,  64,  1 };
 
         // spec constants and tile sizes for quant matmul (Qi_K)
-        l_warptile_mmq_k = { 256, 64, 128, 64,  1 };
-        m_warptile_mmq_k = { 256, 32,  64, 64,  0 };
-        s_warptile_mmq_k = { 256, 32,  32, 128, 0 };
-        l_mmq_wg_denoms_k = { 64, 128, 1 };
-        m_mmq_wg_denoms_k = { 32,  64, 1 };
-        s_mmq_wg_denoms_k = { 32,  32, 1 };
+        l_warptile_mmq_k = { 256, 128, 256, 64, 1 };
+        m_warptile_mmq_k = { 256, 128, 128, 64, 1 };
+        s_warptile_mmq_k = { 256, 32,  64, 128, 0 };
+        l_mmq_wg_denoms_k = { 128, 256, 1 };
+        m_mmq_wg_denoms_k = { 128, 128, 1 };
+        s_mmq_wg_denoms_k = { 32,  64,  1 };
 
         // spec constants and tile sizes for quant matmul_id
         l_warptile_mmqid = { 256, 128, 128, 16, 0 };
@@ -5022,26 +5022,37 @@ static void ggml_vk_buffer_memset(vk_buffer& dst, size_t offset, uint32_t c, siz
     ggml_vk_queue_command_pools_cleanup(dst->device);
 }
 
-static uint32_t ggml_vk_guess_split_k(ggml_backend_vk_context * ctx, int m, int n, int k, const vk_pipeline& pipeline) {
+static uint32_t ggml_vk_guess_split_k(ggml_backend_vk_context * ctx, uint32_t m, uint32_t n, uint32_t k, const vk_pipeline& pipeline) {
     VK_LOG_DEBUG("ggml_vk_guess_split_k(" << m << ", " << n << ", " << k << ")");
 
     uint32_t split_k = 1;
-    if (ctx->device->shader_core_count != 0 && m >= (int)pipeline->wg_denoms[0] && n >= (int)pipeline->wg_denoms[1]) {
+    if (ctx->device->shader_core_count != 0 && m >= pipeline->wg_denoms[0] && n >= pipeline->wg_denoms[1]) {
         // If k is 'large' and the SMs will fill less than halfway, use split_k.
         uint32_t m_tiles = CEIL_DIV(m, pipeline->wg_denoms[0]);
         uint32_t n_tiles = CEIL_DIV(n, pipeline->wg_denoms[1]);
-        if (k >= 2048 && m_tiles * n_tiles < ctx->device->shader_core_count / 2) {
-            split_k = ctx->device->shader_core_count / (m_tiles * n_tiles);
-            // Clamp to 2 or 4
-            split_k = std::min(split_k, 4u);
-            if (split_k == 3) {
-                split_k = 2;
+
+        if (k >= 2048) {
+            if (m_tiles * n_tiles <= ctx->device->shader_core_count / 2) {
+                split_k = ctx->device->shader_core_count / (m_tiles * n_tiles);
+            } else if (m_tiles * n_tiles <= ctx->device->shader_core_count * 2 / 3) {
+                split_k = 3;
             }
-            if (ctx->device->coopmat2) {
-                // coopmat2 shader expects splits to be aligned to 256
-                while (split_k > 1 && ((k / split_k) % 256) != 0) {
-                    split_k /= 2;
+            // Cap the split at 8x. Unless k is huge this is a lot of overhead.
+            split_k = std::min(split_k, 8u);
+
+            // ggml_vk_matmul will align the splits to be a multiple of 256.
+            // If this rounded up size would cause the last split to be empty,
+            // then reduce the split count.
+            while (true) {
+                if (split_k == 1) {
+                    break;
+                }
+                uint32_t k_split = CEIL_DIV(k, split_k);
+                k_split = ROUNDUP_POW2(k_split, 256);
+                if (k_split * (split_k - 1) < k) {
+                    break;
                 }
+                split_k--;
             }
         }
     }
@@ -5053,9 +5064,22 @@ static vk_pipeline ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx,
     VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline(" << m << ", " << n << ", " << aligned << ", " << ggml_type_name(src0_type) << ", " << ggml_type_name(src1_type) << ")");
 
     if (ctx->device->coopmat2) {
+        const uint32_t shader_core_count = ctx->device->shader_core_count;
+        const uint32_t tiles_l = CEIL_DIV(m, mmp->a_l->wg_denoms[0]) * CEIL_DIV(n, mmp->a_l->wg_denoms[1]);
+        const uint32_t tiles_m = CEIL_DIV(m, mmp->a_m->wg_denoms[0]) * CEIL_DIV(n, mmp->a_m->wg_denoms[1]);
+
         // Use large shader when the N dimension is greater than the medium shader's tile size
         uint32_t crossover_large = mmp->m->wg_denoms[1];
-        if ((ctx->device->mul_mat_l[src0_type] && (n > crossover_large)) || (!ctx->device->mul_mat_m[src0_type] && !ctx->device->mul_mat_s[src0_type])) {
+
+        // Prefer large over medium if either:
+        // - medium or large tiles would overfill the GPU
+        // - large tiles with a split_k==3 fits in the GPU and medium tiles with split_k==2 does not
+        //   (medium with split_k==2 is probably better if it fits - more workgroups running and less split_k overhead)
+        bool prefer_large = tiles_m > shader_core_count || tiles_l > shader_core_count ||
+                            // split_k==3 with large tiles likely better than medium tiles with no split_k.
+                            (tiles_l <= shader_core_count / 3 && tiles_m > shader_core_count / 2);
+
+        if ((ctx->device->mul_mat_l[src0_type] && (n > crossover_large && prefer_large)) || (!ctx->device->mul_mat_m[src0_type] && !ctx->device->mul_mat_s[src0_type])) {
             return aligned ? mmp->a_l : mmp->l;
         }
         // Use medium shader when the N dimension is greater than the small shader's tile size
@@ -5099,7 +5123,11 @@ static void ggml_vk_matmul(
 
     GGML_ASSERT(batch_stride_d == m * n);
 
-    const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, CEIL_DIV(k, split_k), ne02, ne12, broadcast2, broadcast3, padded_n };
+    // Round the split size up to a multiple of 256 (k-quant alignment)
+    uint32_t k_split = CEIL_DIV(k, split_k);
+    k_split = ROUNDUP_POW2(k_split, 256);
+
+    const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, k_split, ne02, ne12, broadcast2, broadcast3, padded_n };
     // Make sure enough workgroups get assigned for split k to work
     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, split_k_buffer }, pc1, { (CEIL_DIV(m, pipeline->wg_denoms[0]) * pipeline->wg_denoms[0]) * split_k, n, batch });
     ggml_vk_sync_buffers(subctx);

From 2fd457c022bb7c40e154e4029403232526522e55 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Sat, 2 Aug 2025 17:13:05 +0300
Subject: [PATCH 26/86] cuda, sycl : fix batched gemm when ne02 == 1 && ne03 >
 1 (llama/15038)

* cuda, sycl : fix batched gemm when ne02 == 1 && ne03 > 1

ggml-ci

* cont : fix cont types

ggml-ci

* cont : adopt variable names and comment from the other branch
---
 ggml/src/ggml-cuda/ggml-cuda.cu  | 17 +++++++++++++----
 ggml/src/ggml-sycl/ggml-sycl.cpp | 15 ++++++++++++---
 2 files changed, 25 insertions(+), 7 deletions(-)

diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index 51792794673..8885fb7fbdd 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -1852,6 +1852,9 @@ static void ggml_cuda_mul_mat_batched_cublas_impl(ggml_backend_cuda_context & ct
     ggml_cuda_pool_alloc<cuda_t> src0_alloc(ctx.pool());
     ggml_cuda_pool_alloc<cuda_t> src1_alloc(ctx.pool());
 
+    bool is_src0_cont_2 = ggml_is_contiguous_2(src0);
+    bool is_src1_cont_2 = ggml_is_contiguous_2(src1);
+
     // Handle src0
     src0_ptr = (const cuda_t *) src0->data;
 
@@ -1870,6 +1873,8 @@ static void ggml_cuda_mul_mat_batched_cublas_impl(ggml_backend_cuda_context & ct
         s11 = ne10;
         s12 = ne11*s11;
         s13 = ne12*s12;
+
+        is_src1_cont_2 = true;
     }
 
     // Setup destination buffer
@@ -1918,15 +1923,19 @@ static void ggml_cuda_mul_mat_batched_cublas_impl(ggml_backend_cuda_context & ct
     const int64_t r2 = ne12/ne02;
     const int64_t r3 = ne13/ne03;
 
-    if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
+    if (r2 == 1 && r3 == 1 && is_src0_cont_2 && is_src1_cont_2) {
+        // with a [0, 2, 1, 3] perm. and ne02==1 the matrix strides need to be determined from dim 3:
+        const int64_t sma = ne02 == 1 ? nb03/nb00 : nb02/nb00;
+        const int64_t smb = ne12 == 1 ? s13       : s12;
+
         // there is no broadcast and src0, src1 are contiguous across dims 2, 3
         // use cublasGemmStridedBatchedEx
         CUBLAS_CHECK(
         cublasGemmStridedBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
                 ne01, ne11, ne10,
-                alpha, src0_ptr, cu_data_type_a, nb01/nb00, nb02/nb00, // strideA
-                       src1_ptr, cu_data_type_b, s11,       s12,       // strideB
-                beta,     dst_t, cu_data_type,   ne0,       ne1*ne0,   // strideC
+                alpha, src0_ptr, cu_data_type_a, nb01/nb00, sma,     // strideA
+                       src1_ptr, cu_data_type_b, s11,       smb,     // strideB
+                beta,     dst_t, cu_data_type,   ne0,       ne1*ne0, // strideC
                 ne12*ne13,
                 cu_compute_type,
                 CUBLAS_GEMM_DEFAULT_TENSOR_OP));
diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
index 2acdef98a6a..f68f1739a9f 100644
--- a/ggml/src/ggml-sycl/ggml-sycl.cpp
+++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
@@ -2688,6 +2688,9 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
     const size_t       type_size_src0 = ggml_type_size(src0->type);
     const size_t       type_size_src1 = ggml_type_size(src1->type);
 
+    bool is_src0_cont_2 = ggml_is_contiguous_2(src0);
+    bool is_src1_cont_2 = ggml_is_contiguous_2(src1);
+
     // SRC1 strides
     int64_t                          s11 = nb11 / type_size_src1;
     int64_t                          s12 = nb12 / type_size_src1;
@@ -2737,6 +2740,8 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
         s11      = ne10;
         s12      = ne11 * s11;
         s13      = ne12 * s12;
+
+        is_src1_cont_2 = true;
     }
 
     ggml_sycl_pool_alloc<sycl::half> dst_f16(ctx.pool());
@@ -2852,12 +2857,16 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
     else
 #endif
     {
-        if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
+        if (r2 == 1 && r3 == 1 && is_src0_cont_2 && is_src1_cont_2) {
+            // with a [0, 2, 1, 3] perm. and ne02==1 the matrix strides need to be determined from dim 3:
+            const int64_t sma = ne02 == 1 ? nb03/nb00 : nb02/nb00;
+            const int64_t smb = ne12 == 1 ? s13       : s12;
+
             // there is no broadcast and src0, src1 are contiguous across dims 2, 3
             SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(*queue, oneapi::math::transpose::trans,
                                                         oneapi::math::transpose::nontrans, ne01, ne11, ne10, alpha,
-                                                        src0_f16, dpct::library_data_t::real_half, nb01 / nb00, nb02 / nb00,
-                                                        src1_f16, dpct::library_data_t::real_half, s11, s12, beta, dst_ddf,
+                                                        src0_f16, dpct::library_data_t::real_half, nb01 / nb00, sma,
+                                                        src1_f16, dpct::library_data_t::real_half, s11, smb, beta, dst_ddf,
                                                         mkl_data_type, ne0, ne1 * ne0, ne12 * ne13, mkl_compute_type)));
         } else {
             const int ne23 = ne12 * ne13;

From f35b9e192b5902d7f23a71b649992717edad6296 Mon Sep 17 00:00:00 2001
From: leejet <leejet714@gmail.com>
Date: Sat, 2 Aug 2025 22:15:36 +0800
Subject: [PATCH 27/86] cuda: make im2col a little faster (llama/15025)

---
 ggml/src/ggml-cuda/im2col.cu | 80 ++++++++++++++++++++----------------
 1 file changed, 45 insertions(+), 35 deletions(-)

diff --git a/ggml/src/ggml-cuda/im2col.cu b/ggml/src/ggml-cuda/im2col.cu
index 5bb85b4807b..73b98133438 100644
--- a/ggml/src/ggml-cuda/im2col.cu
+++ b/ggml/src/ggml-cuda/im2col.cu
@@ -1,65 +1,75 @@
 #include "im2col.cuh"
 
+#define MIN(a, b) (a) < (b) ? (a) : (b)
+
+#define MAX_GRIDDIM_Z 65535
+
 template <typename T>
 static  __global__ void im2col_kernel(
-        const float * x, T * dst, int64_t batch_offset,
-        int64_t offset_delta, int64_t IC, int64_t IW, int64_t IH, int64_t OH, int64_t OW, int64_t KW, int64_t KH, int64_t pelements, int64_t CHW,
+        const float * x, T * dst,
+        int64_t IC, int64_t IW, int64_t IH, int64_t OH, int64_t OW, int64_t KW, int64_t KH,
+        int64_t IC_IH_IW, int64_t IH_IW, int64_t N_OH, int64_t KH_KW, int64_t IC_KH_KW,
         int s0, int s1, int p0, int p1, int d0, int d1) {
     const int64_t i = threadIdx.x + blockIdx.x * blockDim.x;
-    if (i >= pelements) {
+    if (i >= IC_KH_KW) {
         return;
     }
 
-    const int64_t  ksize = OW * KH;
-    const int64_t  kx = i / ksize;
-    const int64_t  kd = kx * ksize;
-    const int64_t  ky = (i - kd) / OW;
-    const int64_t  ix = i % OW;
+    const int64_t iic = i / (KH_KW);
+    const int64_t rem = i - iic * KH_KW;
+    const int64_t ikh = rem / KW;
+    const int64_t ikw = rem - ikh * KW;
 
-    const int64_t  oh = blockIdx.y;
-    const int64_t  batch = blockIdx.z / IC;
-    const int64_t  ic = blockIdx.z % IC;
+    const int64_t  iow = blockIdx.y;
+    for (int64_t iz = blockIdx.z; iz < N_OH; iz+=MAX_GRIDDIM_Z) {
+        const int64_t  in = iz / OH;
+        const int64_t  ioh = iz - in * OH;
 
-    const int64_t iiw = ix * s0 + kx * d0 - p0;
-    const int64_t iih = oh * s1 + ky * d1 - p1;
+        const int64_t iiw = iow * s0 + ikw * d0 - p0;
+        const int64_t iih = ioh * s1 + ikh * d1 - p1;
 
-    const int64_t offset_dst =
-        ((batch * OH + oh) * OW + ix) * CHW +
-        (ic * (KW * KH) + ky * KW + kx);
+        const int64_t offset_dst =
+            ((in * OH + ioh) * OW + iow) * IC_KH_KW + iic * KH_KW + ikh * KW + ikw;
 
-    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
-        dst[offset_dst] = 0.0f;
-    } else {
-        const int64_t offset_src = ic * offset_delta + batch * batch_offset;
-        dst[offset_dst] = x[offset_src + iih * IW + iiw];
+        if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
+            dst[offset_dst] = 0.0f;
+        } else {
+            const int64_t offset_src = iic * IC_IH_IW + in * IH_IW;
+            dst[offset_dst] = x[offset_src + iih * IW + iiw];
+        }
     }
 }
 
+// im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW]
 template <typename T>
 static void im2col_cuda(const float * x, T* dst,
     int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW, int64_t KH, int64_t IC,
-    int64_t batch, int64_t batch_offset, int64_t offset_delta,
+    int64_t N, int64_t IC_IH_IW, int64_t IH_IW,
     int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) {
-    const int parallel_elements = OW * KW * KH;
-    const int num_blocks = (parallel_elements + CUDA_IM2COL_BLOCK_SIZE - 1) / CUDA_IM2COL_BLOCK_SIZE;
-    dim3 block_nums(num_blocks, OH, batch * IC);
-    im2col_kernel<<<block_nums, CUDA_IM2COL_BLOCK_SIZE, 0, stream>>>(x, dst, batch_offset, offset_delta, IC, IW, IH, OH, OW, KW, KH, parallel_elements, (IC * KH * KW), s0, s1, p0, p1, d0, d1);
+    const int64_t IC_KH_KW = IC * KH * KW;
+    const int64_t num_blocks = (IC_KH_KW + CUDA_IM2COL_BLOCK_SIZE - 1) / CUDA_IM2COL_BLOCK_SIZE;
+    const int64_t N_OH = N * OH;
+    const int64_t KH_KW = KW*KH;
+    dim3 block_nums(num_blocks, OW, MIN(N_OH, MAX_GRIDDIM_Z));
+    im2col_kernel<<<block_nums, MIN(IC_KH_KW, CUDA_IM2COL_BLOCK_SIZE) , 0, stream>>>(x, dst, IC, IW, IH, OH, OW, KW, KH,
+                                                                                     IC_IH_IW, IH_IW, N_OH, KH_KW, IC_KH_KW,
+                                                                                     s0, s1, p0, p1, d0, d1);
 }
 
 static void im2col_cuda_f16(const float * x, half * dst,
     int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW, int64_t KH, int64_t IC,
-    int64_t batch, int64_t batch_offset, int64_t offset_delta,
+    int64_t N, int64_t IC_IH_IW, int64_t IH_IW,
     int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) {
 
-    im2col_cuda<half>(x, dst, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, offset_delta, s0, s1, p0, p1, d0, d1, stream);
+    im2col_cuda<half>(x, dst, IW, IH, OW, OH, KW, KH, IC, N, IC_IH_IW, IH_IW, s0, s1, p0, p1, d0, d1, stream);
 }
 
 static void im2col_cuda_f32(const float * x, float * dst,
     int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW, int64_t KH, int64_t IC,
-    int64_t batch, int64_t batch_offset, int64_t offset_delta,
+    int64_t N, int64_t IC_IH_IW, int64_t IH_IW,
     int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) {
 
-    im2col_cuda<float>(x, dst, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, offset_delta, s0, s1, p0, p1, d0, d1, stream);
+    im2col_cuda<float>(x, dst, IW, IH, OW, OH, KW, KH, IC, N, IC_IH_IW, IH_IW, s0, s1, p0, p1, d0, d1, stream);
 }
 
 void ggml_cuda_op_im2col(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -91,13 +101,13 @@ void ggml_cuda_op_im2col(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const int64_t OH = is_2D ? dst->ne[2] : 1;
     const int64_t OW =         dst->ne[1];
 
-    const size_t  delta_offset = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
-    const int64_t batch        = src1->ne[is_2D ? 3 : 2];
-    const size_t  batch_offset = src1->nb[is_2D ? 3 : 2] / 4; // nb is byte offset, src is type float32
+    const int64_t IC_IH_IW = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
+    const int64_t N        = src1->ne[is_2D ? 3 : 2];
+    const int64_t IH_IW    = src1->nb[is_2D ? 3 : 2] / 4; // nb is byte offset, src is type float32
 
     if(dst->type == GGML_TYPE_F16) {
-        im2col_cuda_f16(src1_d, (half *) dst_d, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, stream);
+        im2col_cuda_f16(src1_d, (half *) dst_d, IW, IH, OW, OH, KW, KH, IC, N, IC_IH_IW, IH_IW, s0, s1, p0, p1, d0, d1, stream);
     } else {
-        im2col_cuda_f32(src1_d, (float *) dst_d, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, stream);
+        im2col_cuda_f32(src1_d, (float *) dst_d, IW, IH, OW, OH, KW, KH, IC, N, IC_IH_IW, IH_IW, s0, s1, p0, p1, d0, d1, stream);
     }
 }

From 1cbee87e7fa3f991de68f28cec007f4396abafcc Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= <johannesg@5d6.de>
Date: Sat, 2 Aug 2025 16:37:08 +0200
Subject: [PATCH 28/86] CUDA: use mma FA kernel for gqa > 4 on RTX 4000
 (llama/15035)

---
 ggml/src/ggml-cuda/fattn.cu | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu
index a51136f6b8a..039c54e015e 100644
--- a/ggml/src/ggml-cuda/fattn.cu
+++ b/ggml/src/ggml-cuda/fattn.cu
@@ -315,8 +315,9 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
 
     const bool gqa_opt_applies = ((Q->ne[2] / K->ne[2]) % 2 == 0) && mask; // The mma-based kernels have GQA-specific optimizations
     const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16;
-    const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies &&
-        (Q->ne[3] > 1 || cc < GGML_CUDA_CC_ADA_LOVELACE) && !mma_needs_data_conversion;
+    const bool mma_faster_for_rtx4000 = Q->ne[3] > 1 || (Q->ne[2] > 4*K->ne[2] && K->ne[1] >= 8192);
+    const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies && !mma_needs_data_conversion &&
+        (cc < GGML_CUDA_CC_ADA_LOVELACE || mma_faster_for_rtx4000);
     const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % (2*warp_size) == 0;
     if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1) {
         if (prec == GGML_PREC_DEFAULT) {

From 3a868d8714822a56bcbcc1ba73c3bae3773f7fd9 Mon Sep 17 00:00:00 2001
From: lhez <lih@qti.qualcomm.com>
Date: Sat, 2 Aug 2025 10:51:18 -0700
Subject: [PATCH 29/86] opencl: fix adreno compiler detection logic
 (llama/15029)

---
 ggml/src/ggml-opencl/ggml-opencl.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
index 150842f366a..c9316eb7fd3 100644
--- a/ggml/src/ggml-opencl/ggml-opencl.cpp
+++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
@@ -2046,8 +2046,8 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
 
     backend_ctx->adreno_cl_compiler_version = get_adreno_cl_compiler_version(driver_version);
     backend_ctx->has_vector_subgroup_broadcast =
-        backend_ctx->adreno_cl_compiler_version.major >= 47 ||
-        backend_ctx->adreno_cl_compiler_version.major == 17;
+        (backend_ctx->adreno_cl_compiler_version.type == E031 && backend_ctx->adreno_cl_compiler_version.major >= 47) ||
+        (backend_ctx->adreno_cl_compiler_version.type == DX   && backend_ctx->adreno_cl_compiler_version.major >= 17);
     GGML_LOG_INFO("ggml_opencl: vector subgroup broadcast support: %s\n",
         backend_ctx->has_vector_subgroup_broadcast ? "true" : "false");
 

From 883a32e020c0e4f56e2a1f52ef553efe97d3e328 Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Sun, 3 Aug 2025 07:23:57 -0500
Subject: [PATCH 30/86] vulkan: Use coopmat2 for conv2d (llama/14982)

---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp          | 13 +++-
 .../ggml-vulkan/vulkan-shaders/conv2d_mm.comp | 62 ++++++++++++++++---
 .../vulkan-shaders/vulkan-shaders-gen.cpp     |  3 +
 3 files changed, 69 insertions(+), 9 deletions(-)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index e095b26a484..3682ee38047 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -3096,6 +3096,10 @@ static void ggml_vk_load_shaders(vk_device& device) {
         uint32_t conv2d_SHMEM_PAD = 4;
         bool conv2d_UNROLL = true;
 
+        if (device->coopmat2) {
+            conv2d_SHMEM_PAD = 8; // 8 float16_t
+        }
+
         if (device->vendor_id == VK_VENDOR_ID_INTEL) {
             conv2d_SHMEM_PAD = 0;
             conv2d_UNROLL = false;
@@ -3154,7 +3158,14 @@ static void ggml_vk_load_shaders(vk_device& device) {
         std::array<uint32_t, 3> wg_denoms = { conv2d_BS_K, conv2d_BS_NPQ, 1 };
         std::vector<uint32_t> spec_constants = { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives, conv2d_SHMEM_PAD };
 
-        if (conv2d_UNROLL) {
+        if (device->coopmat2) {
+            ggml_vk_create_pipeline(
+                device, device->pipeline_conv2d_f32[s], "conv2d_f32", conv2d_f32_cm2_len, conv2d_f32_cm2_data, "main", 3,
+                sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
+            ggml_vk_create_pipeline(
+                device, device->pipeline_conv2d_f16_f32[s], "conv2d_f16_f32", conv2d_f16_f32_cm2_len, conv2d_f16_f32_cm2_data, "main", 3,
+                sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
+        } else if (conv2d_UNROLL) {
             ggml_vk_create_pipeline(
                 device, device->pipeline_conv2d_f32[s], "conv2d_f32", conv2d_f32_unroll_len, conv2d_f32_unroll_data, "main", 3,
                 sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp
index 04a10c012f4..86bafba4a43 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp
@@ -1,6 +1,11 @@
 #version 450
 
 #extension GL_EXT_control_flow_attributes : enable
+#ifdef COOPMAT2
+#extension GL_NV_cooperative_matrix2 : enable
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#extension GL_KHR_memory_scope_semantics : enable
+#endif
 
 #ifdef USE_COLLECTIVES
 #    extension GL_KHR_shader_subgroup_shuffle : enable
@@ -91,6 +96,12 @@ uint32_t n_elems_out = K * NPQ;
 // Number of blocktiles per input
 uint32_t NB_CRS = splitWork(CRS, BS_CRS);
 
+#ifdef COOPMAT2
+#define SHMEM_TYPE float16_t
+#else
+#define SHMEM_TYPE float
+#endif
+
 const uint32_t Ash_stride = BS_CRS + SHMEM_PAD;
 const uint32_t Bsh_stride = BS_NPQ + SHMEM_PAD;
 
@@ -100,8 +111,8 @@ const uint32_t Bsh_numel = BS_CRS * BS_NPQ;
 const uint32_t Ash_len = BS_K * Ash_stride;
 const uint32_t Bsh_len = BS_CRS * Bsh_stride;
 
-shared float Ash[Ash_len];  // K x CRS
-shared float Bsh[Bsh_len];  // CRS x NPQ
+shared SHMEM_TYPE Ash[Ash_len];  // K x CRS
+shared SHMEM_TYPE Bsh[Bsh_len];  // CRS x NPQ
 
 // Threadtile sizes
 const uint32_t TS_NPQ = BS_K * BS_NPQ / WG_SIZE / TS_K;
@@ -110,10 +121,6 @@ const uint32_t TS_NPQ = BS_K * BS_NPQ / WG_SIZE / TS_K;
 const uint32_t NT_K   = BS_K / TS_K;
 const uint32_t NT_NPQ = BS_NPQ / TS_NPQ;
 
-float regA[TS_K];
-float regB[TS_NPQ];
-float regC[TS_K][TS_NPQ];
-
 /*
 Compute
 KxCRS @ CRSxNPQ = K x NPQ
@@ -145,12 +152,36 @@ uint fastdiv(uint n, uint mp, uint L) {
     return (msbs + n) >> L;
 }
 
+#ifdef COOPMAT2
+#define ACC_TYPE float16_t
+
+ACC_TYPE perElemOpStore(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem)
+{
+    uint32_t K_idx   = B_idx_K * BS_K + r;
+    uint32_t NPQ_idx = B_idx_NPQ * BS_NPQ + c;
+    uint32_t N_idx   = fastdiv(NPQ_idx, p.OWOHmp, p.OWOHL); // divide by p.OH * p.OW;
+    uint32_t OH_idx  = fastdiv(NPQ_idx - N_idx * p.OH * p.OW, p.OWmp, p.OWL); // divide by p.OW;
+    uint32_t OW_idx  = NPQ_idx - N_idx * p.OH * p.OW - OH_idx * p.OW;
+    uint32_t dst_idx = OW_idx + OH_idx * p.nb1 + K_idx * p.nb2 + N_idx * p.nb3;
+    if (K_idx < K && NPQ_idx < NPQ) {
+        dst_data[dst_idx] = D_TYPE(elem);
+    }
+    return elem;
+}
+#endif
+
 void main() {
+#ifdef COOPMAT2
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, BS_K, BS_NPQ, gl_MatrixUseAccumulator> matC;
+    matC = coopmat<ACC_TYPE, gl_ScopeWorkgroup, BS_K, BS_NPQ, gl_MatrixUseAccumulator>(0.0);
+#else
+    float regC[TS_K][TS_NPQ];
     for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
         for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
             regC[T_ly][T_lx] = 0.0;
         }
     }
+#endif
     /* Advance block in CRS dim */
     for (uint32_t B_idx_CRS = 0; B_idx_CRS < NB_CRS; B_idx_CRS++) {
         uint32_t CRS_idx_a;
@@ -199,7 +230,7 @@ void main() {
             if (K_idx >= K || CRS_idx_a >= CRS) {
                 val = 0.0;
             }
-            Ash[B_ly * Ash_stride + B_lx] = val;
+            Ash[B_ly * Ash_stride + B_lx] = SHMEM_TYPE(val);
         }
         /* Load input to B_block: (BS_CRS x BS_NPQ) */
         UNROLL for (uint32_t r_offset = 0; r_offset < BS_CRS; r_offset += BrpWg) {
@@ -244,11 +275,21 @@ void main() {
             if (CRS_idx_b >= CRS || NPQ_idx >= NPQ || H_idx < 0 || H_idx >= p.H || W_idx < 0 || W_idx >= p.W) {
                 val = 0.0;
             }
-            Bsh[B_ly * Bsh_stride + B_lx] = val;
+            Bsh[B_ly * Bsh_stride + B_lx] = SHMEM_TYPE(val);
         }
         barrier();
+#ifdef COOPMAT2
+        coopmat<float16_t, gl_ScopeWorkgroup, BS_K, BS_CRS, gl_MatrixUseA> matA;
+        coopmat<float16_t, gl_ScopeWorkgroup, BS_CRS, BS_NPQ, gl_MatrixUseB> matB;
+
+        coopMatLoad(matA, Ash, 0, Ash_stride, gl_CooperativeMatrixLayoutRowMajor);
+        coopMatLoad(matB, Bsh, 0, Bsh_stride, gl_CooperativeMatrixLayoutRowMajor);
+        matC = coopMatMulAdd(matA, matB, matC);
+#else
         if (T_y * TS_K < K) {
             UNROLL for (uint32_t CRS_lidx = 0; CRS_lidx < BS_CRS; CRS_lidx++) {
+                float regA[TS_K];
+                float regB[TS_NPQ];
                 for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
                     regA[T_ly] = Ash[(T_y * TS_K + T_ly) * Ash_stride + CRS_lidx];
                 }
@@ -262,9 +303,13 @@ void main() {
                 }
             }
         }
+#endif
         barrier();
     }
     /* Save C* */
+#ifdef COOPMAT2
+    coopMatPerElementNV(matC, matC, perElemOpStore);
+#else
     if (T_y * TS_K < K) {
         for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
             for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
@@ -280,4 +325,5 @@ void main() {
             }
         }
     }
+#endif
 }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
index b634e52d64d..83e4a7c723d 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -661,6 +661,9 @@ void process_shaders() {
     string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", ""}});
     string_to_spv("conv2d_f16_f32", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", ""}});
 
+    string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}, {"COOPMAT2", "1"}}, true, false, true);
+    string_to_spv("conv2d_f16_f32", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}, {"COOPMAT2", "1"}}, true, false, true);
+
     string_to_spv("conv2d_dw_whcn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));
     string_to_spv("conv2d_dw_cwhn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}}));
 

From 1071eb77c690d28f1fb21fa2a45d194920087856 Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Mon, 4 Aug 2025 00:09:19 -0500
Subject: [PATCH 31/86] vulkan: fix build when using glslang that does not
 support coopmat2 (llama/15062)

---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp                       | 7 ++++++-
 ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp | 2 ++
 2 files changed, 8 insertions(+), 1 deletion(-)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 3682ee38047..3c1ae084996 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -3096,9 +3096,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
         uint32_t conv2d_SHMEM_PAD = 4;
         bool conv2d_UNROLL = true;
 
+#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
         if (device->coopmat2) {
             conv2d_SHMEM_PAD = 8; // 8 float16_t
         }
+#endif
 
         if (device->vendor_id == VK_VENDOR_ID_INTEL) {
             conv2d_SHMEM_PAD = 0;
@@ -3158,6 +3160,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
         std::array<uint32_t, 3> wg_denoms = { conv2d_BS_K, conv2d_BS_NPQ, 1 };
         std::vector<uint32_t> spec_constants = { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives, conv2d_SHMEM_PAD };
 
+#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
         if (device->coopmat2) {
             ggml_vk_create_pipeline(
                 device, device->pipeline_conv2d_f32[s], "conv2d_f32", conv2d_f32_cm2_len, conv2d_f32_cm2_data, "main", 3,
@@ -3165,7 +3168,9 @@ static void ggml_vk_load_shaders(vk_device& device) {
             ggml_vk_create_pipeline(
                 device, device->pipeline_conv2d_f16_f32[s], "conv2d_f16_f32", conv2d_f16_f32_cm2_len, conv2d_f16_f32_cm2_data, "main", 3,
                 sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
-        } else if (conv2d_UNROLL) {
+        } else
+#endif
+        if (conv2d_UNROLL) {
             ggml_vk_create_pipeline(
                 device, device->pipeline_conv2d_f32[s], "conv2d_f32", conv2d_f32_unroll_len, conv2d_f32_unroll_data, "main", 3,
                 sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
index 83e4a7c723d..c6aa3ea4c79 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -661,8 +661,10 @@ void process_shaders() {
     string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", ""}});
     string_to_spv("conv2d_f16_f32", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", ""}});
 
+#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
     string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}, {"COOPMAT2", "1"}}, true, false, true);
     string_to_spv("conv2d_f16_f32", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}, {"COOPMAT2", "1"}}, true, false, true);
+#endif
 
     string_to_spv("conv2d_dw_whcn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));
     string_to_spv("conv2d_dw_cwhn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}}));

From c9386c1ec17cd91566cbc7bb83672e2b5b08249c Mon Sep 17 00:00:00 2001
From: Christian Kastner <ckk@kvr.at>
Date: Mon, 4 Aug 2025 21:29:14 +0200
Subject: [PATCH 32/86] cmake: Add GGML_BACKEND_DIR option (llama/15074)

* cmake: Add GGML_BACKEND_DIR option

This can be used by distributions to specify where to look for backends
when ggml is built with GGML_BACKEND_DL=ON.

* Fix phrasing
---
 ggml/CMakeLists.txt             |  5 +++--
 ggml/cmake/ggml-config.cmake.in |  2 +-
 ggml/src/CMakeLists.txt         | 13 ++++++++++++-
 ggml/src/ggml-backend-reg.cpp   |  3 +++
 4 files changed, 19 insertions(+), 4 deletions(-)

diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt
index 231250efce0..7c4666b7edb 100644
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -39,8 +39,9 @@ if (WIN32)
     set(CMAKE_SHARED_MODULE_PREFIX  "")
 endif()
 
-option(BUILD_SHARED_LIBS "ggml: build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
-option(GGML_BACKEND_DL   "ggml: build backends as dynamic libraries (requires BUILD_SHARED_LIBS)" OFF)
+option(BUILD_SHARED_LIBS           "ggml: build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
+option(GGML_BACKEND_DL             "ggml: build backends as dynamic libraries (requires BUILD_SHARED_LIBS)" OFF)
+set(GGML_BACKEND_DIR "" CACHE PATH "ggml: directory to load dynamic backends from (requires GGML_BACKEND_DL")
 
 #
 # option list
diff --git a/ggml/cmake/ggml-config.cmake.in b/ggml/cmake/ggml-config.cmake.in
index 2322c6cd9d0..65a75a2f1c8 100644
--- a/ggml/cmake/ggml-config.cmake.in
+++ b/ggml/cmake/ggml-config.cmake.in
@@ -106,7 +106,7 @@ if(NOT TARGET ggml::ggml)
 
     find_library(GGML_LIBRARY ggml
         REQUIRED
-        HINTS ${GGML_LIB_DIR}
+        HINTS ${GGML_LIB_DIR} ${GGML_BACKEND_DIR}
         NO_CMAKE_FIND_ROOT_PATH)
 
     add_library(ggml::ggml UNKNOWN IMPORTED)
diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
index 0425fd60a94..177fb282135 100644
--- a/ggml/src/CMakeLists.txt
+++ b/ggml/src/CMakeLists.txt
@@ -214,6 +214,13 @@ add_library(ggml
             ggml-backend-reg.cpp)
 add_library(ggml::ggml ALIAS ggml)
 
+if (GGML_BACKEND_DIR)
+    if (NOT GGML_BACKEND_DL)
+        message(FATAL_ERROR "GGML_BACKEND_DIR requires GGML_BACKEND_DL")
+    endif()
+    target_compile_definitions(ggml PUBLIC GGML_BACKEND_DIR="${GGML_BACKEND_DIR}")
+endif()
+
 target_link_libraries(ggml PUBLIC ggml-base)
 
 if (CMAKE_SYSTEM_NAME MATCHES "Linux")
@@ -227,7 +234,11 @@ function(ggml_add_backend_library backend)
         set_target_properties(${backend} PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
         target_compile_definitions(${backend} PRIVATE GGML_BACKEND_DL)
         add_dependencies(ggml ${backend})
-        install(TARGETS ${backend} LIBRARY DESTINATION ${CMAKE_INSTALL_BINDIR})
+        if (GGML_BACKEND_DIR)
+            install(TARGETS ${backend} LIBRARY DESTINATION ${GGML_BACKEND_DIR})
+        else()
+            install(TARGETS ${backend} LIBRARY DESTINATION ${CMAKE_INSTALL_BINDIR})
+        endif()
     else()
         add_library(${backend} ${ARGN})
         target_link_libraries(ggml PUBLIC ${backend})
diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp
index f0cdac31eae..6c31513750c 100644
--- a/ggml/src/ggml-backend-reg.cpp
+++ b/ggml/src/ggml-backend-reg.cpp
@@ -498,6 +498,9 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
 
     std::vector<fs::path> search_paths;
     if (user_search_path == nullptr) {
+#ifdef GGML_BACKEND_DIR
+        search_paths.push_back(fs::u8path(GGML_BACKEND_DIR));
+#endif
         // default search paths: executable directory, current directory
         search_paths.push_back(get_executable_path());
         search_paths.push_back(fs::current_path());

From e0e122d1f73dbdfedaff6d0755b2077f4b582ee4 Mon Sep 17 00:00:00 2001
From: Romain Biessy <romain.biessy@codeplay.com>
Date: Tue, 5 Aug 2025 18:39:55 +0200
Subject: [PATCH 33/86] sycl: fix mul_mat selection (llama/15092)

---
 ggml/src/ggml-sycl/ggml-sycl.cpp | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
index f68f1739a9f..f8cdb03c3d4 100644
--- a/ggml/src/ggml-sycl/ggml-sycl.cpp
+++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
@@ -2609,6 +2609,8 @@ static void ggml_sycl_mul_mat_vec_nc(ggml_backend_sycl_context & ctx, const ggml
     GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(src0->buffer));
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->ne[1] == 1);
+    GGML_ASSERT(src1->ne[3] == 1);
 
     const int64_t ne00 = src0->ne[0];
     const int64_t ne01 = src0->ne[1];
@@ -3196,7 +3198,7 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
             // The kernel from the if path is faster for that specific case, but does not support all mul mats.
             ggml_sycl_mul_mat_batched_sycl(ctx, src0, src1, dst);
         }
-    } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
+    } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1 && src1->ne[3] == 1) {
         // KQV single-batch
         ggml_sycl_mul_mat_vec_nc(ctx, src0, src1, dst);
     } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2] * src1->ne[3] > 1) {

From 1141265ab3ec9dd99b9740fc4d62bd89db8f99a3 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Tue, 5 Aug 2025 22:10:36 +0300
Subject: [PATCH 34/86] llama : add gpt-oss (llama/15091)

* oai moe

* compat with new checkpoint

* add attn sink impl

* add rope scaling yarn

* logits match with latest transformers code

* wip chat template

* rm trailing space

* use ggml_scale_bias

* rm redundant is_swa_all

* convert interleaved gate_up

* graph : fix activation function to match reference (llama/7)

* vocab : handle o200k_harmony special tokens

* ggml : add attention sinks support (llama/1)

* llama : add attn sinks

* ggml : add attn sinks

* cuda : add attn sinks

* vulkan : add support for sinks in softmax

remove unnecessary return

* ggml : add fused swiglu_oai op (llama/11)

* ggml : add fused swiglu_oai op

* Update ggml/src/ggml-cpu/ops.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* update CUDA impl

* cont : metal impl

* add vulkan impl

* test-backend-ops : more test cases, clean up

* llama : remove unfused impl

* remove extra lines

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

---------

Co-authored-by: slaren <slarengh@gmail.com>

* repack mxfp4 upon conversion

* clean up a bit

* enable thinking

* add quick hack to render only some special tokens

* fix bf16 conversion

* remove vocab hack

* webui ok

* support chat parsing for gpt-oss

* fix webui

* direct mapping mxfp4, FINALLY

* force using mxfp4

* properly use lazy tensor

* ggml : add mxfp4

ggml : use e8m0 conversion instead of powf

Co-authored-by: Diego Devesa <slarengh@gmail.com>

change kvalues_mxfp4 table to match e2m1 (llama/6)

metal : remove quantization for now (not used)

cuda : fix disabled CUDA graphs due to ffn moe bias

vulkan : add support for mxfp4

cont : add cm2 dequant

* ggml : add ggml_add_id (llama/13)

* ggml : add ggml_add_id

* add cuda impl

* llama : add weight support check for add_id

* perf opt

* add vulkan impl

* rename cuda files

* add metal impl

* allow in-place ggml_add_id

* llama : keep biases on CPU with --cpu-moe

* llama : fix compile error

ggml-ci

* cuda : add fallback for __nv_cvt_e8m0_to_bf16raw

ggml-ci

* cleanup

ggml-ci

* sycl : fix supports_op for MXFP4

ggml-ci

* fix Unknown reasoning format

* ggml-cpu : fix AVX build

ggml-ci

* fix hip build

ggml-ci

* cuda : add mxfp4 dequantization support for cuBLAS

ggml-ci

* ggml-cpu : fix mxfp4 fallback definitions for some architectures

ggml-ci

* cuda : fix version required for __nv_cvt_e8m0_to_bf16raw

---------

Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
Co-authored-by: slaren <slarengh@gmail.com>
---
 ggml/include/ggml.h                           |  38 ++-
 ggml/src/ggml-alloc.c                         |   1 +
 ggml/src/ggml-cann/ggml-cann.cpp              |   8 +
 ggml/src/ggml-common.h                        |  17 ++
 ggml/src/ggml-cpu/arch-fallback.h             |   6 +
 ggml/src/ggml-cpu/arch/arm/quants.c           |  61 ++++
 ggml/src/ggml-cpu/arch/x86/quants.c           | 104 ++++++-
 ggml/src/ggml-cpu/ggml-cpu.c                  |  15 +-
 ggml/src/ggml-cpu/ops.cpp                     | 216 ++++++++++++-
 ggml/src/ggml-cpu/ops.h                       |   9 +-
 ggml/src/ggml-cpu/quants.c                    |  35 +++
 ggml/src/ggml-cpu/quants.h                    |   8 +
 ggml/src/ggml-cpu/vec.h                       |  23 +-
 ggml/src/ggml-cuda/add-id.cu                  |  58 ++++
 ggml/src/ggml-cuda/add-id.cuh                 |   3 +
 ggml/src/ggml-cuda/common.cuh                 |  26 ++
 ggml/src/ggml-cuda/convert.cu                 |  28 ++
 ggml/src/ggml-cuda/fattn-common.cuh           |   5 +-
 ggml/src/ggml-cuda/fattn-mma-f16.cuh          |   4 +-
 ggml/src/ggml-cuda/fattn-tile-f16.cu          |   3 +-
 ggml/src/ggml-cuda/fattn-tile-f32.cu          |   3 +-
 ggml/src/ggml-cuda/fattn-vec-f16.cuh          |  42 ++-
 ggml/src/ggml-cuda/fattn-vec-f32.cuh          |  40 ++-
 ggml/src/ggml-cuda/fattn-wmma-f16.cu          |   3 +-
 ggml/src/ggml-cuda/fattn.cu                   |  21 +-
 ggml/src/ggml-cuda/ggml-cuda.cu               |  25 +-
 ggml/src/ggml-cuda/im2col.cu                  |   5 +-
 ggml/src/ggml-cuda/mmq.cu                     |   4 +
 ggml/src/ggml-cuda/mmq.cuh                    |  82 ++++-
 ggml/src/ggml-cuda/mmvq.cu                    |   9 +
 ggml/src/ggml-cuda/softmax.cu                 |  26 +-
 .../template-instances/mmq-instance-mxfp4.cu  |   5 +
 ggml/src/ggml-cuda/unary.cu                   |  75 +++++
 ggml/src/ggml-cuda/unary.cuh                  |   2 +
 ggml/src/ggml-cuda/vecdotq.cuh                |  68 ++++-
 ggml/src/ggml-cuda/vendors/cuda.h             |   4 +
 ggml/src/ggml-impl.h                          |  61 ++++
 ggml/src/ggml-metal/ggml-metal-impl.h         |  14 +
 ggml/src/ggml-metal/ggml-metal.m              | 118 ++++++-
 ggml/src/ggml-metal/ggml-metal.metal          | 287 +++++++++++++++++-
 ggml/src/ggml-opencl/ggml-opencl.cpp          |   2 +
 ggml/src/ggml-quants.c                        | 116 ++++++-
 ggml/src/ggml-quants.h                        |   6 +
 ggml/src/ggml-sycl/ggml-sycl.cpp              |  20 +-
 ggml/src/ggml-vulkan/ggml-vulkan.cpp          | 139 ++++++++-
 .../ggml-vulkan/vulkan-shaders/add_id.comp    |  42 +++
 .../vulkan-shaders/copy_from_quant.comp       |   4 +-
 .../vulkan-shaders/dequant_funcs.comp         |  18 ++
 .../vulkan-shaders/dequant_funcs_cm2.comp     |  21 ++
 .../vulkan-shaders/dequant_mxfp4.comp         |  32 ++
 .../ggml-vulkan/vulkan-shaders/glu_head.comp  |   2 +
 .../ggml-vulkan/vulkan-shaders/mul_mm.comp    |  15 +
 .../vulkan-shaders/mul_mmq_funcs.comp         |   6 +
 .../ggml-vulkan/vulkan-shaders/soft_max.comp  |  12 +-
 .../vulkan-shaders/swiglu_oai.comp            |  14 +
 .../src/ggml-vulkan/vulkan-shaders/types.comp |  55 ++++
 .../vulkan-shaders/vulkan-shaders-gen.cpp     |   9 +-
 ggml/src/ggml.c                               |  88 +++++-
 58 files changed, 2019 insertions(+), 144 deletions(-)
 create mode 100644 ggml/src/ggml-cuda/add-id.cu
 create mode 100644 ggml/src/ggml-cuda/add-id.cuh
 create mode 100644 ggml/src/ggml-cuda/template-instances/mmq-instance-mxfp4.cu
 create mode 100644 ggml/src/ggml-vulkan/vulkan-shaders/add_id.comp
 create mode 100644 ggml/src/ggml-vulkan/vulkan-shaders/dequant_mxfp4.comp
 create mode 100644 ggml/src/ggml-vulkan/vulkan-shaders/swiglu_oai.comp

diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
index 8a8775be365..2f06e1e39b2 100644
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -304,6 +304,16 @@
     GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne) \
     GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
 
+#define GGML_TENSOR_TERNARY_OP_LOCALS \
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
+    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb) \
+    GGML_TENSOR_LOCALS(int64_t, ne2, src2, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb2, src2, nb) \
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
 #define GGML_TENSOR_BINARY_OP_LOCALS01 \
     GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
     GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
@@ -395,7 +405,8 @@ extern "C" {
         // GGML_TYPE_IQ4_NL_4_4 = 36,
         // GGML_TYPE_IQ4_NL_4_8 = 37,
         // GGML_TYPE_IQ4_NL_8_8 = 38,
-        GGML_TYPE_COUNT   = 39,
+        GGML_TYPE_MXFP4   = 39, // MXFP4 (1 block)
+        GGML_TYPE_COUNT   = 40,
     };
 
     // precision
@@ -430,6 +441,7 @@ extern "C" {
         GGML_FTYPE_MOSTLY_IQ4_XS  = 22, // except 1d tensors
         GGML_FTYPE_MOSTLY_IQ1_M   = 23, // except 1d tensors
         GGML_FTYPE_MOSTLY_BF16    = 24, // except 1d tensors
+        GGML_FTYPE_MOSTLY_MXFP4   = 25, // except 1d tensors
     };
 
     // available tensor operations:
@@ -438,6 +450,7 @@ extern "C" {
 
         GGML_OP_DUP,
         GGML_OP_ADD,
+        GGML_OP_ADD_ID,
         GGML_OP_ADD1,
         GGML_OP_ACC,
         GGML_OP_SUB,
@@ -557,6 +570,7 @@ extern "C" {
         GGML_GLU_OP_REGLU,
         GGML_GLU_OP_GEGLU,
         GGML_GLU_OP_SWIGLU,
+        GGML_GLU_OP_SWIGLU_OAI,
         GGML_GLU_OP_GEGLU_ERF,
         GGML_GLU_OP_GEGLU_QUICK,
 
@@ -831,6 +845,13 @@ extern "C" {
             struct ggml_tensor  * b,
             enum   ggml_type      type);
 
+    // dst[i0, i1, i2] = a[i0, i1, i2] + b[i0, ids[i1, i2]]
+    GGML_API struct ggml_tensor * ggml_add_id(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            struct ggml_tensor  * ids);
+
     GGML_API struct ggml_tensor * ggml_add1(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1198,6 +1219,13 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_swiglu_oai(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            float                 alpha,
+            float                 limit);
+
     // normalize along rows
     GGML_API struct ggml_tensor * ggml_norm(
             struct ggml_context * ctx,
@@ -1570,6 +1598,10 @@ extern "C" {
             float                 scale,
             float                 max_bias);
 
+    GGML_API void ggml_soft_max_add_sinks(
+            struct ggml_tensor * a,
+            struct ggml_tensor * sinks);
+
     GGML_API struct ggml_tensor * ggml_soft_max_ext_back(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -2052,6 +2084,10 @@ extern "C" {
     GGML_API enum ggml_prec ggml_flash_attn_ext_get_prec(
             const struct ggml_tensor * a);
 
+    GGML_API void ggml_flash_attn_ext_add_sinks(
+            struct ggml_tensor * a,
+            struct ggml_tensor * sinks);
+
     // TODO: needs to be adapted to ggml_flash_attn_ext
     GGML_API struct ggml_tensor * ggml_flash_attn_back(
            struct ggml_context * ctx,
diff --git a/ggml/src/ggml-alloc.c b/ggml/src/ggml-alloc.c
index fcc552da519..8b6e6028361 100644
--- a/ggml/src/ggml-alloc.c
+++ b/ggml/src/ggml-alloc.c
@@ -29,6 +29,7 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
         case GGML_OP_DIAG_MASK_ZERO:
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_ADD:
+        case GGML_OP_ADD_ID:
         case GGML_OP_ADD1:
         case GGML_OP_SUB:
         case GGML_OP_MUL:
diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
index 8eb8b1470bb..551c1f7673a 100755
--- a/ggml/src/ggml-cann/ggml-cann.cpp
+++ b/ggml/src/ggml-cann/ggml-cann.cpp
@@ -2340,6 +2340,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             memcpy(&bias, (float*)op->op_params + 1, sizeof(float));
             return bias == 0.0f; // TODO: support bias != 0.0f
         case GGML_OP_SOFT_MAX:
+            // TODO: support attention sinks [TAG_ATTN_SINKS]
+            if (op->src[2]) {
+                return false;
+            }
             // TODO: support broadcast
             // ref: https://github.com/ggml-org/llama.cpp/pull/14435
             return !op->src[1] || (op->src[1]->ne[2] == 1 && op->src[1]->ne[3] == 1);
@@ -2354,6 +2358,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             if(op->type != GGML_TYPE_F16 && op->type != GGML_TYPE_F32 && op->type != GGML_TYPE_BF16){
                 return false;
             }
+            // TODO: support attention sinks [TAG_ATTN_SINKS]
+            if (op->src[4]) {
+                return false;
+            }
             if (op->src[1]->ne[0] != op->src[2]->ne[0]) {
                 // different head sizes of K and V are not supported yet
                 return false;
diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h
index fbb04426abe..93ab7ea446e 100644
--- a/ggml/src/ggml-common.h
+++ b/ggml/src/ggml-common.h
@@ -99,6 +99,9 @@ typedef sycl::half2 ggml_half2;
 #define QI4_1 (QK4_1 / (4 * QR4_1))
 #define QR4_1 2
 
+#define QI_MXFP4 (QK_MXFP4 / (4 * QR_MXFP4))
+#define QR_MXFP4 2
+
 #define QI5_0 (QK5_0 / (4 * QR5_0))
 #define QR5_0 2
 
@@ -184,6 +187,13 @@ typedef struct {
 } block_q4_1;
 static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_half) + QK4_1 / 2, "wrong q4_1 block size/padding");
 
+#define QK_MXFP4 32
+typedef struct {
+    uint8_t e; // E8M0
+    uint8_t qs[QK_MXFP4/2];
+} block_mxfp4;
+static_assert(sizeof(block_mxfp4) == sizeof(uint8_t) + QK_MXFP4/2, "wrong mxfp4 block size/padding");
+
 #define QK5_0 32
 typedef struct {
     ggml_half d;           // delta
@@ -1074,10 +1084,17 @@ GGML_TABLE_BEGIN(uint32_t, iq3s_grid, 512)
     0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101,
 GGML_TABLE_END()
 
+// TODO: fix name to kvalues_iq4_nl
 GGML_TABLE_BEGIN(int8_t, kvalues_iq4nl, 16)
     -127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113,
 GGML_TABLE_END()
 
+// e2m1 values (doubled)
+// ref: https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf
+GGML_TABLE_BEGIN(int8_t, kvalues_mxfp4, 16)
+    0, 1, 2, 3, 4, 6, 8, 12, 0, -1, -2, -3, -4, -6, -8, -12,
+GGML_TABLE_END()
+
 #define NGRID_IQ1S 2048
 #define IQ1S_DELTA 0.125f
 #define IQ1M_DELTA 0.125f
diff --git a/ggml/src/ggml-cpu/arch-fallback.h b/ggml/src/ggml-cpu/arch-fallback.h
index f02cfe8fa5c..b62e3158ee7 100644
--- a/ggml/src/ggml-cpu/arch-fallback.h
+++ b/ggml/src/ggml-cpu/arch-fallback.h
@@ -13,6 +13,7 @@
 #define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
 #define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
 #define ggml_vec_dot_q8_0_q8_0_generic ggml_vec_dot_q8_0_q8_0
+#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
 #define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
 #define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
@@ -68,6 +69,7 @@
 #define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
 #define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
@@ -90,6 +92,7 @@
 #define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
 #define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
@@ -120,6 +123,7 @@
 #define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
 #define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
 #define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
+#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
@@ -149,6 +153,7 @@
 #define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
 #define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
 #define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
@@ -179,6 +184,7 @@
 #define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
 #define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
 #define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
+#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
diff --git a/ggml/src/ggml-cpu/arch/arm/quants.c b/ggml/src/ggml-cpu/arch/arm/quants.c
index c6d1d852e9a..aadbb487ec0 100644
--- a/ggml/src/ggml-cpu/arch/arm/quants.c
+++ b/ggml/src/ggml-cpu/arch/arm/quants.c
@@ -589,6 +589,67 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
     *s = sumf;
 }
 
+void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK_MXFP4 == 0);
+    static_assert(QK_MXFP4 == QK8_0, "QK_MXFP4 and QK8_0 must be the same");
+
+    const block_mxfp4 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_MXFP4;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined __ARM_NEON
+    const int8x16_t values = vld1q_s8(kvalues_mxfp4);
+    const uint8x16_t m4b = vdupq_n_u8(0x0f);
+    uint8x16x2_t q4bits;
+    int8x16x4_t q4b;
+    int8x16x4_t q8b;
+    int32x4_t prod_1;
+    int32x4_t prod_2;
+
+    for (; ib + 1 < nb; ib += 2) {
+        q4bits.val[0] = vld1q_u8(x[ib + 0].qs);
+        q4bits.val[1] = vld1q_u8(x[ib + 1].qs);
+        q8b.val[0]    = vld1q_s8(y[ib + 0].qs);
+        q8b.val[1]    = vld1q_s8(y[ib + 0].qs + 16);
+        q8b.val[2]    = vld1q_s8(y[ib + 1].qs);
+        q8b.val[3]    = vld1q_s8(y[ib + 1].qs + 16);
+
+        q4b.val[0] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[0], m4b));
+        q4b.val[1] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[0], 4));
+        q4b.val[2] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[1], m4b));
+        q4b.val[3] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[1], 4));
+
+        prod_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[0], q8b.val[0]), q4b.val[1], q8b.val[1]);
+        prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]);
+
+        sumf +=
+            GGML_E8M0_TO_FP32_HALF(x[ib + 0].e) * GGML_CPU_FP16_TO_FP32(y[ib + 0].d) * vaddvq_s32(prod_1) +
+            GGML_E8M0_TO_FP32_HALF(x[ib + 1].e) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) * vaddvq_s32(prod_2);
+    }
+
+#endif
+    for (; ib < nb; ++ib) {
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_E8M0_TO_FP32_HALF(x[ib].e);
+        int sumi1 = 0;
+        int sumi2 = 0;
+        for (int j = 0; j < QK_MXFP4/2; ++j) {
+            sumi1 += y[ib].qs[j +          0] * kvalues_mxfp4[x[ib].qs[j] & 0xf];
+            sumi2 += y[ib].qs[j + QK_MXFP4/2] * kvalues_mxfp4[x[ib].qs[j] >>  4];
+        }
+        sumf += d * (sumi1 + sumi2);
+    }
+    *s = sumf;
+}
+
 void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;
diff --git a/ggml/src/ggml-cpu/arch/x86/quants.c b/ggml/src/ggml-cpu/arch/x86/quants.c
index 30fd59f7028..cb49320a67f 100644
--- a/ggml/src/ggml-cpu/arch/x86/quants.c
+++ b/ggml/src/ggml-cpu/arch/x86/quants.c
@@ -66,6 +66,12 @@ static inline int hsum_i32_4(const __m128i a) {
 }
 
 #if defined(__AVX2__) || defined(__AVX512F__)
+static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
+    const __m256i ax = _mm256_sign_epi8(x, x);
+    const __m256i sy = _mm256_sign_epi8(y, x);
+    return _mm256_maddubs_epi16(ax, sy);
+}
+
 // spread 32 bits to 32 bytes { 0x00, 0xFF }
 static inline __m256i bytes_from_bits_32(const uint8_t * x) {
     uint32_t x32;
@@ -261,6 +267,11 @@ static inline __m256 quad_fp16_delta_float(const float x0, const float y0, const
     return _mm256_set_m128(_mm_set1_ps(GGML_CPU_FP16_TO_FP32(x1) * GGML_CPU_FP16_TO_FP32(y1)),
                            _mm_set1_ps(GGML_CPU_FP16_TO_FP32(x0) * GGML_CPU_FP16_TO_FP32(y0)));
 }
+
+static inline __m256 quad_mx_delta_float(const int8_t x0, const float y0, const int8_t x1, const float y1) {
+    return _mm256_set_m128(_mm_set1_ps(GGML_E8M0_TO_FP32_HALF(x1) * GGML_CPU_FP16_TO_FP32(y1)),
+                           _mm_set1_ps(GGML_E8M0_TO_FP32_HALF(x0) * GGML_CPU_FP16_TO_FP32(y0)));
+}
 #endif
 #elif defined(__SSSE3__)
 // horizontally add 4x4 floats
@@ -746,6 +757,91 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 #endif
 }
 
+void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK_MXFP4 == 0);
+    static_assert(QK_MXFP4 == QK8_0, "QK_MXFP4 and QK8_0 must be the same");
+
+    const block_mxfp4 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_MXFP4;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined __AVX2__
+
+    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_mxfp4);
+    const __m128i m4b  = _mm_set1_epi8(0x0f);
+    const __m256i mone = _mm256_set1_epi16(1);
+
+    __m256 accum1 = _mm256_setzero_ps();
+    __m256 accum2 = _mm256_setzero_ps();
+    for (; ib + 1 < nb; ib += 2) {
+        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)x[ib + 0].qs);
+        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)x[ib + 1].qs);
+        const __m256i q8b_1 = _mm256_loadu_si256((const __m256i *)y[ib + 0].qs);
+        const __m256i q8b_2 = _mm256_loadu_si256((const __m256i *)y[ib + 1].qs);
+        const __m256i q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
+                                              _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
+        const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
+                                              _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b)));
+        const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
+        const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
+        const __m256i p_1 = _mm256_madd_epi16(p16_1, mone);
+        const __m256i p_2 = _mm256_madd_epi16(p16_2, mone);
+        accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_E8M0_TO_FP32_HALF(x[ib + 0].e)),
+                _mm256_cvtepi32_ps(p_1), accum1);
+        accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_E8M0_TO_FP32_HALF(x[ib + 1].e)),
+                _mm256_cvtepi32_ps(p_2), accum2);
+    }
+
+    sumf = hsum_float_8(_mm256_add_ps(accum1, accum2));
+
+#elif defined __AVX__
+    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_mxfp4);
+    const __m128i m4b  = _mm_set1_epi8(0x0f);
+
+    __m256 accum = _mm256_setzero_ps();
+    for (; ib + 1 < nb; ib += 2) {
+        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)x[ib + 0].qs);
+        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
+        const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs);
+        const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs + 1);
+        const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
+        const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs + 1);
+
+        const __m128i q4b_1_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b));
+        const __m128i q4b_1_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b));
+        const __m128i q4b_2_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b));
+        const __m128i q4b_2_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b));
+
+        const __m256 p = mul_sum_i8_quad_float(q4b_1_0, q4b_1_1, q4b_2_0, q4b_2_1, q8b_1_0, q8b_1_1, q8b_2_0, q8b_2_1);
+        const __m256 deltas = quad_mx_delta_float(x[ib].e, y[ib].d, x[ib + 1].e, y[ib + 1].d);
+        accum = _mm256_add_ps(_mm256_mul_ps(deltas, p), accum);
+    }
+
+    sumf = hsum_float_8(accum);
+
+#endif
+    for (; ib < nb; ++ib) {
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_E8M0_TO_FP32_HALF(x[ib].e);
+        int sumi1 = 0;
+        int sumi2 = 0;
+        for (int j = 0; j < QK_MXFP4/2; ++j) {
+            sumi1 += y[ib].qs[j +          0] * kvalues_mxfp4[x[ib].qs[j] & 0xf];
+            sumi2 += y[ib].qs[j + QK_MXFP4/2] * kvalues_mxfp4[x[ib].qs[j] >>  4];
+        }
+        sumf += d * (sumi1 + sumi2);
+    }
+    *s = sumf;
+}
+
 void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;
@@ -3206,14 +3302,6 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 #endif
 }
 
-#if defined(__AVX2__)
-static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
-    const __m256i ax = _mm256_sign_epi8(x, x);
-    const __m256i sy = _mm256_sign_epi8(y, x);
-    return _mm256_maddubs_epi16(ax, sy);
-}
-#endif
-
 void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
     assert(nrc == 1);
diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
index c5271b77572..d89cd8f4ef6 100644
--- a/ggml/src/ggml-cpu/ggml-cpu.c
+++ b/ggml/src/ggml-cpu/ggml-cpu.c
@@ -253,6 +253,12 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .vec_dot_type             = GGML_TYPE_Q8_1,
         .nrows                    = 1,
     },
+    [GGML_TYPE_MXFP4] = {
+        .from_float               = quantize_row_mxfp4,
+        .vec_dot                  = ggml_vec_dot_mxfp4_q8_0,
+        .vec_dot_type             = GGML_TYPE_Q8_0,
+        .nrows                    = 1,
+    },
     [GGML_TYPE_Q2_K] = {
         .from_float               = quantize_row_q2_K,
         .vec_dot                  = ggml_vec_dot_q2_K_q8_K,
@@ -1670,6 +1676,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_add(params, tensor);
             } break;
+        case GGML_OP_ADD_ID:
+            {
+                ggml_compute_forward_add_id(params, tensor);
+            } break;
         case GGML_OP_ADD1:
             {
                 ggml_compute_forward_add1(params, tensor);
@@ -1924,7 +1934,7 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             } break;
         case GGML_OP_FLASH_ATTN_EXT:
             {
-                ggml_compute_forward_flash_attn_ext(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor->src[3], tensor);
+                ggml_compute_forward_flash_attn_ext(params, tensor);
             } break;
         case GGML_OP_FLASH_ATTN_BACK:
             {
@@ -2111,6 +2121,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_DUP:
         case GGML_OP_CONT:
         case GGML_OP_ADD:
+        case GGML_OP_ADD_ID:
         case GGML_OP_ADD1:
         case GGML_OP_ACC:
             {
@@ -2172,6 +2183,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 case GGML_GLU_OP_REGLU:
                 case GGML_GLU_OP_GEGLU:
                 case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_SWIGLU_OAI:
                 case GGML_GLU_OP_GEGLU_ERF:
                 case GGML_GLU_OP_GEGLU_QUICK:
                     {
@@ -2673,6 +2685,7 @@ struct ggml_cplan ggml_graph_plan(
                         }
                     } break;
                 case GGML_OP_ADD:
+                case GGML_OP_ADD_ID:
                 case GGML_OP_ADD1:
                     {
                         if (ggml_is_quantized(node->src[0]->type)) {
diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
index 6581d27adde..854f1c2b496 100644
--- a/ggml/src/ggml-cpu/ops.cpp
+++ b/ggml/src/ggml-cpu/ops.cpp
@@ -8,6 +8,7 @@
 #include "vec.h"
 
 #include <float.h>
+#include <algorithm>
 
 // ggml_compute_forward_dup
 
@@ -1283,6 +1284,7 @@ void ggml_compute_forward_add(
         case GGML_TYPE_Q5_0:
         case GGML_TYPE_Q5_1:
         case GGML_TYPE_Q8_0:
+        case GGML_TYPE_MXFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -1309,6 +1311,77 @@ void ggml_compute_forward_add(
     }
 }
 
+// ggml_compute_forward_add_id
+
+static void ggml_compute_forward_add_id_f32(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    const ggml_tensor * src2 = dst->src[2];
+
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(src2->type == GGML_TYPE_I32);
+
+    GGML_ASSERT(src0->nb[0] == sizeof(float));
+    GGML_ASSERT(src1->nb[0] == sizeof(float));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nr  = ggml_nrows(src0);
+
+    GGML_TENSOR_TERNARY_OP_LOCALS
+
+    GGML_ASSERT( nb0 == sizeof(float));
+    GGML_ASSERT(nb10 == sizeof(float));
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int ir = ir0; ir < ir1; ++ir) {
+        // src0 indices
+        const int i3 = ir/(ne2*ne1);
+        const int i2 = (ir - i3*ne2*ne1)/ne1;
+        const int i1 = (ir - i3*ne2*ne1 - i2*ne1);
+
+        // src1 indices
+        const int i11 = *(int32_t *) ((char *) src2->data + i1*nb20 + i2*nb21);
+
+        GGML_ASSERT(i11 >= 0 && i11 < ne11);
+
+        ggml_vec_add_f32(ne0,
+                (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1 ),
+                (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01),
+                (float *) ((char *) src1->data + i11*nb11));
+    }
+}
+
+void ggml_compute_forward_add_id(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_add_id_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("unsupported type for ggml_compute_forward_add_id: %s", ggml_type_name(src0->type));
+            }
+    }
+}
+
 // ggml_compute_forward_add1
 
 static void ggml_compute_forward_add1_f32(
@@ -1660,6 +1733,7 @@ void ggml_compute_forward_add1(
         case GGML_TYPE_Q5_1:
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_Q8_1:
+        case GGML_TYPE_MXFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -1787,6 +1861,7 @@ void ggml_compute_forward_acc(
         case GGML_TYPE_Q5_1:
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_Q8_1:
+        case GGML_TYPE_MXFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -3614,6 +3689,93 @@ static void ggml_compute_forward_swiglu(
     }
 }
 
+// ggml_compute_forward_swiglu_oai
+
+static void ggml_compute_forward_swiglu_oai_f32(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    char * src0_d = (char *) src0->data;
+    char * src1_d = (char *) (src1 ? src1->data : src0->data);
+    const size_t src0_o = src0->nb[1];
+    const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(ggml_is_contiguous_1(dst));
+
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
+    const int nr = ggml_nrows(src0);
+
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_nrows(dst) == nr);
+
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+    const float alpha = ggml_get_op_params_f32(dst, 2);
+    const float limit = ggml_get_op_params_f32(dst, 3);
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int i1 = ir0; i1 < ir1; i1++) {
+        float * src0_p = (float *) (src0_d + i1*src0_o);
+        float * src1_p = (float *) (src1_d + i1*src1_o);
+        float * dst_p  = (float *) ((char *) dst->data + i1*(dst->nb[1]));
+
+        if (!src1) {
+            src0_p += swapped ? nc : 0;
+            src1_p += swapped ? 0 : nc;
+        }
+
+        for (int k = 0; k < nc; k++) {
+            const float x = std::min(src0_p[k], limit);
+            const float y = std::clamp(src1_p[k], -limit, limit);
+            const float out_glu = x / (1.f + expf(alpha * (-x)));
+            dst_p[k] = out_glu * (y + 1.f);
+        }
+
+#ifndef NDEBUG
+        for (int k = 0; k < nc; k++) {
+            const float x = dst_p[k];
+            GGML_UNUSED(x);
+            assert(!isnan(x));
+            assert(!isinf(x));
+        }
+#endif
+    }
+}
+
+static void ggml_compute_forward_swiglu_oai(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_swiglu_oai_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_geglu_erf
 
 static void ggml_compute_forward_geglu_erf_f32(
@@ -4599,6 +4761,7 @@ void ggml_compute_forward_out_prod(
         case GGML_TYPE_Q5_0:
         case GGML_TYPE_Q5_1:
         case GGML_TYPE_Q8_0:
+        case GGML_TYPE_MXFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -4873,6 +5036,7 @@ void ggml_compute_forward_set(
         case GGML_TYPE_Q5_1:
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_Q8_1:
+        case GGML_TYPE_MXFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -5134,6 +5298,7 @@ void ggml_compute_forward_get_rows(
         case GGML_TYPE_Q5_1:
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_Q8_1:
+        case GGML_TYPE_MXFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -5523,6 +5688,7 @@ static void ggml_compute_forward_soft_max_f32(
 
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
+    const ggml_tensor * src2 = dst->src[2];
 
     assert(ggml_is_contiguous(dst));
     assert(ggml_are_same_shape(src0, dst));
@@ -5557,6 +5723,9 @@ static void ggml_compute_forward_soft_max_f32(
 
     const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
 
+    // sinks
+    const float * sk = src2 ? (float *)((char *) src2->data) : nullptr;
+
     for (int64_t i03 = 0; i03 < ne03; i03++) {
         for (int64_t i02 = 0; i02 < ne02; i02++) {
             for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
@@ -5599,9 +5768,18 @@ static void ggml_compute_forward_soft_max_f32(
                 float max = -INFINITY;
                 ggml_vec_max_f32(ne00, &max, wp);
 
+                // if we have sinks, make a correction as if they were included in the softmax
+                if (sk) {
+                    max = MAX(max, sk[i02]);
+                }
+
                 ggml_float sum = ggml_vec_soft_max_f32(ne00, dp, wp, max);
                 assert(sum > 0.0);
 
+                if (sk) {
+                    sum += (ggml_float) expf(sk[i02] - max);
+                }
+
                 sum = 1.0/sum;
                 ggml_vec_scale_f32(ne00, dp, sum);
 
@@ -5836,6 +6014,7 @@ void ggml_compute_forward_clamp(
         case GGML_TYPE_Q5_1:
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_Q8_1:
+        case GGML_TYPE_MXFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -7989,12 +8168,14 @@ void ggml_compute_forward_argsort(
 
 static void ggml_compute_forward_flash_attn_ext_f16(
         const ggml_compute_params * params,
-        const ggml_tensor * q,
-        const ggml_tensor * k,
-        const ggml_tensor * v,
-        const ggml_tensor * mask,
         ggml_tensor * dst) {
 
+    const ggml_tensor * q     = dst->src[0];
+    const ggml_tensor * k     = dst->src[1];
+    const ggml_tensor * v     = dst->src[2];
+    const ggml_tensor * mask  = dst->src[3];
+    const ggml_tensor * sinks = dst->src[4];
+
     GGML_TENSOR_LOCALS(int64_t, neq, q,   ne)
     GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb)
     GGML_TENSOR_LOCALS(int64_t, nek, k,   ne)
@@ -8189,6 +8370,23 @@ static void ggml_compute_forward_flash_attn_ext_f16(
             }
         }
 
+        // sinks
+        if (sinks) {
+            const float s = ((float *)((char *) sinks->data))[h];
+
+            float ms = 1.0f;
+            float vs = 1.0f;
+
+            if (s > M) {
+                ms = expf(M - s);
+                ggml_vec_scale_f32(DV, VKQ32, ms);
+            } else {
+                vs = expf(s - M);
+            }
+
+            S = S*ms + vs;
+        }
+
         // V /= S
         const float S_inv = 1.0f/S;
         ggml_vec_scale_f32(DV, VKQ32, S_inv);
@@ -8208,17 +8406,13 @@ static void ggml_compute_forward_flash_attn_ext_f16(
 
 void ggml_compute_forward_flash_attn_ext(
         const ggml_compute_params * params,
-        const ggml_tensor * q,
-        const ggml_tensor * k,
-        const ggml_tensor * v,
-        const ggml_tensor * mask,
         ggml_tensor * dst) {
     switch (dst->op_params[3]) {
         case GGML_PREC_DEFAULT:
         case GGML_PREC_F32:
             {
                 // uses F32 accumulators
-                ggml_compute_forward_flash_attn_ext_f16(params, q, k, v, mask, dst);
+                ggml_compute_forward_flash_attn_ext_f16(params, dst);
             } break;
         default:
             {
@@ -9080,6 +9274,10 @@ void ggml_compute_forward_glu(
             {
                 ggml_compute_forward_swiglu(params, dst);
             } break;
+        case GGML_GLU_OP_SWIGLU_OAI:
+            {
+                ggml_compute_forward_swiglu_oai(params, dst);
+            } break;
         case GGML_GLU_OP_GEGLU_ERF:
             {
                 ggml_compute_forward_geglu_erf(params, dst);
diff --git a/ggml/src/ggml-cpu/ops.h b/ggml/src/ggml-cpu/ops.h
index 3a32ec20dba..f154afb4624 100644
--- a/ggml/src/ggml-cpu/ops.h
+++ b/ggml/src/ggml-cpu/ops.h
@@ -29,6 +29,7 @@ extern "C" {
 
 void ggml_compute_forward_dup(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_add(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_add_id(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_add1(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_acc(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_sum(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@@ -82,13 +83,7 @@ void ggml_compute_forward_arange(const struct ggml_compute_params * params, stru
 void ggml_compute_forward_timestep_embedding(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_argsort(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_leaky_relu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-void ggml_compute_forward_flash_attn_ext(
-    const struct ggml_compute_params * params,
-    const struct ggml_tensor * q,
-    const struct ggml_tensor * k,
-    const struct ggml_tensor * v,
-    const struct ggml_tensor * mask,
-    struct ggml_tensor * dst);
+void ggml_compute_forward_flash_attn_ext(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_flash_attn_back(
         const struct ggml_compute_params * params,
         const bool masked,
diff --git a/ggml/src/ggml-cpu/quants.c b/ggml/src/ggml-cpu/quants.c
index ee35ab42fda..365cb36d2d7 100644
--- a/ggml/src/ggml-cpu/quants.c
+++ b/ggml/src/ggml-cpu/quants.c
@@ -46,6 +46,10 @@ void quantize_row_q8_1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRI
     quantize_row_q8_1_ref(x, y, k);
 }
 
+void quantize_row_mxfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+    quantize_row_mxfp4_ref(x, y, k);
+}
+
 //
 // 2-6 bit quantization in super-blocks
 //
@@ -181,6 +185,37 @@ void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, c
     *s = sumf;
 }
 
+void ggml_vec_dot_mxfp4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK_MXFP4 == 0);
+    static_assert(QK_MXFP4 == QK8_0, "QK_MXFP4 and QK8_0 must be the same");
+
+    const block_mxfp4 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_MXFP4;
+
+    int ib = 0;
+    float sumf = 0;
+
+    for (; ib < nb; ++ib) {
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_E8M0_TO_FP32_HALF(x[ib].e);
+
+        int sumi1 = 0;
+        int sumi2 = 0;
+        for (int j = 0; j < QK_MXFP4/2; ++j) {
+            sumi1 += y[ib].qs[j +          0] * kvalues_mxfp4[x[ib].qs[j] & 0xf];
+            sumi2 += y[ib].qs[j + QK_MXFP4/2] * kvalues_mxfp4[x[ib].qs[j] >>  4];
+        }
+        sumf += d * (sumi1 + sumi2);
+    }
+    *s = sumf;
+}
+
 void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;
diff --git a/ggml/src/ggml-cpu/quants.h b/ggml/src/ggml-cpu/quants.h
index dc4342c87f5..d83eb1b144d 100644
--- a/ggml/src/ggml-cpu/quants.h
+++ b/ggml/src/ggml-cpu/quants.h
@@ -19,6 +19,8 @@ void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, in
 void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 
+void quantize_row_mxfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
 void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
@@ -39,6 +41,8 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 
+void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
 void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
@@ -67,8 +71,12 @@ void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, c
 void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q5_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q8_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_mxfp4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
 void ggml_vec_dot_tq1_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_tq2_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
 void ggml_vec_dot_q2_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q3_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q4_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
diff --git a/ggml/src/ggml-cpu/vec.h b/ggml/src/ggml-cpu/vec.h
index d18783a00a1..2250d93cb00 100644
--- a/ggml/src/ggml-cpu/vec.h
+++ b/ggml/src/ggml-cpu/vec.h
@@ -55,7 +55,22 @@ inline static void ggml_vec_cpy_i32(const int n, int32_t * y, const int32_t * x)
 
 inline static void ggml_vec_set_f16(const int n, ggml_fp16_t * x, const ggml_fp16_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
 inline static void ggml_vec_set_bf16(const int n, ggml_bf16_t * x, const ggml_bf16_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
-inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i] + y[i]; }
+
+inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) {
+    int i = 0;
+#if defined(__AVX2__)
+    for (; i + 7 < n; i += 8) {
+        __m256 vx = _mm256_loadu_ps(x + i);
+        __m256 vy = _mm256_loadu_ps(y + i);
+        __m256 vz = _mm256_add_ps(vx, vy);
+        _mm256_storeu_ps(z + i, vz);
+    }
+#endif
+    for (; i < n; ++i) {
+        z[i] = x[i] + y[i];
+    }
+}
+
 inline static void ggml_vec_add_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
     for (int i = 0; i < n; ++i) {
         z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) + GGML_CPU_FP16_TO_FP32(y[i]));
@@ -992,9 +1007,9 @@ void ggml_vec_swiglu_f32(const int n, float * y, const float * x, const float *
 
 inline static void ggml_vec_swiglu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
     for (int i = 0; i < n; ++i) {
-        float v = GGML_CPU_FP16_TO_FP32(x[i]);
-        float w = GGML_CPU_FP16_TO_FP32(g[i]);
-        y[i] = GGML_CPU_FP32_TO_FP16((v/(1.0f + expf(-v))) * w);
+        float xi = GGML_CPU_FP16_TO_FP32(x[i]);
+        float gi = GGML_CPU_FP16_TO_FP32(g[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16((xi/(1.0f + expf(-xi))) * gi);
     }
 }
 
diff --git a/ggml/src/ggml-cuda/add-id.cu b/ggml/src/ggml-cuda/add-id.cu
new file mode 100644
index 00000000000..8bed62ac9d2
--- /dev/null
+++ b/ggml/src/ggml-cuda/add-id.cu
@@ -0,0 +1,58 @@
+#include "add-id.cuh"
+
+static __global__ void add_id_kernel(
+        const float * src0, const float * src1, const int32_t * src2, float * dst,
+        int64_t ne0, int64_t ne1,
+        size_t nb01, size_t nb02,
+        size_t nb11,
+        size_t nb21
+    ) {
+
+    const int64_t i1 = blockIdx.x;
+    const int64_t i2 = blockIdx.y;
+
+    const int i11 = *(int32_t *) ((char *) src2 + i1*sizeof(int32_t) + i2*nb21);
+
+    const size_t nb1 = ne0 * sizeof(float);
+    const size_t nb2 = ne1 * nb1;
+
+    float * dst_row = (float *)((char *)dst + i1*nb1 + i2*nb2);
+    const float * src0_row = (const float *)((char *)src0 +  i1*nb01 + i2*nb02);
+    const float * src1_row = (const float *)((char *)src1 + i11*nb11);
+
+    for (int64_t i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) {
+        dst_row[i0] = src0_row[i0] + src1_row[i0];
+    }
+}
+
+void ggml_cuda_op_add_id(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    const ggml_tensor * src2 = dst->src[2];
+
+    GGML_TENSOR_TERNARY_OP_LOCALS
+
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(src2->type == GGML_TYPE_I32);
+
+    GGML_ASSERT(nb00 == sizeof(float));
+    GGML_ASSERT(nb10 == sizeof(float));
+    GGML_ASSERT(nb20 == sizeof(int32_t));
+
+    const float * src0_d = (const float *)src0->data;
+    const float * src1_d = (const float *)src1->data;
+    const int32_t * src2_d = (const int32_t *)src2->data;
+    float * dst_d = (float *)dst->data;
+
+    int threads = std::min((int)ne00, 768); // cols
+    dim3 blocks(ne01, ne02); // n_experts_used, n_tokens
+    add_id_kernel<<<blocks, threads, 0, ctx.stream()>>>(
+        src0_d, src1_d, src2_d, dst_d,
+        ne0, ne1,
+        nb01, nb02,
+        nb11,
+        nb21
+    );
+}
diff --git a/ggml/src/ggml-cuda/add-id.cuh b/ggml/src/ggml-cuda/add-id.cuh
new file mode 100644
index 00000000000..30b1721ac32
--- /dev/null
+++ b/ggml/src/ggml-cuda/add-id.cuh
@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_op_add_id(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
index 7fb04d51b77..8f27255476d 100644
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -1,6 +1,7 @@
 #pragma once
 
 #include "ggml.h"
+#include "ggml-impl.h"
 #include "ggml-cuda.h"
 
 #include <cstdint>
@@ -549,6 +550,24 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
 #endif // defined(GGML_USE_HIP)
 }
 
+static __device__ __forceinline__ float ggml_cuda_e8m0_to_fp32(uint8_t x) {
+#if CUDART_VERSION >= 12080
+    const nv_bfloat16 e = __nv_cvt_e8m0_to_bf16raw(x);
+    return (float) e;
+#else
+    uint32_t bits;
+    if (x == 0) {
+        bits = 0x00400000;
+    } else {
+        bits = (uint32_t) x << 23;
+    }
+
+    float result;
+    memcpy(&result, &bits, sizeof(float));
+    return result;
+#endif // CUDART_VERSION >= 12050
+}
+
 typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);
 
 static __device__ __forceinline__ float get_alibi_slope(
@@ -607,6 +626,13 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q8_0> {
     static constexpr int qi = QI8_0;
 };
 
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_MXFP4> {
+    static constexpr int qk = QK_MXFP4;
+    static constexpr int qr = QR_MXFP4;
+    static constexpr int qi = QI_MXFP4;
+};
+
 template<>
 struct ggml_cuda_type_traits<GGML_TYPE_Q2_K> {
     static constexpr int qk = QK_K;
diff --git a/ggml/src/ggml-cuda/convert.cu b/ggml/src/ggml-cuda/convert.cu
index 15c927861f0..e3beddbc1b2 100644
--- a/ggml/src/ggml-cuda/convert.cu
+++ b/ggml/src/ggml-cuda/convert.cu
@@ -465,6 +465,24 @@ static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst
     }
 }
 
+template<typename dst_t>
+static __global__ void dequantize_block_mxfp4(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i   = blockIdx.x;
+    const block_mxfp4 * x = (const block_mxfp4 *) vx + i*(QK_K/QK_MXFP4);
+
+    const int64_t tid = threadIdx.x;
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 4*il;
+    const uint8_t  * q4 = x[ib].qs + 4*il;
+    const float d = ggml_cuda_e8m0_to_fp32(x[ib].e);
+    for (int j = 0; j < 4; ++j) {
+        y[j+ 0] = d * kvalues_mxfp4[q4[j] & 0xf]*0.5f;
+        y[j+16] = d * kvalues_mxfp4[q4[j] >>  4]*0.5f;
+    }
+}
+
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
 static void dequantize_block_cuda(const void * vx, dst_t * y,
         const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
@@ -588,6 +606,12 @@ static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int64_t
     dequantize_block_iq4_xs<<<nb, 32, 0, stream>>>(vx, y);
 }
 
+template<typename dst_t>
+static void dequantize_row_mxfp4_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = (k + QK_K - 1) / QK_K;
+    dequantize_block_mxfp4<<<nb, 32, 0, stream>>>(vx, y);
+}
+
 template <typename src_t, typename dst_t>
 static __global__ void convert_unary(
         const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t ne00, const int64_t ne01, const int64_t ne02,
@@ -677,6 +701,8 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
             return dequantize_row_iq4_xs_cuda;
         case GGML_TYPE_IQ3_S:
             return dequantize_row_iq3_s_cuda;
+        case GGML_TYPE_MXFP4:
+            return dequantize_row_mxfp4_cuda;
         case GGML_TYPE_F32:
             return convert_unary_cont_cuda<float>;
         case GGML_TYPE_BF16:
@@ -726,6 +752,8 @@ to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
             return dequantize_row_iq4_xs_cuda;
         case GGML_TYPE_IQ3_S:
             return dequantize_row_iq3_s_cuda;
+        case GGML_TYPE_MXFP4:
+            return dequantize_row_mxfp4_cuda;
         case GGML_TYPE_F16:
             return convert_unary_cont_cuda<half>;
         case GGML_TYPE_BF16:
diff --git a/ggml/src/ggml-cuda/fattn-common.cuh b/ggml/src/ggml-cuda/fattn-common.cuh
index b6db446c6fe..e46f0e2081b 100644
--- a/ggml/src/ggml-cuda/fattn-common.cuh
+++ b/ggml/src/ggml-cuda/fattn-common.cuh
@@ -15,6 +15,7 @@ typedef void (* fattn_kernel_t)(
         const char * __restrict__ K,
         const char * __restrict__ V,
         const char * __restrict__ mask,
+        const char * __restrict__ sinks,
         const int  * __restrict__ KV_max,
         float      * __restrict__ dst,
         float2     * __restrict__ dst_meta,
@@ -736,7 +737,8 @@ void launch_fattn(
 
     GGML_ASSERT(V || is_mla);
 
-    const ggml_tensor * mask = dst->src[3];
+    const ggml_tensor * mask  = dst->src[3];
+    const ggml_tensor * sinks = dst->src[4];
 
     ggml_tensor * KQV = dst;
 
@@ -940,6 +942,7 @@ void launch_fattn(
         K_data,
         V_data,
         mask ? ((const char *) mask->data) : nullptr,
+        sinks ? ((const char *) sinks->data) : nullptr,
         KV_max.ptr,
         !stream_k && parallel_blocks > 1 ? dst_tmp.ptr : (float *) KQV->data, dst_tmp_meta.ptr,
         scale, max_bias, m0, m1, n_head_log2, logit_softcap,
diff --git a/ggml/src/ggml-cuda/fattn-mma-f16.cuh b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
index a86b95428f6..e7570f9d3b8 100644
--- a/ggml/src/ggml-cuda/fattn-mma-f16.cuh
+++ b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
@@ -1206,6 +1206,7 @@ static __global__ void flash_attn_ext_f16(
         const char * __restrict__ K,
         const char * __restrict__ V,
         const char * __restrict__ mask,
+        const char * __restrict__ sinks,
         const int  * __restrict__ KV_max,
         float      * __restrict__ dst,
         float2     * __restrict__ dst_meta,
@@ -1267,6 +1268,7 @@ static __global__ void flash_attn_ext_f16(
     // kb0 == k start index when in the output tile.
     int kb0_start = kbc % iter_k;
     int kb0_stop  = min(iter_k, kb0_start + kbc_stop - kbc);
+
     while (kbc < kbc_stop && kb0_stop == iter_k) {
         const int sequence = kbc / (iter_k*iter_j*(ne02/ncols2));
         const int head = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence) / (iter_k*iter_j);
@@ -1340,7 +1342,7 @@ static __global__ void flash_attn_ext_f16(
         (Q_f2, K_h2, V_h2, mask_h2, dstk, dst_meta, scale, slope, logit_softcap,
          ne01, ne02, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel);
 #else
-    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
+    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); GGML_UNUSED(sinks);
     GGML_UNUSED(dst); GGML_UNUSED(dst_meta);
     GGML_UNUSED(scale); GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
     GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
diff --git a/ggml/src/ggml-cuda/fattn-tile-f16.cu b/ggml/src/ggml-cuda/fattn-tile-f16.cu
index 9d0b24ae7ec..0fcfaa32ea4 100644
--- a/ggml/src/ggml-cuda/fattn-tile-f16.cu
+++ b/ggml/src/ggml-cuda/fattn-tile-f16.cu
@@ -13,6 +13,7 @@ static __global__ void flash_attn_tile_ext_f16(
         const char * __restrict__ K,
         const char * __restrict__ V,
         const char * __restrict__ mask,
+        const char * __restrict__ sinks,
         const int  * __restrict__ KV_max,
         float      * __restrict__ dst,
         float2     * __restrict__ dst_meta,
@@ -272,7 +273,7 @@ static __global__ void flash_attn_tile_ext_f16(
         }
     }
 #else
-    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
+    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); GGML_UNUSED(sinks);
     GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
     GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
     GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
diff --git a/ggml/src/ggml-cuda/fattn-tile-f32.cu b/ggml/src/ggml-cuda/fattn-tile-f32.cu
index be72f76fb65..23550cbbd97 100644
--- a/ggml/src/ggml-cuda/fattn-tile-f32.cu
+++ b/ggml/src/ggml-cuda/fattn-tile-f32.cu
@@ -13,6 +13,7 @@ static __global__ void flash_attn_tile_ext_f32(
         const char * __restrict__ K,
         const char * __restrict__ V,
         const char * __restrict__ mask,
+        const char * __restrict__ sinks,
         const int  * __restrict__ KV_max,
         float      * __restrict__ dst,
         float2     * __restrict__ dst_meta,
@@ -37,7 +38,7 @@ static __global__ void flash_attn_tile_ext_f32(
     return;
 #endif // FP16_MMA_AVAILABLE
     if (use_logit_softcap && !(D == 128 || D == 256)) {
-        GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
+        GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); GGML_UNUSED(sinks);
         GGML_UNUSED(dst); GGML_UNUSED(dst_meta);
         GGML_UNUSED(scale); GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
         GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
diff --git a/ggml/src/ggml-cuda/fattn-vec-f16.cuh b/ggml/src/ggml-cuda/fattn-vec-f16.cuh
index a2df2f66be0..b05f682cd3b 100644
--- a/ggml/src/ggml-cuda/fattn-vec-f16.cuh
+++ b/ggml/src/ggml-cuda/fattn-vec-f16.cuh
@@ -16,6 +16,7 @@ static __global__ void flash_attn_vec_ext_f16(
         const char * __restrict__ K,
         const char * __restrict__ V,
         const char * __restrict__ mask,
+        const char * __restrict__ sinks,
         const int  * __restrict__ KV_max,
         float      * __restrict__ dst,
         float2     * __restrict__ dst_meta,
@@ -61,7 +62,8 @@ static __global__ void flash_attn_vec_ext_f16(
     K += nb13*sequence + nb12*(head / gqa_ratio);
     V += nb23*sequence + nb22*(head / gqa_ratio);
 
-    const half * maskh = (const half *) (mask + nb33*(sequence % ne33) + nb31*ic0);
+    const half  * maskh  = (const half  *) (mask + nb33*(sequence % ne33) + nb31*ic0);
+    const float * sinksf = (const float *) (sinks);
 
     const float slopef = get_alibi_slope(max_bias, head, n_head_log2, m0, m1);
     const half  slopeh = __float2half(slopef);
@@ -75,11 +77,12 @@ static __global__ void flash_attn_vec_ext_f16(
     half2 * KQ2 = (half2 *) KQ;
 
     half kqmax[ncols];
+    half kqsum[ncols];
 #pragma unroll
     for (int j = 0; j < ncols; ++j) {
         kqmax[j] = -HALF_MAX_HALF;
+        kqsum[j] = 0.0f;
     }
-    half kqsum[ncols] = {0.0f};
 
     __shared__ half kqmax_shared[ncols][WARP_SIZE];
     __shared__ half kqsum_shared[ncols][WARP_SIZE];
@@ -283,6 +286,39 @@ static __global__ void flash_attn_vec_ext_f16(
         __syncthreads();
     }
 
+    if (sinksf && blockIdx.y == 0) {
+        const half sink = __float2half(sinksf[head]);
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            if (threadIdx.x == 0) {
+                kqmax_shared[j][threadIdx.y] = fmaxf(kqmax[j], sink);
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            half kqmax_new_j = kqmax_shared[j][threadIdx.x];
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+
+            const half KQ_max_scale = hexp(kqmax[j] - kqmax_new_j);
+            kqmax[j] = kqmax_new_j;
+
+            const half val = hexp(sink - kqmax[j]);
+            kqsum[j] = kqsum[j]*KQ_max_scale;
+
+            if (tid == 0) {
+                kqsum[j] += val;
+            }
+
+            VKQ[j] *= __half2half2(KQ_max_scale);
+        }
+
+        __syncthreads();
+    }
+
 #pragma unroll
     for (int j = 0; j < ncols; ++j) {
         kqsum[j] = warp_reduce_sum((float)kqsum[j]);
@@ -313,7 +349,7 @@ static __global__ void flash_attn_vec_ext_f16(
         dst_meta[((sequence*ne01 + ic0 + tid)*ne02 + head)*gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]);
     }
 #else
-    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
+    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); GGML_UNUSED(sinks);
     GGML_UNUSED(dst); GGML_UNUSED(dst_meta);
     GGML_UNUSED(scale); GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
     GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
diff --git a/ggml/src/ggml-cuda/fattn-vec-f32.cuh b/ggml/src/ggml-cuda/fattn-vec-f32.cuh
index 9ab0fc133b7..d6d0bfb744b 100644
--- a/ggml/src/ggml-cuda/fattn-vec-f32.cuh
+++ b/ggml/src/ggml-cuda/fattn-vec-f32.cuh
@@ -16,6 +16,7 @@ static __global__ void flash_attn_vec_ext_f32(
         const char * __restrict__ K,
         const char * __restrict__ V,
         const char * __restrict__ mask,
+        const char * __restrict__ sinks,
         const int  * __restrict__ KV_max,
         float      * __restrict__ dst,
         float2     * __restrict__ dst_meta,
@@ -72,7 +73,8 @@ static __global__ void flash_attn_vec_ext_f32(
     K += nb13*sequence + nb12*(head / gqa_ratio);
     V += nb23*sequence + nb22*(head / gqa_ratio);
 
-    const half * maskh = (const half *) (mask + nb33*(sequence % ne33) + nb31*ic0);
+    const half  * maskh  = (const half  *) (mask + nb33*(sequence % ne33) + nb31*ic0);
+    const float * sinksf = (const float *) (sinks);
 
     const float slope = get_alibi_slope(max_bias, head, n_head_log2, m0, m1);
 
@@ -88,11 +90,12 @@ static __global__ void flash_attn_vec_ext_f32(
     }
 
     float kqmax[ncols];
+    float kqsum[ncols];
 #pragma unroll
     for (int j = 0; j < ncols; ++j) {
         kqmax[j] = -FLT_MAX/2.0f;
+        kqsum[j] = 0.0f;
     }
-    float kqsum[ncols] = {0.0f};
 
     __shared__ float kqmax_shared[ncols][WARP_SIZE];
     __shared__ float kqsum_shared[ncols][WARP_SIZE];
@@ -279,6 +282,39 @@ static __global__ void flash_attn_vec_ext_f32(
         __syncthreads();
     }
 
+    if (sinksf && blockIdx.y == 0) {
+        const float sink = sinksf[head];
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            if (threadIdx.x == 0) {
+                kqmax_shared[j][threadIdx.y] = fmaxf(kqmax[j], sink);
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            float kqmax_new_j = kqmax_shared[j][threadIdx.x];
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+
+            const float KQ_max_scale = expf(kqmax[j] - kqmax_new_j);
+            kqmax[j] = kqmax_new_j;
+
+            const float val = expf(sink - kqmax[j]);
+            kqsum[j] = kqsum[j]*KQ_max_scale;
+
+            if (tid == 0) {
+                kqsum[j] += val;
+            }
+
+            VKQ[j] *= KQ_max_scale;
+        }
+
+        __syncthreads();
+    }
+
 #pragma unroll
     for (int j = 0; j < ncols; ++j) {
         kqsum[j] = warp_reduce_sum(kqsum[j]);
diff --git a/ggml/src/ggml-cuda/fattn-wmma-f16.cu b/ggml/src/ggml-cuda/fattn-wmma-f16.cu
index 40554204b62..93d4d810218 100644
--- a/ggml/src/ggml-cuda/fattn-wmma-f16.cu
+++ b/ggml/src/ggml-cuda/fattn-wmma-f16.cu
@@ -29,6 +29,7 @@ static __global__ void flash_attn_ext_f16(
         const char * __restrict__ K,
         const char * __restrict__ V,
         const char * __restrict__ mask,
+        const char * __restrict__ sinks,
         const int  * __restrict__ KV_max,
         float      * __restrict__ dst,
         float2     * __restrict__ dst_meta,
@@ -423,7 +424,7 @@ static __global__ void flash_attn_ext_f16(
         dst_meta[j_dst_unrolled] = dst_meta_val;
     }
 #else
-    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
+    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); GGML_UNUSED(sinks);
     GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
     GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
     GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu
index 039c54e015e..656e04a4735 100644
--- a/ggml/src/ggml-cuda/fattn.cu
+++ b/ggml/src/ggml-cuda/fattn.cu
@@ -269,17 +269,28 @@ static void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, gg
 }
 
 void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * KQV  = dst;
-    const ggml_tensor * Q    = dst->src[0];
-    const ggml_tensor * K    = dst->src[1];
-    const ggml_tensor * V    = dst->src[2];
-    const ggml_tensor * mask = dst->src[3];
+    const ggml_tensor * KQV   = dst;
+    const ggml_tensor * Q     = dst->src[0];
+    const ggml_tensor * K     = dst->src[1];
+    const ggml_tensor * V     = dst->src[2];
+    const ggml_tensor * mask  = dst->src[3];
+    const ggml_tensor * sinks = dst->src[4];
 
     ggml_cuda_set_device(ctx.device);
     const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
     const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size;
     const enum ggml_prec prec = ggml_flash_attn_ext_get_prec(KQV);
 
+    // TODO: currently only vec implementation for sinks is supported [TAG_ATTN_SINKS]
+    if (sinks) {
+        if (prec == GGML_PREC_DEFAULT && fast_fp16_available(cc)) {
+            ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
+        } else {
+            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+        }
+        return;
+    }
+
 #if defined(GGML_HIP_ROCWMMA_FATTN)
     if (GGML_CUDA_CC_IS_AMD(cc) && fp16_mma_available(cc)) {
         ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index 8885fb7fbdd..60e481b95af 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -4,6 +4,7 @@
 
 #include "ggml-cuda/common.cuh"
 #include "ggml-cuda/acc.cuh"
+#include "ggml-cuda/add-id.cuh"
 #include "ggml-cuda/arange.cuh"
 #include "ggml-cuda/argmax.cuh"
 #include "ggml-cuda/argsort.cuh"
@@ -2259,6 +2260,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_ADD1: // TODO: more efficient implementation
             ggml_cuda_op_add(ctx, dst);
             break;
+        case GGML_OP_ADD_ID:
+            ggml_cuda_op_add_id(ctx, dst);
+            break;
         case GGML_OP_SUB:
             ggml_cuda_op_sub(ctx, dst);
             break;
@@ -2333,6 +2337,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
                 case GGML_GLU_OP_SWIGLU:
                     ggml_cuda_op_swiglu(ctx, dst);
                     break;
+                case GGML_GLU_OP_SWIGLU_OAI:
+                    ggml_cuda_op_swiglu_oai(ctx, dst);
+                    break;
                 case GGML_GLU_OP_GEGLU_ERF:
                     ggml_cuda_op_geglu_erf(ctx, dst);
                     break;
@@ -2607,6 +2614,9 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
 
     const std::string gemma3n_per_layer_proj_src0_name = "inp_per_layer_selected";
     const std::string gemma3n_per_layer_proj_src1_name = "per_layer_proj";
+    const std::string ffn_moe_gate_bias_prefix = "ffn_moe_gate_biased";
+    const std::string ffn_moe_up_bias_prefix = "ffn_moe_up_biased";
+    const std::string ffn_moe_down_bias_prefix = "ffn_moe_down_biased";
 
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_tensor * node = cgraph->nodes[i];
@@ -2629,7 +2639,13 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
 #endif
         }
 
-        if (node->op == GGML_OP_ADD && node->src[1] && node->src[1]->ne[1] > 1 && (node->src[0] ? node->src[0]->name != gemma3n_per_layer_proj_src0_name : true) && (node->src[1] ? node->src[1]->name != gemma3n_per_layer_proj_src1_name : true)) {
+        if (node->op == GGML_OP_ADD &&
+            node->src[1] && node->src[1]->ne[1] > 1 &&
+            (node->src[0] ? node->src[0]->name != gemma3n_per_layer_proj_src0_name : true) &&
+            (node->src[1] ? node->src[1]->name != gemma3n_per_layer_proj_src1_name : true) &&
+            strncmp(node->name, ffn_moe_gate_bias_prefix.c_str(), ffn_moe_gate_bias_prefix.size()) != 0 &&
+            strncmp(node->name, ffn_moe_up_bias_prefix.c_str(), ffn_moe_up_bias_prefix.size()) != 0 &&
+            strncmp(node->name, ffn_moe_down_bias_prefix.c_str(), ffn_moe_down_bias_prefix.size()) != 0) {
             // disable CUDA graphs for batch size > 1 for now while excluding the matrix-matrix addition as part of Gemma3n's `project_per_layer_input` operation
             // by means of matching node names. See
             // https://github.com/ggml-org/llama.cpp/blob/f9a31eea06a859e34cecb88b4d020c7f03d86cc4/src/llama-model.cpp#L10199-L10241 and
@@ -3227,6 +3243,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_GLU_OP_REGLU:
                 case GGML_GLU_OP_GEGLU:
                 case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_SWIGLU_OAI:
                 case GGML_GLU_OP_GEGLU_ERF:
                 case GGML_GLU_OP_GEGLU_QUICK:
                     return ggml_is_contiguous_1(op->src[0]);
@@ -3277,6 +3294,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                     case GGML_TYPE_Q5_0:
                     case GGML_TYPE_Q5_1:
                     case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_MXFP4:
                     case GGML_TYPE_Q2_K:
                     case GGML_TYPE_Q3_K:
                     case GGML_TYPE_Q4_K:
@@ -3423,6 +3441,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_PERMUTE:
         case GGML_OP_TRANSPOSE:
         case GGML_OP_ADD:
+        case GGML_OP_ADD_ID:
         case GGML_OP_ADD1:
         case GGML_OP_SUB:
         case GGML_OP_MUL:
@@ -3503,6 +3522,10 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 const int gqa_ratio = op->src[0]->ne[2] / op->src[1]->ne[2];
                 return op->src[1]->ne[0] == 576 && op->src[2]->ne[0] == 512 && op->src[3] && gqa_ratio % 16 == 0;
             }
+            // TODO: more general-purpose attention sink support [TAG_ATTN_SINKS]
+            if (op->src[4] && op->src[0]->ne[0] != 64 && op->src[0]->ne[0] != 128) { // currently only sinks for head_size 64 and 128 are supported
+                return false;
+            }
             if (op->src[0]->ne[0] == 192) {
                 return false;
             }
diff --git a/ggml/src/ggml-cuda/im2col.cu b/ggml/src/ggml-cuda/im2col.cu
index 73b98133438..16bb9bec97d 100644
--- a/ggml/src/ggml-cuda/im2col.cu
+++ b/ggml/src/ggml-cuda/im2col.cu
@@ -1,7 +1,5 @@
 #include "im2col.cuh"
 
-#define MIN(a, b) (a) < (b) ? (a) : (b)
-
 #define MAX_GRIDDIM_Z 65535
 
 template <typename T>
@@ -38,6 +36,9 @@ static  __global__ void im2col_kernel(
             dst[offset_dst] = x[offset_src + iih * IW + iiw];
         }
     }
+
+    GGML_UNUSED(IC);
+    GGML_UNUSED(KH);
 }
 
 // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW]
diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cu
index d4954fbe69e..8954a383104 100644
--- a/ggml/src/ggml-cuda/mmq.cu
+++ b/ggml/src/ggml-cuda/mmq.cu
@@ -20,6 +20,9 @@ static void ggml_cuda_mul_mat_q_switch_type(ggml_backend_cuda_context & ctx, con
         case GGML_TYPE_Q8_0:
             mul_mat_q_case<GGML_TYPE_Q8_0>(ctx, args, stream);
             break;
+        case GGML_TYPE_MXFP4:
+            mul_mat_q_case<GGML_TYPE_MXFP4>(ctx, args, stream);
+            break;
         case GGML_TYPE_Q2_K:
             mul_mat_q_case<GGML_TYPE_Q2_K>(ctx, args, stream);
             break;
@@ -282,6 +285,7 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
         case GGML_TYPE_Q5_0:
         case GGML_TYPE_Q5_1:
         case GGML_TYPE_Q8_0:
+        case GGML_TYPE_MXFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh
index 04a8d80e121..1634725c20a 100644
--- a/ggml/src/ggml-cuda/mmq.cuh
+++ b/ggml/src/ggml-cuda/mmq.cuh
@@ -58,6 +58,8 @@ static mmq_q8_1_ds_layout mmq_get_q8_1_ds_layout(const ggml_type type_x) {
             return MMQ_Q8_1_DS_LAYOUT_DS4;
         case GGML_TYPE_Q8_0:
             return MMQ_Q8_1_DS_LAYOUT_D4;
+        case GGML_TYPE_MXFP4:
+            return MMQ_Q8_1_DS_LAYOUT_D4;
         case GGML_TYPE_Q2_K:
             return MMQ_Q8_1_DS_LAYOUT_D2S6;
         case GGML_TYPE_Q3_K:
@@ -170,6 +172,7 @@ static constexpr __host__ __device__ tile_x_sizes mmq_get_dp4a_tile_x_sizes(ggml
         case GGML_TYPE_Q5_0:    return MMQ_DP4A_TXS_Q8_0;
         case GGML_TYPE_Q5_1:    return MMQ_DP4A_TXS_Q8_1;
         case GGML_TYPE_Q8_0:    return MMQ_DP4A_TXS_Q8_0;
+        case GGML_TYPE_MXFP4:   return MMQ_DP4A_TXS_Q8_1;
         case GGML_TYPE_Q2_K:    return MMQ_DP4A_TXS_Q2_K;
         case GGML_TYPE_Q3_K:    return MMQ_DP4A_TXS_Q3_K;
         case GGML_TYPE_Q4_K:    return MMQ_DP4A_TXS_Q4_K;
@@ -206,6 +209,7 @@ static constexpr __host__ __device__ int mmq_get_mma_tile_x_k(ggml_type type) {
         case GGML_TYPE_Q5_0:    return MMQ_MMA_TILE_X_K_Q8_0;
         case GGML_TYPE_Q5_1:    return MMQ_MMA_TILE_X_K_Q8_1;
         case GGML_TYPE_Q8_0:    return MMQ_MMA_TILE_X_K_Q8_0;
+        case GGML_TYPE_MXFP4:   return MMQ_MMA_TILE_X_K_Q8_1;
         case GGML_TYPE_Q2_K:    return MMQ_MMA_TILE_X_K_Q2_K;
         case GGML_TYPE_Q3_K:    return MMQ_MMA_TILE_X_K_Q3_K;
         case GGML_TYPE_Q4_K:    return MMQ_MMA_TILE_X_K_Q8_1;
@@ -692,6 +696,71 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     }
 }
 
+template <int mmq_y, bool need_check> static __device__ __forceinline__ void load_tiles_mxfp4(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+    constexpr int nwarps = mmq_get_nwarps_device();
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+
+#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_MXFP4, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+
+    constexpr int threads_per_row = MMQ_ITER_K / (4 * QR_MXFP4);
+    constexpr int nrows = warp_size / threads_per_row;
+    const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x;
+    const int kbx  = txi / QI_MXFP4;
+    const int kqsx = txi % QI_MXFP4;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) {
+        int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row);
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_mxfp4 * bxi = (const block_mxfp4 *) x + kbx0 + i*stride + kbx;
+
+        const int aux_q4 = get_int_b1(bxi->qs, kqsx);
+        const int2 v = get_int_from_table_16(aux_q4, kvalues_mxfp4);
+        const int k0 = kbx * (2 * QI_MXFP4) + kqsx;
+
+#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + k0 + 0]        = v.x;
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + k0 + QI_MXFP4] = v.y;
+#else
+        x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0]        = v.x;
+        x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + QI_MXFP4] = v.y;
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+    }
+
+    constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI_MXFP4;
+    constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row;
+    const int kbxd = threadIdx.x % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * rows_per_warp) {
+        int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_mxfp4 * bxi = (const block_mxfp4 *) x + kbx0 + i*stride + kbxd;
+
+#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_1                 + kbxd] = ggml_cuda_e8m0_to_fp32(bxi->e)*0.5f;
+#else
+        x_df[i*(MMQ_TILE_NE_K/QI_MXFP4) + i/QI_MXFP4 + kbxd] = ggml_cuda_e8m0_to_fp32(bxi->e)*0.5f;
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+    }
+}
+
 template <int mmq_x, int mmq_y>
 static __device__ __forceinline__ void vec_dot_q8_0_q8_1_dp4a(
     const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
@@ -2268,7 +2337,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         const block_iq4_nl * bxi = (const block_iq4_nl *) x + kbx0 + i*stride + kbx;
 
         const int aux_q4 = get_int_b2(bxi->qs, kqsx);
-        const int2 v = get_int_from_table_16(aux_q4);
+        const int2 v = get_int_from_table_16(aux_q4, kvalues_iq4nl);
         const int k0 = kbx * (2 * QI4_NL) + kqsx;
 
 #if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
@@ -2707,7 +2776,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         const block_iq4_xs * bxi = (const block_iq4_xs *) x + kbx0 + i*stride;
 
         const int aux_q4 = get_int_b4(bxi->qs, kqsx);
-        const int2 v = get_int_from_table_16(aux_q4);
+        const int2 v = get_int_from_table_16(aux_q4, kvalues_iq4nl);
         const int k0 = 8 * (kqsx / 4) + kqsx % 4;
 
 #if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
@@ -2863,6 +2932,14 @@ struct mmq_type_traits<mmq_x, mmq_y, need_check, GGML_TYPE_Q8_0> {
     static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a<mmq_x, mmq_y>;
 };
 
+template <int mmq_x, int mmq_y, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, need_check, GGML_TYPE_MXFP4> {
+    static constexpr int              vdr          = VDR_MXFP4_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_mxfp4<mmq_y, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, MMQ_Q8_1_DS_LAYOUT_D4>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a<mmq_x, mmq_y>;
+};
+
 template <int mmq_x, int mmq_y, bool need_check>
 struct mmq_type_traits<mmq_x, mmq_y, need_check, GGML_TYPE_Q2_K> {
     static constexpr int              vdr          = VDR_Q2_K_Q8_1_MMQ;
@@ -3642,6 +3719,7 @@ extern DECL_MMQ_CASE(GGML_TYPE_Q4_1);
 extern DECL_MMQ_CASE(GGML_TYPE_Q5_0);
 extern DECL_MMQ_CASE(GGML_TYPE_Q5_1);
 extern DECL_MMQ_CASE(GGML_TYPE_Q8_0);
+extern DECL_MMQ_CASE(GGML_TYPE_MXFP4);
 extern DECL_MMQ_CASE(GGML_TYPE_Q2_K);
 extern DECL_MMQ_CASE(GGML_TYPE_Q3_K);
 extern DECL_MMQ_CASE(GGML_TYPE_Q4_K);
diff --git a/ggml/src/ggml-cuda/mmvq.cu b/ggml/src/ggml-cuda/mmvq.cu
index dc7adf509fa..5c8e5c4a7ee 100644
--- a/ggml/src/ggml-cuda/mmvq.cu
+++ b/ggml/src/ggml-cuda/mmvq.cu
@@ -13,6 +13,7 @@ static constexpr __device__ vec_dot_q_cuda_t get_vec_dot_q_cuda(ggml_type type)
         case GGML_TYPE_Q5_0:    return vec_dot_q5_0_q8_1;
         case GGML_TYPE_Q5_1:    return vec_dot_q5_1_q8_1;
         case GGML_TYPE_Q8_0:    return vec_dot_q8_0_q8_1;
+        case GGML_TYPE_MXFP4:   return vec_dot_mxfp4_q8_1;
         case GGML_TYPE_Q2_K:    return vec_dot_q2_K_q8_1;
         case GGML_TYPE_Q3_K:    return vec_dot_q3_K_q8_1;
         case GGML_TYPE_Q4_K:    return vec_dot_q4_K_q8_1;
@@ -38,6 +39,7 @@ static constexpr __device__ int get_vdr_mmvq(ggml_type type) {
         case GGML_TYPE_Q5_0:    return VDR_Q5_0_Q8_1_MMVQ;
         case GGML_TYPE_Q5_1:    return VDR_Q5_1_Q8_1_MMVQ;
         case GGML_TYPE_Q8_0:    return VDR_Q8_0_Q8_1_MMVQ;
+        case GGML_TYPE_MXFP4:   return VDR_MXFP4_Q8_1_MMVQ;
         case GGML_TYPE_Q2_K:    return VDR_Q2_K_Q8_1_MMVQ;
         case GGML_TYPE_Q3_K:    return VDR_Q3_K_Q8_1_MMVQ;
         case GGML_TYPE_Q4_K:    return VDR_Q4_K_Q8_1_MMVQ;
@@ -384,6 +386,13 @@ static void mul_mat_vec_q_switch_type(
                  nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
                  stream);
             break;
+        case GGML_TYPE_MXFP4:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_MXFP4>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
         case GGML_TYPE_Q2_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q2_K>
                 (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
diff --git a/ggml/src/ggml-cuda/softmax.cu b/ggml/src/ggml-cuda/softmax.cu
index 14543e978cf..eeacde0bdb1 100644
--- a/ggml/src/ggml-cuda/softmax.cu
+++ b/ggml/src/ggml-cuda/softmax.cu
@@ -45,7 +45,7 @@ struct soft_max_params {
 #endif // __clang__
 template <bool use_shared, int ncols_template, int block_size_template, typename T>
 static __global__ void soft_max_f32(
-        const float * x, const T * mask, float * dst, const soft_max_params p) {
+        const float * x, const T * mask, const float * sinks, float * dst, const soft_max_params p) {
     const int ncols = ncols_template == 0 ? p.ncols : ncols_template;
 
     const int tid  = threadIdx.x;
@@ -77,7 +77,7 @@ static __global__ void soft_max_f32(
     // shared memory buffer to cache values between iterations:
     float * vals = use_shared ? buf_iw + WARP_SIZE : dst;
 
-    float max_val = -INFINITY;
+    float max_val = sinks ? sinks[i02] : -INFINITY;
 
 #pragma unroll
     for (int col0 = 0; col0 < ncols; col0 += block_size) {
@@ -143,6 +143,10 @@ static __global__ void soft_max_f32(
         tmp = warp_reduce_sum(tmp);
     }
 
+    if (sinks) {
+        tmp += expf(sinks[i02] - max_val);
+    }
+
     const float inv_sum = 1.0f / tmp;
 
 #pragma unroll
@@ -183,7 +187,7 @@ static __global__ void soft_max_back_f32(
 }
 
 template<int... Ns, typename T>
-static void launch_soft_max_kernels(const float * x, const T * mask, float * dst,
+static void launch_soft_max_kernels(const float * x, const T * mask, const float * sinks, float * dst,
                              const soft_max_params & p, cudaStream_t stream, dim3 block_dims, dim3 block_nums, size_t nbytes_shared)
 {
     const int id       = ggml_cuda_get_device();
@@ -196,7 +200,7 @@ static void launch_soft_max_kernels(const float * x, const T * mask, float * dst
         if (p.ncols == ncols) {
             CUDA_SET_SHARED_MEMORY_LIMIT((soft_max_f32<true, ncols, block, T>), smpbo);
             soft_max_f32<true, ncols, block><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, mask, dst, p);
+                (x, mask, sinks, dst, p);
             return true;
         }
         return false;
@@ -209,12 +213,12 @@ static void launch_soft_max_kernels(const float * x, const T * mask, float * dst
 
     //default case
     CUDA_SET_SHARED_MEMORY_LIMIT((soft_max_f32<true, 0, 0, T>), smpbo);
-    soft_max_f32<true, 0, 0><<<block_nums, block_dims, nbytes_shared, stream>>>(x, mask, dst, p);
+    soft_max_f32<true, 0, 0><<<block_nums, block_dims, nbytes_shared, stream>>>(x, mask, sinks, dst, p);
 }
 
 
 template<typename T>
-static void soft_max_f32_cuda(const float * x, const T * mask, float * dst, const soft_max_params & params, cudaStream_t stream) {
+static void soft_max_f32_cuda(const float * x, const T * mask, const float * sinks, float * dst, const soft_max_params & params, cudaStream_t stream) {
     int nth = WARP_SIZE;
     const int64_t ncols_x = params.ncols;
 
@@ -230,10 +234,10 @@ static void soft_max_f32_cuda(const float * x, const T * mask, float * dst, cons
 
 
     if (nbytes_shared <= smpbo) {
-        launch_soft_max_kernels<32, 64, 128, 256, 512, 1024, 2048, 4096>(x, mask, dst, params, stream, block_dims, block_nums, nbytes_shared);
+        launch_soft_max_kernels<32, 64, 128, 256, 512, 1024, 2048, 4096>(x, mask, sinks, dst, params, stream, block_dims, block_nums, nbytes_shared);
     } else {
         const size_t nbytes_shared_low = WARP_SIZE*sizeof(float);
-        soft_max_f32<false, 0, 0><<<block_nums, block_dims, nbytes_shared_low, stream>>>(x, mask, dst, params);
+        soft_max_f32<false, 0, 0><<<block_nums, block_dims, nbytes_shared_low, stream>>>(x, mask, sinks, dst, params);
     }
 }
 
@@ -249,9 +253,11 @@ static void soft_max_back_f32_cuda(
 void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
+    const ggml_tensor * src2 = dst->src[2];
 
     const float * src0_d = (const float *) src0->data;
     const void  * src1_d = src1 ? (const void *) src1->data : nullptr;
+    const void  * src2_d = src2 ? (const void *) src2->data : nullptr;
     float       *  dst_d = (float *) dst->data;
 
     cudaStream_t stream = ctx.stream();
@@ -309,9 +315,9 @@ void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     params.m1 = m1;
 
     if (use_f16) {
-        soft_max_f32_cuda(src0_d, (const half  *) src1_d, dst_d, params, stream);
+        soft_max_f32_cuda(src0_d, (const half  *) src1_d, (const float *) src2_d, dst_d, params, stream);
     } else {
-        soft_max_f32_cuda(src0_d, (const float *) src1_d, dst_d, params, stream);
+        soft_max_f32_cuda(src0_d, (const float *) src1_d, (const float *) src2_d, dst_d, params, stream);
     }
 }
 
diff --git a/ggml/src/ggml-cuda/template-instances/mmq-instance-mxfp4.cu b/ggml/src/ggml-cuda/template-instances/mmq-instance-mxfp4.cu
new file mode 100644
index 00000000000..c14624c52ca
--- /dev/null
+++ b/ggml/src/ggml-cuda/template-instances/mmq-instance-mxfp4.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_MXFP4);
diff --git a/ggml/src/ggml-cuda/unary.cu b/ggml/src/ggml-cuda/unary.cu
index 91c830c4dac..5aff8a876af 100644
--- a/ggml/src/ggml-cuda/unary.cu
+++ b/ggml/src/ggml-cuda/unary.cu
@@ -300,6 +300,81 @@ void ggml_cuda_op_geglu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst
     ggml_cuda_op_unary_gated<op_gelu_quick>(ctx, dst);
 }
 
+// swiglu_oai
+
+template <typename T>
+static __global__ void swiglu_oai_kernel(const T * x, const T * g, T * dst, const int64_t k, const int64_t n, const int64_t o0, const int64_t o1, float alpha, float limit) {
+    const int64_t i = int64_t(blockDim.x)*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    // perform base op and multiply with gate (either offset in same tensor or a separate one)
+    const int64_t j0 = (i / n) * o0 + (i % n);
+    const int64_t j1 = o0 == o1 ? j0 : (i / n) * o1 + (i % n);
+
+    float xi = x[j0];
+    float gi = g[j1];
+    xi = fminf(xi, limit);
+    gi = fmaxf(fminf(gi, limit), -limit);
+
+    float out_glu = xi / (1.0f + expf(-xi * alpha));
+    out_glu = out_glu * (1.0f + gi);
+
+    dst[i] = out_glu;
+}
+
+template <typename T>
+static void swiglu_oai_cuda(const T * x, const T * g, T * dst, const int64_t k, const int64_t n, const int64_t o0, const int64_t o1, const float alpha, const float limit, cudaStream_t stream) {
+    const int64_t num_blocks = (k + CUDA_GLU_BLOCK_SIZE - 1) / CUDA_GLU_BLOCK_SIZE;
+    swiglu_oai_kernel<<<num_blocks, CUDA_GLU_BLOCK_SIZE, 0, stream>>>(x, g, dst, k, n, o0, o1, alpha, limit);
+}
+
+void ggml_cuda_op_swiglu_oai(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    void * src0_d = src0->data;
+    void * src1_d = src1 ? src1->data : src0->data;
+    const int64_t src0_o = src0->nb[1];
+    const int64_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+    void * dst_d = dst->data;
+    const int64_t nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(src0->nb[0] == ggml_element_size(src0));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->type == dst->type);
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_nrows(dst) == ggml_nrows(src0));
+
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src1->nb[0] == ggml_element_size(src1));
+        GGML_ASSERT(src1->ne[0] == nc);
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    //const int32_t swapped = ((const int32_t *) dst->op_params)[1];
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+    const float alpha = ggml_get_op_params_f32(dst, 2);
+    const float limit = ggml_get_op_params_f32(dst, 3);
+
+    float * src0_p = (float *) src0_d;
+    float * src1_p = (float *) src1_d;
+
+    if (!src1) {
+        src0_p += swapped ? nc : 0;
+        src1_p += swapped ? 0 : nc;
+    }
+
+    swiglu_oai_cuda(src0_p, src1_p, (float *)dst_d, ggml_nelements(dst), nc, src0_o / sizeof(float), src1_o / sizeof(float), alpha, limit, stream);
+}
+
 /* silu_back */
 
 static __device__ __forceinline__ float op_silu_back(float grad, float x) {
diff --git a/ggml/src/ggml-cuda/unary.cuh b/ggml/src/ggml-cuda/unary.cuh
index cb14d16f8f3..da3caf1d896 100644
--- a/ggml/src/ggml-cuda/unary.cuh
+++ b/ggml/src/ggml-cuda/unary.cuh
@@ -67,6 +67,8 @@ void ggml_cuda_op_geglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_swiglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
+void ggml_cuda_op_swiglu_oai(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
 void ggml_cuda_op_geglu_erf(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_geglu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/ggml/src/ggml-cuda/vecdotq.cuh b/ggml/src/ggml-cuda/vecdotq.cuh
index ba195e1d100..d8f9aa5ba62 100644
--- a/ggml/src/ggml-cuda/vecdotq.cuh
+++ b/ggml/src/ggml-cuda/vecdotq.cuh
@@ -1,8 +1,20 @@
 #pragma once
 
 #include "common.cuh"
+
 #include <cstdint>
 
+static __device__ __forceinline__ int get_int_b1(const void * x, const int & i32) {
+    const uint8_t * x8 = (const uint8_t *) x;
+
+    int x32  = x8[4*i32 + 0] <<  0;
+    x32     |= x8[4*i32 + 1] <<  8;
+    x32     |= x8[4*i32 + 2] << 16;
+    x32     |= x8[4*i32 + 3] << 24;
+
+    return x32;
+}
+
 static __device__ __forceinline__ int get_int_b2(const void * x, const int & i32) {
     const uint16_t * x16 = (const uint16_t *) x; // assume at least 2 byte alignment
 
@@ -16,6 +28,20 @@ static __device__ __forceinline__ int get_int_b4(const void * x, const int & i32
     return ((const int *) x)[i32]; // assume at least 4 byte alignment
 }
 
+static __device__ __forceinline__ int2 get_int_from_table_16(const int & q4, const int8_t * table) {
+    const int      q0_32  = (q4 >> 0) & 0x0F0F0F0F;
+    const int8_t * q0_8   = (const int8_t *) &q0_32;
+    const char4    val0_8 = make_char4(
+        table[q0_8[0]], table[q0_8[1]], table[q0_8[2]], table[q0_8[3]]);
+
+    const int      q1_32  = (q4 >> 4) & 0x0F0F0F0F;
+    const int8_t * q1_8   = (const int8_t *) &q1_32;
+    const char4    val1_8 = make_char4(
+        table[q1_8[0]], table[q1_8[1]], table[q1_8[2]], table[q1_8[3]]);
+
+    return make_int2(*((const int *) &val0_8), *((const int *) &val1_8));
+}
+
 // VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called
 // MMVQ = mul_mat_vec_q, MMQ = mul_mat_q
 
@@ -211,6 +237,30 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q8_0_16_q8_1_
     return d8_1*sumf;
 }
 
+#define VDR_MXFP4_Q8_1_MMVQ 2
+#define VDR_MXFP4_Q8_1_MMQ  4
+
+static __device__ __forceinline__ float vec_dot_mxfp4_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_mxfp4 * bq4 = (const block_mxfp4 *) vbq + kbx;
+
+    const int * q8 = (const int *) bq8_1->qs + iqs;
+
+    int sumi = 0;
+#pragma unroll
+    for (int l = 0; l < VDR_MXFP4_Q8_1_MMVQ; ++l) {
+        const int aux_q4 = get_int_b1(bq4->qs, iqs + l);
+        const int2 v = get_int_from_table_16(aux_q4, kvalues_mxfp4);
+
+        sumi = ggml_cuda_dp4a(v.x, q8[l + 0], sumi);
+        sumi = ggml_cuda_dp4a(v.y, q8[l + 4], sumi);
+    }
+
+    const float d = ggml_cuda_e8m0_to_fp32(bq4->e) * 0.5f * __low2float(bq8_1->ds);
+    return d * sumi;
+}
+
 #define VDR_Q2_K_Q8_1_MMVQ 1
 #define VDR_Q2_K_Q8_1_MMQ  4
 
@@ -1068,20 +1118,6 @@ static __device__ __forceinline__ float vec_dot_iq1_m_q8_1(
     return d * ((sumi[0] + sumf[0]) * sc0 + (sumi[1] + sumf[1]) * sc1);
 }
 
-static __device__ __forceinline__ int2 get_int_from_table_16(const int & q4) {
-    const int      q0_32  = (q4 >> 0) & 0x0F0F0F0F;
-    const int8_t * q0_8   = (const int8_t *) &q0_32;
-    const char4    val0_8 = make_char4(
-        kvalues_iq4nl[q0_8[0]], kvalues_iq4nl[q0_8[1]], kvalues_iq4nl[q0_8[2]], kvalues_iq4nl[q0_8[3]]);
-
-    const int      q1_32  = (q4 >> 4) & 0x0F0F0F0F;
-    const int8_t * q1_8   = (const int8_t *) &q1_32;
-    const char4    val1_8 = make_char4(
-        kvalues_iq4nl[q1_8[0]], kvalues_iq4nl[q1_8[1]], kvalues_iq4nl[q1_8[2]], kvalues_iq4nl[q1_8[3]]);
-
-    return make_int2(*((const int *) &val0_8), *((const int *) &val1_8));
-}
-
 #define VDR_IQ4_NL_Q8_1_MMVQ 2
 #define VDR_IQ4_NL_Q8_1_MMQ  4
 
@@ -1096,7 +1132,7 @@ static __device__ __forceinline__ float vec_dot_iq4_nl_q8_1(
 #pragma unroll
     for (int l = 0; l < VDR_Q4_0_Q8_1_MMVQ; ++l) {
         const int aux_q4 = get_int_b2(bq4->qs, iqs + l);
-        const int2 v = get_int_from_table_16(aux_q4);
+        const int2 v = get_int_from_table_16(aux_q4, kvalues_iq4nl);
 
         sumi = ggml_cuda_dp4a(v.x, q8[l + 0], sumi);
         sumi = ggml_cuda_dp4a(v.y, q8[l + 4], sumi);
@@ -1118,7 +1154,7 @@ static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1(
 #pragma unroll
     for (int j = 0; j < 4; ++j) {
         const int aux_q4 = get_int_b4(bq4->qs, iqs + j);
-        const int2 v = get_int_from_table_16(aux_q4);
+        const int2 v = get_int_from_table_16(aux_q4, kvalues_iq4nl);
 
         const int u0 = get_int_b4(bq8_1[iqs/4].qs, j + 0);
         const int u1 = get_int_b4(bq8_1[iqs/4].qs, j + 4);
diff --git a/ggml/src/ggml-cuda/vendors/cuda.h b/ggml/src/ggml-cuda/vendors/cuda.h
index 1746b073203..3b3086778ee 100644
--- a/ggml/src/ggml-cuda/vendors/cuda.h
+++ b/ggml/src/ggml-cuda/vendors/cuda.h
@@ -6,6 +6,10 @@
 #include <cuda_bf16.h>
 #include <cuda_fp16.h>
 
+#if CUDART_VERSION >= 12050
+#include <cuda_fp8.h>
+#endif // CUDART_VERSION >= 12050
+
 #if CUDART_VERSION < 11020
 #define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED
 #define CUBLAS_TF32_TENSOR_OP_MATH CUBLAS_TENSOR_OP_MATH
diff --git a/ggml/src/ggml-impl.h b/ggml/src/ggml-impl.h
index a2e30994c46..19a7adb2d10 100644
--- a/ggml/src/ggml-impl.h
+++ b/ggml/src/ggml-impl.h
@@ -410,6 +410,67 @@ static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
 #define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
 #define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
 
+static inline float ggml_e8m0_to_fp32(uint8_t x) {
+    uint32_t bits;  // Stores the raw bit representation of the float
+
+    // Handle special case for minimum exponent (denormalized float)
+    if (x == 0) {
+        // Bit pattern for 2^(-127):
+        // - Sign bit: 0 (positive)
+        // - Exponent: 0 (denormalized number)
+        // - Mantissa: 0x400000 (0.5 in fractional form)
+        // Value = 0.5 * 2^(-126) = 2^(-127)
+        bits = 0x00400000;
+    }
+    // note: disabled as we don't need to handle NaNs
+    //// Handle special case for NaN (all bits set)
+    //else if (x == 0xFF) {
+    //    // Standard quiet NaN pattern:
+    //    // - Sign bit: 0
+    //    // - Exponent: all 1s (0xFF)
+    //    // - Mantissa: 0x400000 (quiet NaN flag)
+    //    bits = 0x7FC00000;
+    //}
+    // Normalized values (most common case)
+    else {
+        // Construct normalized float by shifting exponent into position:
+        // - Exponent field: 8 bits (positions 30-23)
+        // - Mantissa: 0 (implicit leading 1)
+        // Value = 2^(x - 127)
+        bits = (uint32_t) x << 23;
+    }
+
+    float result;  // Final float value
+                   // Safely reinterpret bit pattern as float without type-punning issues
+    memcpy(&result, &bits, sizeof(float));
+    return result;
+}
+
+// Equal to ggml_e8m0_to_fp32/2
+// Useful with MXFP4 quantization since the E0M2 values are doubled
+static inline float ggml_e8m0_to_fp32_half(uint8_t x) {
+    uint32_t bits;
+
+    // For x < 2: use precomputed denormal patterns
+    if (x < 2) {
+        // 0x00200000 = 2^(-128), 0x00400000 = 2^(-127)
+        bits = 0x00200000 << x;
+    }
+    // For x >= 2: normalized exponent adjustment
+    else {
+        // 0.5 * 2^(x-127) = 2^(x-128) = normalized with exponent (x-1)
+        bits = (uint32_t)(x - 1) << 23;
+    }
+    // Note: NaNs are not handled here
+
+    float result;
+    memcpy(&result, &bits, sizeof(float));
+    return result;
+}
+
+#define GGML_E8M0_TO_FP32(x) ggml_e8m0_to_fp32(x)
+#define GGML_E8M0_TO_FP32_HALF(x) ggml_e8m0_to_fp32_half(x)
+
 /**
  * Converts brain16 to float32.
  *
diff --git a/ggml/src/ggml-metal/ggml-metal-impl.h b/ggml/src/ggml-metal/ggml-metal-impl.h
index 8424464d8ca..fc6526d6d5d 100644
--- a/ggml/src/ggml-metal/ggml-metal-impl.h
+++ b/ggml/src/ggml-metal/ggml-metal-impl.h
@@ -23,6 +23,9 @@
 #define N_R0_Q8_0 4
 #define N_SG_Q8_0 2
 
+#define N_R0_MXFP4 2
+#define N_SG_MXFP4 2
+
 #define N_R0_Q2_K 4
 #define N_SG_Q2_K 2
 
@@ -129,6 +132,15 @@ typedef struct {
     uint64_t o1[8];
 } ggml_metal_kargs_bin;
 
+typedef struct {
+    int64_t ne0;
+    int64_t ne1;
+    size_t nb01;
+    size_t nb02;
+    size_t nb11;
+    size_t nb21;
+} ggml_metal_kargs_add_id;
+
 typedef struct {
     int32_t  ne00;
     int32_t  ne01;
@@ -444,6 +456,8 @@ typedef struct{
     uint64_t nb1;
     int32_t  i00;
     int32_t  i10;
+    float    alpha;
+    float    limit;
 } ggml_metal_kargs_glu;
 
 typedef struct {
diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
index 337f7985bad..cb8eff4a772 100644
--- a/ggml/src/ggml-metal/ggml-metal.m
+++ b/ggml/src/ggml-metal/ggml-metal.m
@@ -195,6 +195,7 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
     GGML_METAL_KERNEL_TYPE_MUL_ROW_C4,
     GGML_METAL_KERNEL_TYPE_DIV,
     GGML_METAL_KERNEL_TYPE_DIV_ROW_C4,
+    GGML_METAL_KERNEL_TYPE_ADD_ID,
     GGML_METAL_KERNEL_TYPE_REPEAT_F32,
     GGML_METAL_KERNEL_TYPE_REPEAT_F16,
     GGML_METAL_KERNEL_TYPE_REPEAT_I32,
@@ -234,6 +235,7 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
     GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_MXFP4,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K,
@@ -286,6 +288,7 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
     GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_MXFP4_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_2,
     GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_3,
     GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_4,
@@ -310,6 +313,10 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
     GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_3,
     GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_4,
     GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_5,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_2,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_3,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_4,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_5,
     GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_2,
     GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_3,
     GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_4,
@@ -351,6 +358,7 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_MXFP4_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32,
@@ -373,6 +381,7 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
     GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_MXFP4_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32,
@@ -397,6 +406,7 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F16,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F16,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MXFP4_F16,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F16,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F16,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F16,
@@ -579,6 +589,7 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
     GGML_METAL_KERNEL_TYPE_REGLU,
     GGML_METAL_KERNEL_TYPE_GEGLU,
     GGML_METAL_KERNEL_TYPE_SWIGLU,
+    GGML_METAL_KERNEL_TYPE_SWIGLU_OAI,
     GGML_METAL_KERNEL_TYPE_GEGLU_ERF,
     GGML_METAL_KERNEL_TYPE_GEGLU_QUICK,
     GGML_METAL_KERNEL_TYPE_SUM_ROWS,
@@ -1199,6 +1210,7 @@ @implementation GGMLMetalClass
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW_C4,                      mul_row_c4,                      true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV,                             div,                             true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW_C4,                      div_row_c4,                      true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ID,                          add_id,                          true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F32,                      repeat_f32,                      true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F16,                      repeat_f16,                      true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_I32,                      repeat_i32,                      true);
@@ -1238,6 +1250,7 @@ @implementation GGMLMetalClass
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0,                   get_rows_q5_0,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1,                   get_rows_q5_1,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0,                   get_rows_q8_0,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_MXFP4,                  get_rows_mxfp4,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K,                   get_rows_q2_K,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K,                   get_rows_q3_K,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K,                   get_rows_q4_K,                   true);
@@ -1290,6 +1303,7 @@ @implementation GGMLMetalClass
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32,                 mul_mv_q5_0_f32,                 has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32,                 mul_mv_q5_1_f32,                 has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32,                 mul_mv_q8_0_f32,                 has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_MXFP4_F32,                mul_mv_mxfp4_f32,                has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_2,         mul_mv_ext_f16_f32_r1_2,         has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_3,         mul_mv_ext_f16_f32_r1_3,         has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_4,         mul_mv_ext_f16_f32_r1_4,         has_simdgroup_reduction);
@@ -1314,6 +1328,10 @@ @implementation GGMLMetalClass
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_3,        mul_mv_ext_q8_0_f32_r1_3,        has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_4,        mul_mv_ext_q8_0_f32_r1_4,        has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_5,        mul_mv_ext_q8_0_f32_r1_5,        has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_2,       mul_mv_ext_mxfp4_f32_r1_2,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_3,       mul_mv_ext_mxfp4_f32_r1_3,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_4,       mul_mv_ext_mxfp4_f32_r1_4,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_5,       mul_mv_ext_mxfp4_f32_r1_5,       has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_2,        mul_mv_ext_q4_K_f32_r1_2,        has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_3,        mul_mv_ext_q4_K_f32_r1_3,        has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_4,        mul_mv_ext_q4_K_f32_r1_4,        has_simdgroup_reduction);
@@ -1355,6 +1373,7 @@ @implementation GGMLMetalClass
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32,              mul_mv_id_q5_0_f32,              has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32,              mul_mv_id_q5_1_f32,              has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32,              mul_mv_id_q8_0_f32,              has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_MXFP4_F32,             mul_mv_id_mxfp4_f32,             has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32,              mul_mv_id_q2_K_f32,              has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32,              mul_mv_id_q3_K_f32,              has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32,              mul_mv_id_q4_K_f32,              has_simdgroup_reduction);
@@ -1377,6 +1396,8 @@ @implementation GGMLMetalClass
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32,                 mul_mm_q5_0_f32,                 has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32,                 mul_mm_q5_1_f32,                 has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32,                 mul_mm_q8_0_f32,                 has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_MXFP4_F32,                mul_mm_mxfp4_f32,                has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_MXFP4_F32,                mul_mm_mxfp4_f32,                has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32,                 mul_mm_q2_K_f32,                 has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32,                 mul_mm_q3_K_f32,                 has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32,                 mul_mm_q4_K_f32,                 has_simdgroup_mm);
@@ -1401,6 +1422,7 @@ @implementation GGMLMetalClass
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F16,              mul_mm_id_q5_0_f16,              has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F16,              mul_mm_id_q5_1_f16,              has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F16,              mul_mm_id_q8_0_f16,              has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MXFP4_F16,             mul_mm_id_mxfp4_f16,             has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F16,              mul_mm_id_q2_K_f16,              has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F16,              mul_mm_id_q3_K_f16,              has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F16,              mul_mm_id_q4_K_f16,              has_simdgroup_mm);
@@ -1583,6 +1605,7 @@ @implementation GGMLMetalClass
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REGLU,                           reglu,                           true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GEGLU,                           geglu,                           true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SWIGLU,                          swiglu,                          true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SWIGLU_OAI,                      swiglu_oai,                      true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GEGLU_ERF,                       geglu_erf,                       true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GEGLU_QUICK,                     geglu_quick,                     true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS,                        sum_rows,                        true);
@@ -1774,6 +1797,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
                 case GGML_GLU_OP_REGLU:
                 case GGML_GLU_OP_GEGLU:
                 case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_SWIGLU_OAI:
                 case GGML_GLU_OP_GEGLU_ERF:
                 case GGML_GLU_OP_GEGLU_QUICK:
                     return ggml_is_contiguous_1(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
@@ -1791,6 +1815,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
         case GGML_OP_SUB:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
+        case GGML_OP_ADD_ID:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_ACC:
         case GGML_OP_REPEAT:
@@ -2042,6 +2067,7 @@ static int ggml_metal_encode_node(
 
     const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
     const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
+    const enum ggml_type src2t = src2 ? src2->type : GGML_TYPE_COUNT;
     const enum ggml_type dstt  = dst  ? dst->type  : GGML_TYPE_COUNT;
 
     size_t offs_src0 = 0;
@@ -2291,6 +2317,38 @@ static int ggml_metal_encode_node(
                     [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                 }
             } break;
+        case GGML_OP_ADD_ID:
+            {
+                GGML_ASSERT(src0t == GGML_TYPE_F32);
+                GGML_ASSERT(src1t == GGML_TYPE_F32);
+                GGML_ASSERT(src2t == GGML_TYPE_I32);
+                GGML_ASSERT(dstt  == GGML_TYPE_F32);
+
+                GGML_ASSERT(ggml_is_contiguous_rows(src0));
+
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_ID].pipeline;
+
+                ggml_metal_kargs_add_id args = {
+                    /*.ne0  =*/ ne0,
+                    /*.ne1  =*/ ne1,
+                    /*.nb01 =*/ nb01,
+                    /*.nb02 =*/ nb02,
+                    /*.nb11 =*/ nb11,
+                    /*.nb21 =*/ nb21,
+
+                };
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
+                [encoder setBuffer:id_src2 offset:offs_src2 atIndex:3];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:4];
+
+                const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne00);
+
+                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+            } break;
         case GGML_OP_REPEAT:
             {
                 id<MTLComputePipelineState> pipeline;
@@ -2710,6 +2768,9 @@ static int ggml_metal_encode_node(
                     case GGML_GLU_OP_SWIGLU:
                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SWIGLU].pipeline;
                         break;
+                    case GGML_GLU_OP_SWIGLU_OAI:
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SWIGLU_OAI].pipeline;
+                        break;
                     case GGML_GLU_OP_GEGLU_ERF:
                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GEGLU_ERF].pipeline;
                         break;
@@ -2720,7 +2781,9 @@ static int ggml_metal_encode_node(
                         GGML_ABORT("fatal error");
                 }
 
-                const int32_t swp = ((const int32_t *) dst->op_params)[1];
+                const int32_t swp = ggml_get_op_params_i32(dst, 1);
+                const float alpha = ggml_get_op_params_f32(dst, 2);
+                const float limit = ggml_get_op_params_f32(dst, 3);
 
                 const int32_t i00 = swp ? ne0 : 0;
                 const int32_t i10 = swp ? 0 : ne0;
@@ -2734,6 +2797,8 @@ static int ggml_metal_encode_node(
                     /*.nb1  =*/ nb1,
                     /*.i00  =*/ src1 ? 0 : i00,
                     /*.i10  =*/ src1 ? 0 : i10,
+                    /*.alpha=*/ alpha,
+                    /*.limit=*/ limit
                 };
 
                 [encoder setComputePipelineState:pipeline];
@@ -2992,8 +3057,13 @@ static int ggml_metal_encode_node(
                 } else {
                     [encoder setBuffer:h_src0 offset:offs_src0  atIndex:1];
                 }
-                [encoder setBuffer:id_dst offset:offs_dst       atIndex:2];
-                [encoder setBytes:&args   length:sizeof(args)   atIndex:3];
+                if (id_src2) {
+                    [encoder setBuffer:id_src2 offset:offs_src2 atIndex:2];
+                } else {
+                    [encoder setBuffer:h_src0 offset:offs_src0  atIndex:2];
+                }
+                [encoder setBuffer:id_dst offset:offs_dst       atIndex:3];
+                [encoder setBytes:&args   length:sizeof(args)   atIndex:4];
 
                 [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
 
@@ -3291,6 +3361,7 @@ static int ggml_metal_encode_node(
                        src0t == GGML_TYPE_Q5_0 ||
                        src0t == GGML_TYPE_Q5_1 ||
                        src0t == GGML_TYPE_Q8_0 ||
+                       src0t == GGML_TYPE_MXFP4 ||
                        src0t == GGML_TYPE_IQ4_NL ||
                        false) && (ne11 >= 2 && ne11 <= 8)
                      ) ||
@@ -3383,6 +3454,14 @@ static int ggml_metal_encode_node(
                                 case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_5].pipeline; break;
                                 default: GGML_ABORT("not implemented");
                             } break;
+                        case GGML_TYPE_MXFP4:
+                            switch (r1ptg) {
+                                case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_2].pipeline; break;
+                                case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_3].pipeline; break;
+                                case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_4].pipeline; break;
+                                case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_MXFP4_F32_R1_5].pipeline; break;
+                                default: GGML_ABORT("not implemented");
+                            } break;
                         case GGML_TYPE_Q4_K:
                             switch (r1ptg) {
                                 case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_2].pipeline; break;
@@ -3481,6 +3560,7 @@ static int ggml_metal_encode_node(
                         case GGML_TYPE_Q5_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32   ].pipeline; break;
                         case GGML_TYPE_Q5_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32   ].pipeline; break;
                         case GGML_TYPE_Q8_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32   ].pipeline; break;
+                        case GGML_TYPE_MXFP4:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_MXFP4_F32   ].pipeline; break;
                         case GGML_TYPE_Q2_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32   ].pipeline; break;
                         case GGML_TYPE_Q3_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32   ].pipeline; break;
                         case GGML_TYPE_Q4_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32   ].pipeline; break;
@@ -3623,6 +3703,13 @@ static int ggml_metal_encode_node(
                                 nr0 = N_R0_Q8_0;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32].pipeline;
                             } break;
+                        case GGML_TYPE_MXFP4:
+                            {
+                                nsg = N_SG_MXFP4;
+                                nr0 = N_R0_MXFP4;
+                                smem = 32*sizeof(float);
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_MXFP4_F32].pipeline;
+                            } break;
                         case GGML_TYPE_Q2_K:
                             {
                                 nsg = N_SG_Q2_K;
@@ -3756,8 +3843,6 @@ static int ggml_metal_encode_node(
         case GGML_OP_MUL_MAT_ID:
             {
                 // src2 = ids
-                const enum ggml_type src2t = src2->type; GGML_UNUSED(src2t);
-
                 GGML_ASSERT(src2t == GGML_TYPE_I32);
 
                 GGML_ASSERT(!ggml_is_transposed(src0));
@@ -3883,6 +3968,7 @@ static int ggml_metal_encode_node(
                             case GGML_TYPE_Q5_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F16   ].pipeline; break;
                             case GGML_TYPE_Q5_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F16   ].pipeline; break;
                             case GGML_TYPE_Q8_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F16   ].pipeline; break;
+                            case GGML_TYPE_MXFP4:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MXFP4_F16  ].pipeline; break;
                             case GGML_TYPE_Q2_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F16   ].pipeline; break;
                             case GGML_TYPE_Q3_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F16   ].pipeline; break;
                             case GGML_TYPE_Q4_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F16   ].pipeline; break;
@@ -4018,6 +4104,13 @@ static int ggml_metal_encode_node(
                                 nr0 = N_R0_Q8_0;
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32].pipeline;
                             } break;
+                        case GGML_TYPE_MXFP4:
+                            {
+                                nsg = N_SG_MXFP4;
+                                nr0 = N_R0_MXFP4;
+                                smem = 32*sizeof(float);
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_MXFP4_F32].pipeline;
+                            } break;
                         case GGML_TYPE_Q2_K:
                             {
                                 nsg = N_SG_Q2_K;
@@ -4170,6 +4263,7 @@ static int ggml_metal_encode_node(
                     case GGML_TYPE_Q5_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0   ].pipeline; break;
                     case GGML_TYPE_Q5_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1   ].pipeline; break;
                     case GGML_TYPE_Q8_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0   ].pipeline; break;
+                    case GGML_TYPE_MXFP4:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_MXFP4  ].pipeline; break;
                     case GGML_TYPE_Q2_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K   ].pipeline; break;
                     case GGML_TYPE_Q3_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K   ].pipeline; break;
                     case GGML_TYPE_Q4_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K   ].pipeline; break;
@@ -4980,11 +5074,14 @@ static int ggml_metal_encode_node(
                 GGML_ASSERT(ne11 == ne21);
                 GGML_ASSERT(ne12 == ne22);
 
-                struct ggml_tensor * src3 = node->src[3];
+                struct ggml_tensor * src3 = node->src[3]; // mask
+                struct ggml_tensor * src4 = node->src[4]; // sinks
 
                 size_t offs_src3 = 0;
+                size_t offs_src4 = 0;
 
                 id<MTLBuffer> id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil;
+                id<MTLBuffer> id_src4 = src4 ? ggml_metal_get_buffer(src4, &offs_src4) : nil;
 
                 GGML_ASSERT(!src3 || src3->type == GGML_TYPE_F16);
                 GGML_ASSERT(!src3 || src3->ne[1] >= GGML_PAD(src0->ne[1], 8) &&
@@ -5000,8 +5097,6 @@ static int ggml_metal_encode_node(
                 const uint64_t nb32 = src3 ? src3->nb[2] : 0; GGML_UNUSED(nb32);
                 const uint64_t nb33 = src3 ? src3->nb[3] : 0; GGML_UNUSED(nb33);
 
-                const enum ggml_type src2t = src2 ? src2->type : GGML_TYPE_COUNT; GGML_UNUSED(src2t);
-
                 float scale;
                 float max_bias;
                 float logit_softcap;
@@ -5389,7 +5484,12 @@ static int ggml_metal_encode_node(
                 } else {
                     [encoder setBuffer:id_src0 offset:offs_src0 atIndex:4];
                 }
-                [encoder setBuffer:id_dst offset:offs_dst       atIndex:5];
+                if (id_src4) {
+                    [encoder setBuffer:id_src4 offset:offs_src4 atIndex:5];
+                } else {
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:5];
+                }
+                [encoder setBuffer:id_dst offset:offs_dst       atIndex:6];
 
                 if (!use_vec_kernel) {
                     // half8x8 kernel
diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
index 99a453090f6..b35a3bbdc31 100644
--- a/ggml/src/ggml-metal/ggml-metal.metal
+++ b/ggml/src/ggml-metal/ggml-metal.metal
@@ -35,6 +35,10 @@ constexpr constant static float kvalues_iq4nl_f[16] = {
     -127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f, 1.f, 13.f, 25.f, 38.f, 53.f, 69.f, 89.f, 113.f
 };
 
+constexpr constant static float kvalues_mxfp4_f[16] = {
+    0, .5f, 1.f, 1.5f, 2.f, 3.f, 4.f, 6.f, -0, -.5f, -1.f, -1.5f, -2.f, -3.f, -4.f, -6.f
+};
+
 static inline int best_index_int8(int n, constant float * val, float x) {
     if (x <= val[0]) return 0;
     if (x >= val[n-1]) return n-1;
@@ -46,6 +50,18 @@ static inline int best_index_int8(int n, constant float * val, float x) {
     return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
 }
 
+static inline float e8m0_to_fp32(uint8_t x) {
+    uint32_t bits;
+
+    if (x == 0) {
+        bits = 0x00400000;
+    } else {
+        bits = (uint32_t) x << 23;
+    }
+
+    return as_type<float>(bits);
+}
+
 // NOTE: this is not dequantizing - we are simply fitting the template
 template <typename type4x4>
 void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) {
@@ -242,6 +258,27 @@ void quantize_q5_1(device const float * src, device block_q5_1 & dst) {
     }
 }
 
+void quantize_q8_0(device const float * src, device block_q8_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+
+    for (int j = 0; j < QK8_0; j++) {
+        const float v = src[j];
+        amax = MAX(amax, fabs(v));
+    }
+
+    const float d = amax / ((1 << 7) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    for (int j = 0; j < QK8_0; ++j) {
+        const float x0 = src[j]*id;
+
+        dst.qs[j] = round(x0);
+    }
+}
+
 void quantize_iq4_nl(device const float * src, device block_iq4_nl & dst) {
 #pragma METAL fp math_mode(safe)
     float amax = 0.0f; // absolute max
@@ -462,25 +499,34 @@ void dequantize_q8_0_t4(device const block_q8_0 *xb, short il, thread type4 & re
     }
 }
 
-void quantize_q8_0(device const float * src, device block_q8_0 & dst) {
-#pragma METAL fp math_mode(safe)
-    float amax = 0.0f; // absolute max
+template <typename type4x4>
+void dequantize_mxfp4(device const block_mxfp4 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * q2 = (device const uint8_t *)xb->qs;
 
-    for (int j = 0; j < QK8_0; j++) {
-        const float v = src[j];
-        amax = MAX(amax, fabs(v));
+    const float d = e8m0_to_fp32(xb->e);
+    const uint8_t shr = il >= 1 ? 4 : 0;
+
+    for (int i = 0; i < 4; ++i) {
+        reg[i][0] = d * kvalues_mxfp4_f[(q2[4*i + 0] >> shr) & 0x0F];
+        reg[i][1] = d * kvalues_mxfp4_f[(q2[4*i + 1] >> shr) & 0x0F];
+        reg[i][2] = d * kvalues_mxfp4_f[(q2[4*i + 2] >> shr) & 0x0F];
+        reg[i][3] = d * kvalues_mxfp4_f[(q2[4*i + 3] >> shr) & 0x0F];
     }
+}
 
-    const float d = amax / ((1 << 7) - 1);
-    const float id = d ? 1.0f/d : 0.0f;
+template <typename type4>
+void dequantize_mxfp4_t4(device const block_mxfp4 * xb, short il, thread type4 & reg) {
+    device const uint8_t * q2 = (device const uint8_t *)xb->qs;
 
-    dst.d = d;
+    const float d = e8m0_to_fp32(xb->e);
+    const short il4 = il%4;
 
-    for (int j = 0; j < QK8_0; ++j) {
-        const float x0 = src[j]*id;
+    const uint8_t shr = il >= 4 ? 4 : 0;
 
-        dst.qs[j] = round(x0);
-    }
+    reg[0] = d * kvalues_mxfp4_f[(q2[4*il4 + 0] >> shr) & 0x0F];
+    reg[1] = d * kvalues_mxfp4_f[(q2[4*il4 + 1] >> shr) & 0x0F];
+    reg[2] = d * kvalues_mxfp4_f[(q2[4*il4 + 2] >> shr) & 0x0F];
+    reg[3] = d * kvalues_mxfp4_f[(q2[4*il4 + 3] >> shr) & 0x0F];
 }
 
 template <typename type4x4>
@@ -960,6 +1006,32 @@ kernel void kernel_div(
     }
 }
 
+kernel void kernel_add_id(
+        constant ggml_metal_kargs_add_id & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i1 = tgpig.x;
+    const int i2 = tgpig.y;
+
+    const int i11 = *((device const int32_t *) (src2 + i1*sizeof(int32_t) + i2*args.nb21));
+
+    const size_t nb1 = args.ne0 * sizeof(float);
+    const size_t nb2 = args.ne1 * nb1;
+
+    device       float * dst_row  = (device       float *)((device char *)dst + i1*nb1 + i2*nb2);
+    device const float * src0_row = (device const float *)((device char *)src0 +  i1*args.nb01 + i2*args.nb02);
+    device const float * src1_row = (device const float *)((device char *)src1 + i11*args.nb11);
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        dst_row[i0] = src0_row[i0] + src1_row[i0];
+    }
+}
+
 template<typename T>
 kernel void kernel_repeat(
         constant ggml_metal_kargs_repeat & args,
@@ -1431,6 +1503,32 @@ kernel void kernel_swiglu(
     }
 }
 
+kernel void kernel_swiglu_oai(
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        constant ggml_metal_kargs_glu & args,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const float * src0_row = (device const float *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const float * src1_row = (device const float *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       float * dst_row  = (device       float *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        float x0 = src0_row[i0];
+        float x1 = src1_row[i0];
+
+        x0 = min(x0, args.limit);
+        x1 = max(min(x1, args.limit), -args.limit);
+
+        float out_glu = x0 / (1.0f + exp(-x0 * args.alpha));
+        out_glu = out_glu * (1.0f + x1);
+
+        dst_row[i0] = out_glu;
+    }
+}
+
 kernel void kernel_geglu_erf(
         device const char * src0,
         device const char * src1,
@@ -1534,6 +1632,7 @@ template<typename T>
 kernel void kernel_soft_max(
         device const  char * src0,
         device const  char * src1,
+        device const  char * src2,
         device        char * dst,
         constant ggml_metal_kargs_soft_max & args,
         threadgroup  float * buf [[threadgroup(0)]],
@@ -1552,6 +1651,7 @@ kernel void kernel_soft_max(
 
     device const float * psrc0 =                (device const float *) (src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
     device const     T * pmask = src1 != src0 ? (device const T *    ) (src1 + i11*args.nb11 + i12*args.nb12 + i13*args.nb13) : nullptr;
+    device const float * psrc2 = src2 != src0 ? (device const float *) (src2)                                                 : nullptr;
     device       float * pdst  =                (device       float *) (dst  + i01*args.nb1  + i02*args.nb2  + i03*args.nb3);
 
     float slope = 1.0f;
@@ -1567,7 +1667,7 @@ kernel void kernel_soft_max(
     }
 
     // parallel max
-    float lmax = -INFINITY;
+    float lmax = psrc2 ? psrc2[i02] : -INFINITY;
 
     for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
         lmax = MAX(lmax, psrc0[i00]*args.scale + (pmask ? slope*pmask[i00] : 0.0f));
@@ -1623,6 +1723,10 @@ kernel void kernel_soft_max(
         sum = simd_sum(sum);
     }
 
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max_val);
+    }
+
     const float inv_sum = 1.0f/sum;
 
     for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
@@ -1634,6 +1738,7 @@ template<typename T>
 kernel void kernel_soft_max_4(
         device const  char * src0,
         device const  char * src1,
+        device const  char * src2,
         device        char * dst,
         constant ggml_metal_kargs_soft_max & args,
         threadgroup  float * buf [[threadgroup(0)]],
@@ -1652,6 +1757,7 @@ kernel void kernel_soft_max_4(
 
     device const float4 * psrc4 =                (device const float4 *) (src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
     device const      T * pmask = src1 != src0 ? (device const T *     ) (src1 + i11*args.nb11 + i12*args.nb12 + i13*args.nb13) : nullptr;
+    device const float *  psrc2 = src2 != src0 ? (device const float * ) (src2)                                                 : nullptr;
     device       float4 * pdst4 =                (device       float4 *) (dst  + i01*args.nb1  + i02*args.nb2  + i03*args.nb3);
 
     float slope = 1.0f;
@@ -1666,7 +1772,7 @@ kernel void kernel_soft_max_4(
     }
 
     // parallel max
-    float4 lmax4 = -INFINITY;
+    float4 lmax4 = psrc2 ? psrc2[i02] : -INFINITY;
 
     for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
         lmax4 = fmax(lmax4, psrc4[i00]*args.scale + (float4)((pmask ? slope*pmask[i00] : 0.0f)));
@@ -1725,6 +1831,10 @@ kernel void kernel_soft_max_4(
         sum = simd_sum(sum);
     }
 
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max_val);
+    }
+
     const float inv_sum = 1.0f/sum;
 
     for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
@@ -3106,6 +3216,11 @@ template [[host_name("kernel_mul_mv_ext_q8_0_f32_r1_3")]]   kernel mul_mv_ext_q4
 template [[host_name("kernel_mul_mv_ext_q8_0_f32_r1_4")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, block_q8_0,   32, dequantize_q8_0_t4>;
 template [[host_name("kernel_mul_mv_ext_q8_0_f32_r1_5")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<5, block_q8_0,   32, dequantize_q8_0_t4>;
 
+template [[host_name("kernel_mul_mv_ext_mxfp4_f32_r1_2")]]  kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, block_mxfp4,  32, dequantize_mxfp4_t4>;
+template [[host_name("kernel_mul_mv_ext_mxfp4_f32_r1_3")]]  kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, block_mxfp4,  32, dequantize_mxfp4_t4>;
+template [[host_name("kernel_mul_mv_ext_mxfp4_f32_r1_4")]]  kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, block_mxfp4,  32, dequantize_mxfp4_t4>;
+template [[host_name("kernel_mul_mv_ext_mxfp4_f32_r1_5")]]  kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<5, block_mxfp4,  32, dequantize_mxfp4_t4>;
+
 template [[host_name("kernel_mul_mv_ext_iq4_nl_f32_r1_2")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, block_iq4_nl, 32, dequantize_iq4_nl_t4>;
 template [[host_name("kernel_mul_mv_ext_iq4_nl_f32_r1_3")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, block_iq4_nl, 32, dequantize_iq4_nl_t4>;
 template [[host_name("kernel_mul_mv_ext_iq4_nl_f32_r1_4")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, block_iq4_nl, 32, dequantize_iq4_nl_t4>;
@@ -4092,6 +4207,7 @@ kernel void kernel_flash_attn_ext(
         device const char * k,
         device const char * v,
         device const char * mask,
+        device const char * sinks,
         device       char * dst,
         threadgroup  half * shmem_f16 [[threadgroup(0)]],
         uint3   tgpig[[threadgroup_position_in_grid]],
@@ -4407,6 +4523,35 @@ kernel void kernel_flash_attn_ext(
             }
         }
 
+        if (sinks != q && sgitg == 0) {
+            for (ushort j = 0; j < Q; ++j) {
+                const float m = M[j];
+                const float s = tiisg == 0 ? ((device const float *) sinks)[iq2] : -FLT_MAX/2;
+
+                M[j] = simd_max(max(M[j], s));
+
+                const float ms = exp(m - M[j]);
+                const float vs = exp(s - M[j]);
+
+                S[j] = S[j]*ms + simd_sum(vs);
+
+                if (tiisg == j) {
+                    ss[j*TS + 2*C + j] = ms;
+                }
+            }
+
+            // O = diag(ms)*O
+            {
+                s8x8_t ms;
+                simdgroup_load(ms, ss + 2*C, TS, 0, false);
+
+                #pragma unroll(DV8)
+                for (short i = 0; i < DV8; ++i) {
+                    simdgroup_multiply(lo[i], ms, lo[i]);
+                }
+            }
+        }
+
         // these are needed for reducing the results from the simdgroups (reuse the ss buffer)
         for (short j = tiisg; j < Q; j += NW) {
             ss[j*TS + 0] = S[j];
@@ -4618,6 +4763,7 @@ kernel void kernel_flash_attn_ext_vec(
         device const char * k,
         device const char * v,
         device const char * mask,
+        device const char * sinks,
         device       char * dst,
         threadgroup  half * shmem_f16 [[threadgroup(0)]],
         uint3   tgpig[[threadgroup_position_in_grid]],
@@ -4835,6 +4981,23 @@ kernel void kernel_flash_attn_ext_vec(
             }
         }
 
+        if (sinks != q && sgitg == 0) {
+            const float m = M;
+            const float s = tiisg == 0 ? ((device const float *) sinks)[iq2] : -FLT_MAX/2;
+
+            M = simd_max(max(M, s));
+
+            const float ms = exp(m - M);
+            const float vs = exp(s - M);
+
+            S = S*ms + simd_sum(vs);
+
+#pragma unroll(DV4/NL)
+            for (short ii = 0; ii < DV4; ii += NL) {
+                lo[ii/NL] *= ms;
+            }
+        }
+
         // these are needed for reducing the results from the simdgroups (reuse the ss buffer)
         if (tiisg == 0) {
             ss[0] = (s_t) S;
@@ -6940,6 +7103,95 @@ kernel void kernel_mul_mv_iq4_xs_f32(
     kernel_mul_mv_iq4_xs_f32_impl<N_R0_IQ4_XS, N_SG_IQ4_XS, N_SIMDWIDTH, constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
 }
 
+template<int nr0, int nsg, int nw, typename args_t>
+void kernel_mul_mv_mxfp4_f32_impl(
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    threadgroup float * shmem_f32 = (threadgroup float *) shmem;
+    const int nb = args.ne00/QK_MXFP4;
+
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+
+    const int first_row = (r0 * nsg + sgitg) * nr0;
+
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
+
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
+
+    device const block_mxfp4 * x = (device const block_mxfp4 *) (src0 + offset0);
+    device const float       * y = (device const float       *) (src1 + offset1);
+
+    const short ix = tiisg/2;  // 0...15
+    const short it = tiisg%2;  // 0 or 1
+
+    shmem_f32[tiisg] = kvalues_mxfp4_f[tiisg%16];
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    float4 yl[4];
+    float sumf[nr0]={0.f};
+
+    device const float * yb = y + ix * QK_MXFP4 + it * 8;
+
+    for (int ib = ix; ib < nb; ib += 16) {
+        device const float4 * y4 = (device const float4 *)yb;
+        yl[0] = y4[0];
+        yl[1] = y4[4];
+        yl[2] = y4[1];
+        yl[3] = y4[5];
+
+#pragma unroll(nr0)
+        for (short row = 0; row < nr0; row++) {
+            device const block_mxfp4 & xb = x[row*nb + ib];
+            device const uint8_t     * q2 = (device const uint8_t *)(xb.qs + 8*it);
+
+            float4 acc1 = yl[0]*float4(shmem_f32[q2[0] &  0x0F], shmem_f32[q2[1] &  0x0F], shmem_f32[q2[2] &  0x0F], shmem_f32[q2[3] &  0x0F]);
+            float4 acc2 = yl[1]*float4(shmem_f32[q2[0] >> 4   ], shmem_f32[q2[1] >> 4   ], shmem_f32[q2[2] >> 4   ], shmem_f32[q2[3] >> 4   ]);
+            float4 acc3 = yl[2]*float4(shmem_f32[q2[4] &  0x0F], shmem_f32[q2[5] &  0x0F], shmem_f32[q2[6] &  0x0F], shmem_f32[q2[7] &  0x0F]);
+            float4 acc4 = yl[3]*float4(shmem_f32[q2[4] >> 4   ], shmem_f32[q2[5] >> 4   ], shmem_f32[q2[6] >> 4   ], shmem_f32[q2[7] >> 4   ]);
+
+            acc1 = (acc1 + acc3) + (acc2 + acc4);
+
+            sumf[row] += e8m0_to_fp32(xb.e) * ((acc1[0] + acc1[1]) + (acc1[2] + acc1[3]));
+        }
+
+        yb += 16 * QK_MXFP4;
+    }
+
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
+    for (int row = 0; row < nr0 && first_row + row < args.ne0; ++row) {
+        float sum_all = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst_f32[first_row + row] = sum_all;
+        }
+    }
+}
+
+[[host_name("kernel_mul_mv_mxfp4_f32")]]
+kernel void kernel_mul_mv_mxfp4_f32(
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_mxfp4_f32_impl<N_R0_MXFP4, N_SG_MXFP4, N_SIMDWIDTH, constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
+}
+
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
 kernel void kernel_get_rows_q(
         constant ggml_metal_kargs_get_rows & args,
@@ -7475,6 +7727,7 @@ template [[host_name("kernel_get_rows_q4_1")]]    kernel get_rows_q_t kernel_get
 template [[host_name("kernel_get_rows_q5_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_0,    2, dequantize_q5_0>;
 template [[host_name("kernel_get_rows_q5_1")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_1,    2, dequantize_q5_1>;
 template [[host_name("kernel_get_rows_q8_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q8_0,    2, dequantize_q8_0>;
+template [[host_name("kernel_get_rows_mxfp4")]]   kernel get_rows_q_t kernel_get_rows_q<block_mxfp4,   2, dequantize_mxfp4>;
 template [[host_name("kernel_get_rows_q2_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q2_K,    QK_NL, dequantize_q2_K>;
 template [[host_name("kernel_get_rows_q3_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q3_K,    QK_NL, dequantize_q3_K>;
 template [[host_name("kernel_get_rows_q4_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_K,    QK_NL, dequantize_q4_K>;
@@ -7527,6 +7780,7 @@ template [[host_name("kernel_mul_mm_q4_1_f32")]]    kernel mul_mm_t kernel_mul_m
 template [[host_name("kernel_mul_mm_q5_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0>;
 template [[host_name("kernel_mul_mm_q5_1_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1>;
 template [[host_name("kernel_mul_mm_q8_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0>;
+template [[host_name("kernel_mul_mm_mxfp4_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4>;
 template [[host_name("kernel_mul_mm_q2_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K>;
 template [[host_name("kernel_mul_mm_q3_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K>;
 template [[host_name("kernel_mul_mm_q4_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K>;
@@ -7558,6 +7812,7 @@ template [[host_name("kernel_mul_mm_id_q4_1_f16")]]    kernel mul_mm_id kernel_m
 template [[host_name("kernel_mul_mm_id_q5_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0>;
 template [[host_name("kernel_mul_mm_id_q5_1_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1>;
 template [[host_name("kernel_mul_mm_id_q8_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0>;
+template [[host_name("kernel_mul_mm_id_mxfp4_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4>;
 template [[host_name("kernel_mul_mm_id_q2_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K>;
 template [[host_name("kernel_mul_mm_id_q3_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K>;
 template [[host_name("kernel_mul_mm_id_q4_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K>;
@@ -7703,6 +7958,8 @@ template [[host_name("kernel_mul_mv_id_q4_1_f32")]]    kernel kernel_mul_mv_id_t
 template [[host_name("kernel_mul_mv_id_q5_0_f32")]]    kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<mul_vec_q_n_f32_impl<block_q5_0, N_R0_Q5_0, N_SG_Q5_0, N_SIMDWIDTH>>>;
 template [[host_name("kernel_mul_mv_id_q5_1_f32")]]    kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<mul_vec_q_n_f32_impl<block_q5_1, N_R0_Q5_1, N_SG_Q5_1, N_SIMDWIDTH>>>;
 
+template [[host_name("kernel_mul_mv_id_mxfp4_f32")]]   kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_mxfp4_f32_impl<N_R0_MXFP4, N_SG_MXFP4, N_SIMDWIDTH>>>;
+
 template [[host_name("kernel_mul_mv_id_q2_K_f32")]]    kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q2_K_f32_impl   <N_R0_Q2_K,    N_SG_Q2_K,    N_SIMDWIDTH>>>;
 template [[host_name("kernel_mul_mv_id_q3_K_f32")]]    kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q3_K_f32_impl   <N_R0_Q3_K,    N_SG_Q3_K,    N_SIMDWIDTH>>>;
 template [[host_name("kernel_mul_mv_id_q4_K_f32")]]    kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q4_K_f32_impl   <N_R0_Q4_K,    N_SG_Q4_K,    N_SIMDWIDTH>>>;
diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
index c9316eb7fd3..bb8b310b983 100644
--- a/ggml/src/ggml-opencl/ggml-opencl.cpp
+++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
@@ -2497,6 +2497,8 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
         case GGML_OP_CLAMP:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_SOFT_MAX:
+            // TODO: support attention sinks [TAG_ATTN_SINKS]
+            return op->src[2] == nullptr;
         case GGML_OP_NORM:
         case GGML_OP_RMS_NORM:
             return true;
diff --git a/ggml/src/ggml-quants.c b/ggml/src/ggml-quants.c
index 9a7d1b22d79..a57d2a16d6c 100644
--- a/ggml/src/ggml-quants.c
+++ b/ggml/src/ggml-quants.c
@@ -21,6 +21,17 @@
 
 #define UNUSED GGML_UNUSED
 
+static inline int best_index_int8(int n, const int8_t * val, float x) {
+    if (x <= val[0]) return 0;
+    if (x >= val[n-1]) return n-1;
+    int ml = 0, mu = n-1;
+    while (mu-ml > 1) {
+        int mav = (ml+mu)/2;
+        if (x < val[mav]) mu = mav; else ml = mav;
+    }
+    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
+}
+
 // reference implementation for deterministic creation of model files
 void quantize_row_q4_0_ref(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k) {
     static const int qk = QK4_0;
@@ -246,6 +257,53 @@ void quantize_row_q8_1_ref(const float * GGML_RESTRICT x, block_q8_1 * GGML_REST
     }
 }
 
+static inline int best_index_mxfp4(float x, float e) {
+    int best_index = 0;
+    float best_err = fabsf(kvalues_mxfp4[0]*e - x);
+    for (int i = 1; i < 16; i++) {
+        float err = fabsf(kvalues_mxfp4[i]*e - x);
+        if (err < best_err) {
+            best_index = i;
+            best_err = err;
+        }
+    }
+    return best_index;
+}
+
+void quantize_row_mxfp4_ref(const float * GGML_RESTRICT x, block_mxfp4 * GGML_RESTRICT y, int64_t k) {
+    static const int qk = QK_MXFP4;
+
+    assert(k % qk == 0);
+
+    const int nb = k / qk;
+
+    for (int i = 0; i < nb; i++) {
+        float amax = 0.0f; // absolute max
+
+        for (int j = 0; j < qk; j++) {
+            const float v = x[i*qk + j];
+
+            if (amax < fabsf(v)) {
+                amax = fabsf(v);
+            }
+        }
+
+        const uint8_t e = (uint8_t) (floorf(log2f(amax)) - 2 + 127);
+
+        const float d = GGML_E8M0_TO_FP32_HALF(e);
+
+        y[i].e = e;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t x0 = best_index_mxfp4(x[i*qk + 0    + j], d);
+            const uint8_t x1 = best_index_mxfp4(x[i*qk + qk/2 + j], d);
+
+            y[i].qs[j]  = x0;
+            y[i].qs[j] |= x1 << 4;
+        }
+    }
+}
+
 void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     static const int qk = QK4_0;
 
@@ -356,6 +414,26 @@ void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRI
     }
 }
 
+void dequantize_row_mxfp4(const block_mxfp4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+    static const int qk = QK_MXFP4;
+
+    assert(k % qk == 0);
+
+    const int nb = k / qk;
+
+    for (int i = 0; i < nb; i++) {
+        const float d = GGML_E8M0_TO_FP32_HALF(x[i].e);
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int8_t x0 = kvalues_mxfp4[x[i].qs[j] & 0x0F];
+            const int8_t x1 = kvalues_mxfp4[x[i].qs[j] >>   4];
+
+            y[i*qk + j + 0   ] = x0*d;
+            y[i*qk + j + qk/2] = x1*d;
+        }
+    }
+}
+
 //
 // 2-6 bit quantization in super-blocks
 //
@@ -2014,6 +2092,12 @@ size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst,
     return nrow * row_size;
 }
 
+size_t quantize_mxfp4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    GGML_UNUSED(quant_weights);
+    quantize_row_mxfp4_ref(src, dst, (int64_t)nrow*n_per_row);
+    return nrow * ggml_row_size(GGML_TYPE_MXFP4, n_per_row);
+}
+
 // ====================== Ternary (de)-quantization (BitNet b1.58 and TriLMs)
 
 void quantize_row_tq1_0_ref(const float * GGML_RESTRICT x, block_tq1_0 * GGML_RESTRICT y, int64_t k) {
@@ -4551,17 +4635,6 @@ size_t quantize_iq1_m(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst,
 
 // ============================ 4-bit non-linear quants
 
-static inline int best_index_int8(int n, const int8_t * val, float x) {
-    if (x <= val[0]) return 0;
-    if (x >= val[n-1]) return n-1;
-    int ml = 0, mu = n-1;
-    while (mu-ml > 1) {
-        int mav = (ml+mu)/2;
-        if (x < val[mav]) mu = mav; else ml = mav;
-    }
-    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
-}
-
 static void quantize_row_iq4_nl_impl(const int super_block_size, const int block_size, const float * GGML_RESTRICT x,
         ggml_fp16_t * dh, uint8_t * q4, uint16_t * scales_h, uint8_t * scales_l,
         float * scales, float * weight, uint8_t * L,
@@ -4961,6 +5034,15 @@ static bool validate_fp16(ggml_fp16_t f, size_t i) {
     return true;
 }
 
+static bool validate_e_e8m0(uint8_t e, size_t i) {
+    if (e == 0xff) {
+        fprintf(stderr, "ggml_validate_row_data: found invalid e value %d at block %zu\n", e, i);
+        return false;
+    }
+
+    return true;
+}
+
 #define VALIDATE_ROW_DATA_D_F16_IMPL(type, data, nb) \
     const type * q = (const type *) (data); \
     for (size_t i = 0; i < (nb); ++i) { \
@@ -4977,6 +5059,14 @@ static bool validate_fp16(ggml_fp16_t f, size_t i) {
         } \
     }
 
+#define VALIDATE_ROW_DATA_E_E8M0_IMPL(type, data, nb) \
+    const type * q = (const type *) (data); \
+    for (size_t i = 0; i < (nb); ++i) { \
+        if (!validate_e_e8m0(q[i].e, i)) { \
+            return false; \
+        } \
+    }
+
 #define VALIDATE_ROW_DATA_DVEC_F16_IMPL(type, data, nb, nr) \
     const type * q = (const type *) (data); \
     for (size_t i = 0; i < (nb); ++i) { \
@@ -5130,6 +5220,10 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte
             {
                 VALIDATE_ROW_DATA_D_F16_IMPL(block_q8_0, data, nb);
             } break;
+        case GGML_TYPE_MXFP4:
+            {
+                VALIDATE_ROW_DATA_E_E8M0_IMPL(block_mxfp4, data, nb);
+            } break;
         case GGML_TYPE_Q2_K:
             {
                 VALIDATE_ROW_DATA_DM_F16_IMPL(block_q2_K, data, nb, d, dmin);
diff --git a/ggml/src/ggml-quants.h b/ggml/src/ggml-quants.h
index d09173e1116..3b688f31c21 100644
--- a/ggml/src/ggml-quants.h
+++ b/ggml/src/ggml-quants.h
@@ -21,6 +21,8 @@ GGML_API void quantize_row_q5_1_ref(const float * GGML_RESTRICT x, block_q5_1 *
 GGML_API void quantize_row_q8_0_ref(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k);
 GGML_API void quantize_row_q8_1_ref(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k);
 
+GGML_API void quantize_row_mxfp4_ref(const float * GGML_RESTRICT x, block_mxfp4 * GGML_RESTRICT y, int64_t k);
+
 GGML_API void quantize_row_q2_K_ref(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k);
 GGML_API void quantize_row_q3_K_ref(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k);
 GGML_API void quantize_row_q4_K_ref(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t k);
@@ -45,6 +47,8 @@ GGML_API void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GG
 GGML_API void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 //GGML_API void dequantize_row_q8_1(const block_q8_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 
+GGML_API void dequantize_row_mxfp4(const block_mxfp4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+
 GGML_API void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 GGML_API void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 GGML_API void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
@@ -90,6 +94,8 @@ GGML_API size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTR
 GGML_API size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 GGML_API size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 
+GGML_API size_t quantize_mxfp4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
 GGML_API void iq2xs_init_impl(enum ggml_type type);
 GGML_API void iq2xs_free_impl(enum ggml_type type);
 GGML_API void iq3xs_init_impl(int grid_size);
diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
index f8cdb03c3d4..6fa27418ce8 100644
--- a/ggml/src/ggml-sycl/ggml-sycl.cpp
+++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
@@ -4193,15 +4193,9 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
             {
-                struct ggml_tensor * a;
-                struct ggml_tensor * b;
-                if (op->op == GGML_OP_MUL_MAT) {
-                    a = op->src[0];
-                    b = op->src[1];
-                } else {
-                    a = op->src[2];
-                    b = op->src[1];
-                }
+                struct ggml_tensor * a = op->src[0];
+                struct ggml_tensor * b = op->src[1];
+
                 if (a->ne[3] != b->ne[3]) {
                     return false;
                 }
@@ -4216,7 +4210,9 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                     }
                 }
                 ggml_type src0_type = op->src[0]->type;
-                if (src0_type == GGML_TYPE_BF16) {
+                if (src0_type == GGML_TYPE_BF16 || src0_type == GGML_TYPE_MXFP4) {
+                    // TODO: support MXFP4
+                    // FIXME: keep a list of supported types to avoid breaking the backend when a new type is added
                     return false;
                 }
                 return true;
@@ -4361,6 +4357,10 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
             if (op->src[0]->ne[3] != 1) {
                 return false;
             }
+            // TODO: support attention sinks [TAG_ATTN_SINKS]
+            if (op->src[2]) {
+                return false;
+            }
             // TODO: support broadcast
             // ref: https://github.com/ggml-org/llama.cpp/pull/14435
             return !op->src[1] || (op->src[1]->ne[2] == 1 && op->src[1]->ne[3] == 1);
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 3c1ae084996..165933a7294 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -449,6 +449,8 @@ struct vk_device_struct {
     vk_pipeline pipeline_div[2][2][2];
     vk_pipeline pipeline_div_norepeat[2][2][2];
 
+    vk_pipeline pipeline_add_id_f32;
+
     vk_pipeline pipeline_concat_f32, pipeline_concat_f16, pipeline_concat_i32;
     vk_pipeline pipeline_upscale_nearest_f32, pipeline_upscale_bilinear_f32, pipeline_upscale_bilinear_ac_f32;
     vk_pipeline pipeline_scale_f32;
@@ -483,6 +485,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_geglu[2];
     vk_pipeline pipeline_reglu[2];
     vk_pipeline pipeline_swiglu[2];
+    vk_pipeline pipeline_swiglu_oai[2];
     vk_pipeline pipeline_geglu_erf[2];
     vk_pipeline pipeline_geglu_quick[2];
 
@@ -705,6 +708,8 @@ struct vk_op_glu_push_constants {
     uint32_t ne00;
     uint32_t ne20;
     uint32_t mode;  // 0: default, 1: swapped, 2: split
+    float alpha; // for swiglu_oai
+    float limit;
 };
 
 struct vk_op_unary_push_constants {
@@ -794,6 +799,15 @@ struct vk_op_binary_push_constants {
     float param1; float param2; int32_t param3;
 };
 
+struct vk_op_add_id_push_constants {
+    uint32_t ne0;
+    uint32_t ne1;
+    uint32_t s01;
+    uint32_t s02;
+    uint32_t s11;
+    uint32_t s21;
+};
+
 struct vk_op_diag_mask_push_constants {
     uint32_t ncols;
     uint32_t rows_per_channel;
@@ -835,6 +849,7 @@ struct vk_op_soft_max_push_constants {
     float m1;
     uint32_t n_head_log2;
     uint32_t nrows_x;
+    uint32_t has_sinks;
 };
 
 struct vk_op_argsort_push_constants {
@@ -1977,6 +1992,7 @@ static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vec
         break;
     case GGML_TYPE_IQ4_NL:
     case GGML_TYPE_IQ4_XS:
+    case GGML_TYPE_MXFP4:
         lut_size = 4*16;
         break;
     default:
@@ -2353,6 +2369,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ3_S],   matmul_iq3_s_f16,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
         CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ4_XS],  matmul_iq4_xs_f16,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
         CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ4_NL],  matmul_iq4_nl_f16,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_MXFP4],   matmul_mxfp4_f16,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
 
         CREATE_MM2(pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_id_push_constants, 4)
 #if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
@@ -2379,6 +2396,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc,   matmul_id_iq3_s_f16,   , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
         CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc,  matmul_id_iq4_xs_f16,  , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
         CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc,  matmul_id_iq4_nl_f16,  , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f16acc,   matmul_id_mxfp4_f16,   , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
 #undef CREATE_MM
 #undef CREATE_MM2
     } else
@@ -2440,6 +2458,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
             CREATE_MM2(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_S],   matmul_iq3_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
             CREATE_MM2(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_XS],  matmul_iq4_xs_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
             CREATE_MM2(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL],  matmul_iq4_nl_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM2(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat[GGML_TYPE_MXFP4],   matmul_mxfp4_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         } else {
             CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
             CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
@@ -2461,6 +2480,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
             CREATE_MM(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_S].f32acc,   matmul_iq3_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
             CREATE_MM(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_XS].f32acc,  matmul_iq4_xs_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
             CREATE_MM(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc,  matmul_iq4_nl_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat[GGML_TYPE_MXFP4].f32acc,   matmul_mxfp4_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         }
 
         CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
@@ -2493,6 +2513,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
             CREATE_MM(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc,   matmul_id_iq3_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
             CREATE_MM(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc,  matmul_id_iq4_xs_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
             CREATE_MM(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc,  matmul_id_iq4_nl_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f16acc,   matmul_id_mxfp4_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         } else {
             CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
             CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
@@ -2514,6 +2535,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
             CREATE_MM(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc,   matmul_id_iq3_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
             CREATE_MM(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc,  matmul_id_iq4_xs_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
             CREATE_MM(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc,  matmul_id_iq4_nl_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f16acc,   matmul_id_mxfp4_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         }
 #undef CREATE_MM2
 #undef CREATE_MM
@@ -2581,6 +2603,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM2(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_S],   matmul_iq3_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         CREATE_MM2(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_XS],  matmul_iq4_xs_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         CREATE_MM2(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL],  matmul_iq4_nl_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat[GGML_TYPE_MXFP4],   matmul_mxfp4_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
 
 #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
         if (device->integer_dot_product) {
@@ -2618,6 +2641,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc,   matmul_id_iq3_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc,  matmul_id_iq4_xs_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc,  matmul_id_iq4_nl_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f16acc,   matmul_id_mxfp4_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
 #undef CREATE_MM2
 #undef CREATE_MMQ
 #undef CREATE_MM
@@ -2672,6 +2696,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_S].f32acc,   matmul_iq3_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         CREATE_MM(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_XS].f32acc,  matmul_iq4_xs_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         CREATE_MM(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc,  matmul_iq4_nl_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat[GGML_TYPE_MXFP4].f32acc,   matmul_mxfp4_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
 
 #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
         if (device->integer_dot_product) {
@@ -2709,6 +2734,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f32acc,   matmul_id_iq3_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f32acc,  matmul_id_iq4_xs_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc,  matmul_id_iq4_nl_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f32acc,   matmul_id_mxfp4_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
     }
     // reusing CREATE_MM from the fp32 path
     if ((device->coopmat2 || device->coopmat_support)
@@ -2767,6 +2793,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq3_s_f32_f32_len,   mul_mat_vec_iq3_s_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f32_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_xs_f32_f32_len,  mul_mat_vec_iq4_xs_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f32_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_nl_f32_f32_len,  mul_mat_vec_iq4_nl_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_MXFP4][i],   "mul_mat_vec_mxfp4_f32_f32_"+std::to_string(i+1),   mul_mat_vec_mxfp4_f32_f32_len,   mul_mat_vec_mxfp4_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
 
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f16_f32_"+std::to_string(i+1),  mul_mat_vec_f32_f16_f32_len,  mul_mat_vec_f32_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f16_f32_"+std::to_string(i+1),  mul_mat_vec_f16_f16_f32_len,  mul_mat_vec_f16_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
@@ -2790,6 +2817,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq3_s_f16_f32_len,   mul_mat_vec_iq3_s_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f16_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_xs_f16_f32_len,  mul_mat_vec_iq4_xs_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f16_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_nl_f16_f32_len,  mul_mat_vec_iq4_nl_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_MXFP4][i],   "mul_mat_vec_mxfp4_f16_f32_"+std::to_string(i+1),   mul_mat_vec_mxfp4_f16_f32_len,   mul_mat_vec_mxfp4_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
     }
 
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32",  mul_mat_vec_id_f32_f32_len,  mul_mat_vec_id_f32_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
@@ -2814,6 +2842,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ3_S],   "mul_mat_vec_id_iq3_s_f32",   mul_mat_vec_id_iq3_s_f32_len,   mul_mat_vec_id_iq3_s_f32_data,   "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_XS],  "mul_mat_vec_id_iq4_xs_f32",  mul_mat_vec_id_iq4_xs_f32_len,  mul_mat_vec_id_iq4_xs_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL],  "mul_mat_vec_id_iq4_nl_f32",  mul_mat_vec_id_iq4_nl_f32_len,  mul_mat_vec_id_iq4_nl_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_MXFP4],   "mul_mat_vec_id_mxfp4_f32",   mul_mat_vec_id_mxfp4_f32_len,   mul_mat_vec_id_mxfp4_f32_data,   "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
 
     // dequant shaders
     ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_F32 ], "f32_to_f16",   dequant_f32_len,  dequant_f32_data,  "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
@@ -2836,6 +2865,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ3_S],   "dequant_iq3_s",   dequant_iq3_s_len,   dequant_iq3_s_data,   "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ4_XS],  "dequant_iq4_xs",  dequant_iq4_xs_len,  dequant_iq4_xs_data,  "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ4_NL],  "dequant_iq4_nl",  dequant_iq4_nl_len,  dequant_iq4_nl_data,  "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_MXFP4],   "dequant_mxfp4",   dequant_mxfp4_len,   dequant_mxfp4_data,   "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
 
     // get_rows
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_F32 ], "get_rows_f32",  get_rows_f32_len,  get_rows_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
@@ -2855,6 +2885,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ3_S],   "get_rows_iq3_s",   get_rows_iq3_s_len,   get_rows_iq3_s_data,   "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ4_XS],  "get_rows_iq4_xs",  get_rows_iq4_xs_len,  get_rows_iq4_xs_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ4_NL],  "get_rows_iq4_nl",  get_rows_iq4_nl_len,  get_rows_iq4_nl_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_MXFP4],   "get_rows_mxfp4",   get_rows_mxfp4_len,   get_rows_mxfp4_data,   "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_F32 ], "get_rows_f32_f32",  get_rows_f32_f32_len,  get_rows_f32_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_F16 ], "get_rows_f16_f32",  get_rows_f16_f32_len,  get_rows_f16_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
@@ -2873,6 +2904,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ3_S],   "get_rows_iq3_s_f32",   get_rows_iq3_s_f32_len,   get_rows_iq3_s_f32_data,   "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ4_XS],  "get_rows_iq4_xs_f32",  get_rows_iq4_xs_f32_len,  get_rows_iq4_xs_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ4_NL],  "get_rows_iq4_nl_f32",  get_rows_iq4_nl_f32_len,  get_rows_iq4_nl_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_MXFP4],   "get_rows_mxfp4_f32",   get_rows_mxfp4_f32_len,   get_rows_mxfp4_f32_data,   "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_matmul_split_k_reduce, "split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256 * 4, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_flash_attn_split_k_reduce, "fa_split_k_reduce", fa_split_k_reduce_len, fa_split_k_reduce_data, "main", 2, 4 * sizeof(uint32_t), {1, device->subgroup_size, 1}, {device->subgroup_size}, 1, true);
@@ -2976,6 +3008,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
     CREATE_BINARY(div, _norepeat, {1})
 #undef CREATE_BINARY
 
+    ggml_vk_create_pipeline(device, device->pipeline_add_id_f32, "add_id_f32", add_id_f32_len, add_id_f32_data, "main", 4, sizeof(vk_op_add_id_push_constants), {1, 1, 1}, {}, 1);
+
     ggml_vk_create_pipeline(device, device->pipeline_acc_f32, "acc_f32", acc_f32_len, acc_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_concat_f32, "concat_f32", concat_f32_len, concat_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
@@ -3026,6 +3060,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     CREATE_GLU(geglu)
     CREATE_GLU(reglu)
     CREATE_GLU(swiglu)
+    CREATE_GLU(swiglu_oai)
     CREATE_GLU(geglu_erf)
     CREATE_GLU(geglu_quick)
 #undef CREATE_GLU
@@ -3035,10 +3070,10 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {1, 512, 1}, {}, 1, true);
 
-    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32, "soft_max_f32", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_wg512, "soft_max_f32_wg512", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16, "soft_max_f32_f16", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16_wg512, "soft_max_f32_f16_wg512", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32, "soft_max_f32", soft_max_f32_len, soft_max_f32_data, "main", 4, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_wg512, "soft_max_f32_wg512", soft_max_f32_len, soft_max_f32_data, "main", 4, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16, "soft_max_f32_f16", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 4, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16_wg512, "soft_max_f32_f16_wg512", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 4, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1);
     ggml_vk_create_pipeline(device, device->pipeline_soft_max_back_f32, "soft_max_back_f32", soft_max_back_f32_len, soft_max_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f32, "rope_norm_f32", rope_norm_f32_len, rope_norm_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
@@ -4244,6 +4279,7 @@ static vk_pipeline ggml_vk_get_to_fp16(ggml_backend_vk_context * ctx, ggml_type
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_MXFP4:
             break;
         default:
             return nullptr;
@@ -4314,6 +4350,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_MXFP4:
             break;
         default:
             return nullptr;
@@ -4357,6 +4394,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context *
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_MXFP4:
             break;
         default:
             return nullptr;
@@ -4411,6 +4449,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_MXFP4:
             break;
         default:
             return nullptr;
@@ -4446,6 +4485,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_MXFP4:
             break;
         default:
             return nullptr;
@@ -4631,6 +4671,7 @@ static void ggml_vk_dispatch_pipeline(ggml_backend_vk_context* ctx, vk_context&
     std::cerr << "}, (" << wg0 << "," << wg1 << "," << wg2 << "))");
     GGML_ASSERT(ctx->descriptor_set_idx < ctx->descriptor_sets.size());
     GGML_ASSERT(descriptor_buffer_infos.size() <= MAX_PARAMETER_COUNT);
+    GGML_ASSERT(pipeline->parameter_count == descriptor_buffer_infos.size());
 
     vk::DescriptorSet& descriptor_set = ctx->descriptor_sets[ctx->descriptor_set_idx++];
     vk::WriteDescriptorSet write_descriptor_set{ descriptor_set, 0, 0, pipeline->parameter_count, vk::DescriptorType::eStorageBuffer, nullptr, descriptor_buffer_infos.begin() };
@@ -6847,6 +6888,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             break;
         }
         return nullptr;
+    case GGML_OP_ADD_ID:
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && src2->type == GGML_TYPE_I32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_add_id_f32;
+        }
+        return nullptr;
     case GGML_OP_CONCAT:
         if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_concat_f32;
@@ -6992,6 +7038,8 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
                 return ctx->device->pipeline_reglu[dst->type == GGML_TYPE_F16];
             case GGML_GLU_OP_SWIGLU:
                 return ctx->device->pipeline_swiglu[dst->type == GGML_TYPE_F16];
+            case GGML_GLU_OP_SWIGLU_OAI:
+                return ctx->device->pipeline_swiglu_oai[dst->type == GGML_TYPE_F16];
             case GGML_GLU_OP_GEGLU_ERF:
                 return ctx->device->pipeline_geglu_erf[dst->type == GGML_TYPE_F16];
             case GGML_GLU_OP_GEGLU_QUICK:
@@ -7007,6 +7055,7 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         return nullptr;
     case GGML_OP_SOFT_MAX:
         GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);
+        GGML_ASSERT(!src2 || src2->type == GGML_TYPE_F32);
 
         if (src0->type == GGML_TYPE_F32 && (src1 == nullptr || src1->type == GGML_TYPE_F32) && dst->type == GGML_TYPE_F32) {
             return src0->ne[0] > 1024 ? ctx->device->pipeline_soft_max_f32_wg512 : ctx->device->pipeline_soft_max_f32;
@@ -7177,6 +7226,7 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
     case GGML_OP_SUB:
     case GGML_OP_MUL:
     case GGML_OP_DIV:
+    case GGML_OP_ADD_ID:
     case GGML_OP_CONCAT:
     case GGML_OP_UPSCALE:
     case GGML_OP_SQR:
@@ -7523,6 +7573,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
                 elements = { ne, 1, 1 };
             }
         } break;
+    case GGML_OP_ADD_ID:
+        {
+            elements = { (uint32_t)ne01, (uint32_t)ne02, 1 };
+        } break;
     case GGML_OP_SET_ROWS:
         {
             uint32_t ne = ggml_nelements(src0);
@@ -7562,8 +7616,8 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         }
     }
 
-    if (op == GGML_OP_SOFT_MAX || op == GGML_OP_GLU) {
-        // Empty src1 is possible in soft_max, but the shader needs a buffer
+    if (op == GGML_OP_GLU) {
+        // Empty src1 is possible in glu, but the shader needs a buffer
         vk_subbuffer subbuf_y;
         if (use_src1) {
             subbuf_y = { d_Y, y_buf_offset, y_sz };
@@ -7573,6 +7627,24 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
 
         ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, subbuf_y, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
+    } else if (op == GGML_OP_SOFT_MAX) {
+        // Empty src1 and src2 is possible in soft_max, but the shader needs a buffer
+        vk_subbuffer subbuf_y;
+        if (use_src1) {
+            subbuf_y = { d_Y, y_buf_offset, y_sz };
+        } else {
+            subbuf_y = { d_X, 0, x_sz };
+        }
+
+        vk_subbuffer subbuf_z;
+        if (use_src2) {
+            subbuf_z = { d_Z, z_buf_offset, z_sz };
+        } else {
+            subbuf_z = { d_X, 0, x_sz };
+        }
+
+        ggml_vk_sync_buffers(subctx);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, subbuf_y, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else if (op == GGML_OP_ROPE || op == GGML_OP_ROPE_BACK) {
         // Empty src2 is possible in rope, but the shader needs a buffer
         vk_subbuffer subbuf_z;
@@ -7701,6 +7773,21 @@ static void ggml_vk_div(ggml_backend_vk_context * ctx, vk_context& subctx, const
     }, dryrun);
 }
 
+static void ggml_vk_add_id(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool dryrun = false) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t src1_type_size = ggml_type_size(src1->type);
+    const uint32_t src2_type_size = ggml_type_size(src2->type);
+
+    ggml_vk_op_f32<vk_op_add_id_push_constants>(ctx, subctx, src0, src1, src2, dst, GGML_OP_ADD_ID, {
+        (uint32_t)dst->ne[0],
+        (uint32_t)dst->ne[1],
+        (uint32_t)src0->nb[1] / src0_type_size,
+        (uint32_t)src0->nb[2] / src0_type_size,
+        (uint32_t)src1->nb[1] / src1_type_size,
+        (uint32_t)src2->nb[1] / src2_type_size,
+    }, dryrun);
+}
+
 static void ggml_vk_op_f32_wkv(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, const vk_op_rwkv_wkv6_push_constants&& pc, int version, bool dryrun = false) {
     GGML_ASSERT(version == 6 || version == 7);
     int num_srcs = version == 6 ? 6 : 7;
@@ -8119,8 +8206,12 @@ static void ggml_vk_unary(ggml_backend_vk_context * ctx, vk_context& subctx, con
 }
 
 static void ggml_vk_glu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    const float * op_params_f = (const float *)dst->op_params;
+
     const bool swapped = (bool)dst->op_params[1];
     const bool split = src1 != nullptr;
+    const float alpha = op_params_f[2];
+    const float limit = op_params_f[3];
 
     GGML_ASSERT(ggml_is_contiguous(src0));
 
@@ -8134,7 +8225,15 @@ static void ggml_vk_glu(ggml_backend_vk_context * ctx, vk_context& subctx, const
 
     const uint32_t mode = split ? 2 : (swapped ? 1 : 0);
 
-    ggml_vk_op_f32<vk_op_glu_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_GLU, { (uint32_t)ggml_nelements(dst), (uint32_t)src0->ne[0], (uint32_t)dst->ne[0], mode }, dryrun);
+    ggml_vk_op_f32<vk_op_glu_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_GLU,
+        {
+            (uint32_t)ggml_nelements(dst),
+            (uint32_t)src0->ne[0],
+            (uint32_t)dst->ne[0],
+            mode,
+            alpha,
+            limit
+        }, dryrun);
 }
 
 static void ggml_vk_diag_mask_inf(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
@@ -8142,7 +8241,7 @@ static void ggml_vk_diag_mask_inf(ggml_backend_vk_context * ctx, vk_context& sub
     ggml_vk_op_f32<vk_op_diag_mask_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_DIAG_MASK_INF, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0] }, dryrun);
 }
 
-static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool dryrun = false) {
     float * op_params = (float *)dst->op_params;
 
     float scale = op_params[0];
@@ -8164,7 +8263,7 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx,
     const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
-    ggml_vk_op_f32<vk_op_soft_max_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SOFT_MAX, {
+    ggml_vk_op_f32<vk_op_soft_max_push_constants>(ctx, subctx, src0, src1, src2, dst, GGML_OP_SOFT_MAX, {
         ncols,
         src1 != nullptr ? nrows_y : (uint32_t)0,
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],
@@ -8174,6 +8273,7 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx,
         m0, m1,
         n_head_log2,
         nrows_x,
+        src2 != nullptr
     }, dryrun);
 }
 
@@ -9413,6 +9513,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         case GGML_GLU_OP_GEGLU:
         case GGML_GLU_OP_REGLU:
         case GGML_GLU_OP_SWIGLU:
+        case GGML_GLU_OP_SWIGLU_OAI:
         case GGML_GLU_OP_GEGLU_ERF:
         case GGML_GLU_OP_GEGLU_QUICK:
             break;
@@ -9424,6 +9525,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_REPEAT_BACK:
     case GGML_OP_GET_ROWS:
     case GGML_OP_ADD:
+    case GGML_OP_ADD_ID:
     case GGML_OP_ACC:
     case GGML_OP_SUB:
     case GGML_OP_MUL:
@@ -9578,6 +9680,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_DIV:
         ggml_vk_div(ctx, compute_ctx, src0, src1, node, dryrun);
 
+        break;
+    case GGML_OP_ADD_ID:
+        ggml_vk_add_id(ctx, compute_ctx, src0, src1, src2, node, dryrun);
+
         break;
     case GGML_OP_CONCAT:
         ggml_vk_concat(ctx, compute_ctx, src0, src1, node, dryrun);
@@ -9675,6 +9781,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         case GGML_GLU_OP_GEGLU:
         case GGML_GLU_OP_REGLU:
         case GGML_GLU_OP_SWIGLU:
+        case GGML_GLU_OP_SWIGLU_OAI:
         case GGML_GLU_OP_GEGLU_ERF:
         case GGML_GLU_OP_GEGLU_QUICK:
             ggml_vk_glu(ctx, compute_ctx, src0, src1, node, dryrun);
@@ -9688,7 +9795,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
 
         break;
     case GGML_OP_SOFT_MAX:
-        ggml_vk_soft_max(ctx, compute_ctx, src0, src1, node, dryrun);
+        ggml_vk_soft_max(ctx, compute_ctx, src0, src1, src2, node, dryrun);
 
         break;
     case GGML_OP_SOFT_MAX_BACK:
@@ -9834,6 +9941,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
     case GGML_OP_SUB:
     case GGML_OP_MUL:
     case GGML_OP_DIV:
+    case GGML_OP_ADD_ID:
     case GGML_OP_CONCAT:
     case GGML_OP_UPSCALE:
     case GGML_OP_SCALE:
@@ -9903,6 +10011,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
         case GGML_GLU_OP_GEGLU:
         case GGML_GLU_OP_REGLU:
         case GGML_GLU_OP_SWIGLU:
+        case GGML_GLU_OP_SWIGLU_OAI:
         case GGML_GLU_OP_GEGLU_ERF:
         case GGML_GLU_OP_GEGLU_QUICK:
             buf = tensor->buffer;
@@ -10752,6 +10861,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 case GGML_GLU_OP_GEGLU:
                 case GGML_GLU_OP_REGLU:
                 case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_SWIGLU_OAI:
                 case GGML_GLU_OP_GEGLU_ERF:
                 case GGML_GLU_OP_GEGLU_QUICK:
                     return ggml_is_contiguous(op->src[0]) &&
@@ -10797,6 +10907,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     case GGML_TYPE_IQ3_S:
                     case GGML_TYPE_IQ4_XS:
                     case GGML_TYPE_IQ4_NL:
+                    case GGML_TYPE_MXFP4:
                         break;
                     default:
                         return false;
@@ -10834,6 +10945,10 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 if (head_sizes == FA_HEAD_SIZE_UNSUPPORTED) {
                     return false;
                 }
+                // TODO: support attention sinks [TAG_ATTN_SINKS]
+                if (op->src[4]) {
+                    return false;
+                }
                 if (op->src[0]->type != GGML_TYPE_F32) {
                     return false;
                 }
@@ -10906,6 +11021,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     case GGML_TYPE_IQ3_S:
                     case GGML_TYPE_IQ4_XS:
                     case GGML_TYPE_IQ4_NL:
+                    case GGML_TYPE_MXFP4:
                         return true;
                     default:
                         return false;
@@ -11004,6 +11120,9 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
             return (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
                    (op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F16) &&
                    (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16);
+        case GGML_OP_ADD_ID:
+            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->src[2]->type == GGML_TYPE_I32 &&
+                   op->type == GGML_TYPE_F32;
         case GGML_OP_SILU_BACK:
         case GGML_OP_RMS_NORM_BACK:
         case GGML_OP_SQR:
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/add_id.comp b/ggml/src/ggml-vulkan/vulkan-shaders/add_id.comp
new file mode 100644
index 00000000000..3ae8f0116c2
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/add_id.comp
@@ -0,0 +1,42 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : require
+
+#include "types.comp"
+
+layout (push_constant) uniform parameter
+{
+    uint ne0;
+    uint ne1;
+    uint s01;
+    uint s02;
+    uint s11;
+    uint s21;
+} p;
+
+#define BLOCK_SIZE 512
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer Y {B_TYPE data_b[];};
+layout (binding = 2) readonly buffer Z {int32_t data_c[];};
+layout (binding = 3) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i1 = gl_WorkGroupID.x;
+    const uint i2 = gl_WorkGroupID.y;
+
+    const uint i11 = data_c[i1 + i2 * p.s21];
+
+    const uint s1 = p.ne0;
+    const uint s2 = p.ne0 * p.ne1;
+
+    const uint d0 = i1 * s1 + i2 * s2;
+    const uint a0 = i1 * p.s01 + i2 * p.s02;
+    const uint b0 = i11 * p.s11;
+
+    for (uint i0 = gl_LocalInvocationID.x; i0 < p.ne0; i0 += BLOCK_SIZE) {
+        data_d[d0 + i0] = data_a[a0 + i0] + data_b[b0 + i0];
+    }
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp b/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp
index dbc7daa3328..978d4300307 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp
@@ -4,8 +4,8 @@
 #include "generic_unary_head.comp"
 #include "dequant_funcs.comp"
 
-#if defined(DATA_A_IQ4_NL)
-// 16 invocations needed for init_iq4nl_shmem
+#if defined(DATA_A_IQ4_NL) || defined(DATA_A_MXFP4)
+// 16 invocations needed for init_iq_shmem
 layout(local_size_x = 16, local_size_y = 1, local_size_z = 1) in;
 #else
 layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp
index 0d9739d4060..d3127fbd986 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp
@@ -434,6 +434,18 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
 }
 #endif
 
+#if defined(DATA_A_MXFP4)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
+    return vec2(kvalues_mxfp4[vui & 0xF], kvalues_mxfp4[vui >> 4]);
+}
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    vec2 v0 = dequantize(ib, iqs, a_offset);
+    vec2 v1 = dequantize(ib, iqs + 1, a_offset);
+    return vec4(v0.x, v0.y, v1.x, v1.y);
+}
+#endif
+
 #if defined(DATA_A_F32) || defined(DATA_A_F16) || defined(DATA_A_BF16)
 vec2 get_dm(uint ib, uint a_offset) {
     return vec2(0, 0);
@@ -455,6 +467,12 @@ vec2 get_dm(uint ib, uint a_offset) {
 }
 #endif
 
+#if defined(DATA_A_MXFP4)
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(e8m0_to_fp32(data_a[a_offset + ib].e), 0);
+}
+#endif
+
 #if defined(DATA_A_Q4_1) || defined(DATA_A_Q5_1)
 vec2 get_dm(uint ib, uint a_offset) {
     return vec2(float(data_a[a_offset + ib].d), float(data_a[a_offset + ib].m));
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp
index 9cb7da2daab..706540fd851 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp
@@ -654,6 +654,25 @@ float16_t dequantFuncIQ4_NL(const in decodeBufIQ4_NL bl, const in uint blockCoor
 }
 #endif
 
+#if defined(DATA_A_MXFP4)
+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufMXFP4 {
+   block_mxfp4 block;
+};
+
+float16_t dequantFuncMXFP4(const in decodeBufMXFP4 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const float d = e8m0_to_fp32(bl.block.e);
+    const uint idx = coordInBlock[1];
+    const uint iqs = idx & 0xF;
+    const uint shift = (idx & 0x10) >> 2;
+    uint32_t qs = bl.block.qs[iqs];
+    qs >>= shift;
+    qs &= 0xF;
+    float16_t ret = float16_t(kvalues_mxfp4[qs] * d);
+    return ret;
+}
+#endif
+
 #if defined(DATA_A_Q4_0)
 #define dequantFuncA dequantFuncQ4_0
 #elif defined(DATA_A_Q4_1)
@@ -696,4 +715,6 @@ float16_t dequantFuncIQ4_NL(const in decodeBufIQ4_NL bl, const in uint blockCoor
 #define dequantFuncA dequantFuncIQ4_XS
 #elif defined(DATA_A_IQ4_NL)
 #define dequantFuncA dequantFuncIQ4_NL
+#elif defined(DATA_A_MXFP4)
+#define dequantFuncA dequantFuncMXFP4
 #endif
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_mxfp4.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_mxfp4.comp
new file mode 100644
index 00000000000..ee496e9d568
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_mxfp4.comp
@@ -0,0 +1,32 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_mxfp4 data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    const uint tid = gl_LocalInvocationID.x % 64;
+    const uint il  = tid/32;
+    const uint ir  = tid%32;
+    const uint ib = 32*i + ir;
+    if (ib >= p.nel / 32) {
+        return;
+    }
+
+    const uint q_idx = 8*il;
+    const uint b_idx = 1024*i + 32*ir + q_idx;
+
+    const float d = e8m0_to_fp32(data_a[ib].e);
+
+    [[unroll]] for (uint l = 0; l < 8; ++l) {
+        data_b[b_idx + l +  0] = D_TYPE(d * kvalues_mxfp4[data_a[ib].qs[q_idx + l] & 0xF]);
+        data_b[b_idx + l + 16] = D_TYPE(d * kvalues_mxfp4[data_a[ib].qs[q_idx + l] >>  4]);
+    }
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.comp b/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.comp
index 004a61fc162..51d70869d95 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.comp
@@ -14,4 +14,6 @@ layout (push_constant) uniform parameter
     uint ne00;
     uint ne20;
     uint mode;
+    float alpha;
+    float limit;
 } p;
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
index f481549911b..8c5114a79d2 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
@@ -747,6 +747,21 @@ void main() {
             buf_a[buf_idx + 1 ] = FLOAT_TYPE(kvalues_iq4nl[bitfieldExtract(vui, 8, 4)]) * d;
             buf_a[buf_idx + 16] = FLOAT_TYPE(kvalues_iq4nl[bitfieldExtract(vui, 4, 4)]) * d;
             buf_a[buf_idx + 17] = FLOAT_TYPE(kvalues_iq4nl[vui >> 12]) * d;
+#elif defined(DATA_A_MXFP4)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + 2 * loadr_a;
+
+            const uint ib = idx / 8;
+            const uint iqs = (idx & 0x07) * 2;
+
+            const float d = e8m0_to_fp32(data_a[ib].e);
+            const uint vui = uint(data_a[ib].qs[iqs]);
+            const uint vui2 = uint(data_a[ib].qs[iqs+1]);
+
+            buf_a[buf_idx     ] = FLOAT_TYPE(kvalues_mxfp4[vui & 0xF] * d);
+            buf_a[buf_idx + 16] = FLOAT_TYPE(kvalues_mxfp4[vui >>  4] * d);
+            buf_a[buf_idx +  1] = FLOAT_TYPE(kvalues_mxfp4[vui2 & 0xF] * d);
+            buf_a[buf_idx + 17] = FLOAT_TYPE(kvalues_mxfp4[vui2 >>  4] * d);
 #endif
         }
         [[unroll]] for (uint l = 0; l < BN; l += loadstride_b) {
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.comp
index 63b15471bd3..34e8db97704 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.comp
@@ -92,6 +92,12 @@ FLOAT_TYPE get_d(uint ib) {
 }
 #endif
 
+#if defined(DATA_A_MXFP4)
+FLOAT_TYPE get_d(uint ib) {
+    return FLOAT_TYPE(e8m0_to_fp32(data_a[ib].e));
+}
+#endif
+
 #if defined(DATA_A_Q4_1) || defined(DATA_A_Q5_1)
 FLOAT_TYPE_VEC2 get_dm(uint ib) {
     return FLOAT_TYPE_VEC2(data_a_packed32[ib].dm);
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp b/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp
index 5bcd3b1e3dd..5f20a1ee7d5 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp
@@ -20,6 +20,7 @@ layout (push_constant) uniform parameter
     float m1;
     uint n_head_log2;
     uint nrows_x;
+    uint has_sinks;
 } p;
 
 #include "types.comp"
@@ -29,7 +30,8 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
 layout (binding = 1) readonly buffer Y {B_TYPE data_b[];};
-layout (binding = 2) buffer D {D_TYPE data_d[];};
+layout (binding = 2) readonly buffer Z {float data_c[];};
+layout (binding = 3) buffer D {D_TYPE data_d[];};
 
 shared FLOAT_TYPE vals[BLOCK_SIZE];
 
@@ -60,13 +62,13 @@ void soft_max(uint num_iters) {
         const uint h = (rowx / p.ne01) % p.ne02; // head index
 
         const float base = h < p.n_head_log2 ? p.m0 : p.m1;
-        const uint   exp  = h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1;
+        const uint   exp = h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1;
 
         slope = pow(base, exp);
     }
 
     // Find max
-    FLOAT_TYPE max_val = uintBitsToFloat(0xFF800000);
+    FLOAT_TYPE max_val = p.has_sinks == 0 ? uintBitsToFloat(0xFF800000) : data_c[i02];
 
     // Cache values while we compute the max, so we don't need to read them
     // again when we're ready to compute exp(x-max).
@@ -148,6 +150,10 @@ void soft_max(uint num_iters) {
     }
     sum = vals[0];
 
+    if (p.has_sinks != 0) {
+        sum += FLOAT_TYPE(exp(FLOAT_TYPE(data_c[i02]) - max_val));
+    }
+
     FLOAT_TYPE rcpdivisor = 1.0/sum;
 
     [[unroll]] for (uint col0 = 0, idx = 0; idx < num_iters; col0 += BLOCK_SIZE, ++idx) {
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/swiglu_oai.comp b/ggml/src/ggml-vulkan/vulkan-shaders/swiglu_oai.comp
new file mode 100644
index 00000000000..970750eec08
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/swiglu_oai.comp
@@ -0,0 +1,14 @@
+#version 450
+
+#include "glu_head.comp"
+
+float op(float a, float b) {
+    float xi = min(a, p.limit);
+    float gi = max(min(b, p.limit), -p.limit);
+
+    float out_glu = xi / (1.0f + exp(-xi * p.alpha));
+    out_glu = out_glu * (1.0f + gi);
+    return out_glu;
+}
+
+#include "glu_main.comp"
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp
index 3bde717832b..a36c33e2676 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp
@@ -1337,6 +1337,29 @@ struct block_iq4_nl_packed16
 #define A_TYPE_PACKED16 block_iq4_nl_packed16
 #endif
 
+#define QUANT_K_MXFP4 32
+#define QUANT_R_MXFP4 2
+
+struct block_mxfp4
+{
+    uint8_t e;
+    uint8_t qs[QUANT_K_MXFP4/2];
+};
+
+//struct block_mxfp4_packed16
+//{
+//    uint8_t e;
+//    uint16_t qs[QUANT_K_MXFP4/2/2];
+//};
+
+#if defined(DATA_A_MXFP4)
+#define QUANT_K QUANT_K_MXFP4
+#define QUANT_R QUANT_R_MXFP4
+#define QUANT_AUXF 1
+#define A_TYPE block_mxfp4
+//#define A_TYPE_PACKED16 block_mxfp4_packed16
+#endif
+
 #if defined(DATA_A_IQ4_NL) || defined(DATA_A_IQ4_XS)
 const int8_t kvalues_iq4nl_const[16] = {
     int8_t(-127), int8_t(-104), int8_t(-83), int8_t(-65), int8_t(-49), int8_t(-35), int8_t(-22), int8_t(-10),
@@ -1356,6 +1379,25 @@ void init_iq_shmem(uvec3 wgsize)
 }
 #endif
 
+#if defined(DATA_A_MXFP4)
+const FLOAT_TYPE kvalues_mxfp4_const[16] = {
+    FLOAT_TYPE(0.0f), FLOAT_TYPE(0.5f), FLOAT_TYPE(1.0f), FLOAT_TYPE(1.5f), FLOAT_TYPE(2.0f), FLOAT_TYPE(3.0f), FLOAT_TYPE(4.0f), FLOAT_TYPE(6.0f),
+    FLOAT_TYPE(-0.0f), FLOAT_TYPE(-0.5f), FLOAT_TYPE(-1.0f), FLOAT_TYPE(-1.5f), FLOAT_TYPE(-2.0f), FLOAT_TYPE(-3.0f), FLOAT_TYPE(-4.0f), FLOAT_TYPE(-6.0f)
+};
+
+shared FLOAT_TYPE kvalues_mxfp4[16];
+
+#define NEEDS_INIT_IQ_SHMEM
+void init_iq_shmem(uvec3 wgsize)
+{
+    // copy the table into shared memory and sync
+    for (uint i = gl_LocalInvocationIndex.x; i < kvalues_mxfp4.length(); i += wgsize.x) {
+        kvalues_mxfp4[i] = kvalues_mxfp4_const[i];
+    }
+    barrier();
+}
+#endif
+
 // returns the bfloat value in the low 16b.
 // See ggml_compute_fp32_to_bf16
 uint32_t fp32_to_bf16(float f)
@@ -1370,4 +1412,17 @@ float bf16_to_fp32(uint32_t u)
     return uintBitsToFloat(u << 16);
 }
 
+float e8m0_to_fp32(uint8_t x) {
+    uint32_t bits;
+
+    if (x == 0) {
+        bits = 0x00400000;
+    } else {
+        bits = x;
+        bits = bits << 23;
+    }
+
+    return uintBitsToFloat(bits);
+}
+
 #endif // !defined(GGML_TYPES_COMP)
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
index c6aa3ea4c79..4cd94c51e3f 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -64,6 +64,7 @@ const std::vector<std::string> type_names = {
     "iq3_s",
     "iq4_xs",
     "iq4_nl",
+    "mxfp4",
     "bf16",
 };
 
@@ -118,7 +119,7 @@ void execute_command(const std::string& command, std::string& stdout_str, std::s
     CloseHandle(pi.hProcess);
     CloseHandle(pi.hThread);
 #else
-int stdout_pipe[2];
+    int stdout_pipe[2];
     int stderr_pipe[2];
 
     if (pipe(stdout_pipe) != 0 || pipe(stderr_pipe) != 0) {
@@ -362,7 +363,7 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
         std::string load_vec_quant = "2";
         if ((tname == "q4_0") || (tname == "q4_1") || (tname == "iq1_s") || (tname == "iq1_m") || (tname == "iq2_xxs") || (tname == "iq2_xs") || (tname == "iq2_s"))
             load_vec_quant = "8";
-        else if ((tname == "q5_0") || (tname == "q5_1") || (tname == "q8_0") || (tname == "iq3_xxs") || (tname == "iq3_s") || (tname == "iq4_nl"))
+        else if ((tname == "q5_0") || (tname == "q5_1") || (tname == "q8_0") || (tname == "iq3_xxs") || (tname == "iq3_s") || (tname == "iq4_nl") || (tname == "mxfp4"))
             load_vec_quant = "4";
 
         if (tname == "bf16") {
@@ -602,6 +603,8 @@ void process_shaders() {
         string_to_spv("reglu_f32" + suffix,      "reglu.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},       {"RTE16", rte ? "1" : "0"}});
         string_to_spv("swiglu_f16" + suffix,     "swiglu.comp",      {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
         string_to_spv("swiglu_f32" + suffix,     "swiglu.comp",      {{"A_TYPE", "float"},       {"D_TYPE", "float"},       {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("swiglu_oai_f16" + suffix, "swiglu_oai.comp",  {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("swiglu_oai_f32" + suffix, "swiglu_oai.comp",  {{"A_TYPE", "float"},       {"D_TYPE", "float"},       {"RTE16", rte ? "1" : "0"}});
         string_to_spv("geglu_erf_f16" + suffix,  "geglu_erf.comp",   {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
         string_to_spv("geglu_erf_f32" + suffix,  "geglu_erf.comp",   {{"A_TYPE", "float"},       {"D_TYPE", "float"},       {"RTE16", rte ? "1" : "0"}});
         string_to_spv("geglu_quick_f16" + suffix,"geglu_quick.comp", {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
@@ -671,6 +674,8 @@ void process_shaders() {
 
     string_to_spv("roll_f32", "roll.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 
+    string_to_spv("add_id_f32", "add_id.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
+
     for (auto &c : compiles) {
         c.wait();
     }
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 124cf3e8b60..55a76f8248c 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -582,9 +582,6 @@ FILE * ggml_fopen(const char * fname, const char * mode) {
 #endif
 
 }
-static void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * GGML_RESTRICT x, size_t bx, const float * GGML_RESTRICT y, size_t by, int nrc);
-static void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * GGML_RESTRICT x, size_t bx, ggml_fp16_t * GGML_RESTRICT y, size_t by, int nrc);
-static void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t * GGML_RESTRICT x, size_t bx, ggml_bf16_t * GGML_RESTRICT y, size_t by, int nrc);
 
 static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
     [GGML_TYPE_I8] = {
@@ -690,6 +687,14 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .is_quantized             = true,
         .from_float_ref           = (ggml_from_float_t) quantize_row_q8_1_ref,
     },
+    [GGML_TYPE_MXFP4] = {
+        .type_name                = "mxfp4",
+        .blck_size                = QK_MXFP4,
+        .type_size                = sizeof(block_mxfp4),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_mxfp4,
+        .from_float_ref           = (ggml_from_float_t)quantize_row_mxfp4_ref,
+    },
     [GGML_TYPE_Q2_K] = {
         .type_name                = "q2_K",
         .blck_size                = QK_K,
@@ -917,6 +922,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
 
     "DUP",
     "ADD",
+    "ADD_ID",
     "ADD1",
     "ACC",
     "SUB",
@@ -1010,13 +1016,14 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "GLU",
 };
 
-static_assert(GGML_OP_COUNT == 86, "GGML_OP_COUNT != 86");
+static_assert(GGML_OP_COUNT == 87, "GGML_OP_COUNT != 87");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
 
     "x",
     "x+y",
+    "x[i]+y",
     "x+y",
     "view(x,nb,offset)+=y->x",
     "x-y",
@@ -1110,7 +1117,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "glu(x)",
 };
 
-static_assert(GGML_OP_COUNT == 86, "GGML_OP_COUNT != 86");
+static_assert(GGML_OP_COUNT == 87, "GGML_OP_COUNT != 87");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
@@ -1140,11 +1147,12 @@ static const char * GGML_GLU_OP_NAME[GGML_GLU_OP_COUNT] = {
     "REGLU",
     "GEGLU",
     "SWIGLU",
+    "SWIGLU_OAI",
     "GEGLU_ERF",
     "GEGLU_QUICK",
 };
 
-static_assert(GGML_GLU_OP_COUNT == 5, "GGML_GLU_OP_COUNT != 5");
+static_assert(GGML_GLU_OP_COUNT == 6, "GGML_GLU_OP_COUNT != 6");
 
 
 static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN");
@@ -1312,6 +1320,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
         case GGML_FTYPE_MOSTLY_Q5_0:          wtype = GGML_TYPE_Q5_0;  break;
         case GGML_FTYPE_MOSTLY_Q5_1:          wtype = GGML_TYPE_Q5_1;  break;
         case GGML_FTYPE_MOSTLY_Q8_0:          wtype = GGML_TYPE_Q8_0;  break;
+        case GGML_FTYPE_MOSTLY_MXFP4:         wtype = GGML_TYPE_MXFP4; break;
         case GGML_FTYPE_MOSTLY_Q2_K:          wtype = GGML_TYPE_Q2_K;  break;
         case GGML_FTYPE_MOSTLY_Q3_K:          wtype = GGML_TYPE_Q3_K;  break;
         case GGML_FTYPE_MOSTLY_Q4_K:          wtype = GGML_TYPE_Q4_K;  break;
@@ -1962,6 +1971,27 @@ struct ggml_tensor * ggml_add_cast(
     return ggml_add_cast_impl(ctx, a, b, type);
 }
 
+struct ggml_tensor * ggml_add_id(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            struct ggml_tensor  * ids) {
+
+    GGML_ASSERT(a->ne[0] == b->ne[0]);
+    GGML_ASSERT(a->ne[1] == ids->ne[0]);
+    GGML_ASSERT(a->ne[2] == ids->ne[1]);
+    GGML_ASSERT(ids->type == GGML_TYPE_I32);
+
+    struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
+
+    result->op     = GGML_OP_ADD_ID;
+    result->src[0] = a;
+    result->src[1] = b;
+    result->src[2] = ids;
+
+    return result;
+}
+
 // ggml_add1
 
 static struct ggml_tensor * ggml_add1_impl(
@@ -2812,6 +2842,19 @@ struct ggml_tensor * ggml_geglu_quick_split(
     return ggml_glu_impl(ctx, a, b, GGML_GLU_OP_GEGLU_QUICK, false);
 }
 
+struct ggml_tensor * ggml_swiglu_oai(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        float                 alpha,
+        float                 limit) {
+    struct ggml_tensor * result = ggml_glu_impl(ctx, a, b, GGML_GLU_OP_SWIGLU_OAI, false);
+    ggml_set_op_params_f32(result, 2, alpha);
+    ggml_set_op_params_f32(result, 3, limit);
+
+    return result;
+}
+
 // ggml_norm
 
 static struct ggml_tensor * ggml_norm_impl(
@@ -3779,6 +3822,22 @@ struct ggml_tensor * ggml_soft_max_ext(
     return ggml_soft_max_impl(ctx, a, mask, scale, max_bias, false);
 }
 
+void ggml_soft_max_add_sinks(
+        struct ggml_tensor * a,
+        struct ggml_tensor * sinks) {
+    if (!sinks) {
+        a->src[2] = NULL;
+        return;
+    }
+
+    GGML_ASSERT(a->op == GGML_OP_SOFT_MAX);
+    GGML_ASSERT(a->src[2] == NULL);
+    GGML_ASSERT(a->src[0]->ne[2] == sinks->ne[0]);
+    GGML_ASSERT(sinks->type == GGML_TYPE_F32);
+
+    a->src[2] = sinks;
+}
+
 // ggml_soft_max_ext_back
 
 static struct ggml_tensor * ggml_soft_max_ext_back_impl(
@@ -4812,6 +4871,22 @@ enum ggml_prec ggml_flash_attn_ext_get_prec(
     return (enum ggml_prec) prec_i32;
 }
 
+void ggml_flash_attn_ext_add_sinks(
+        struct ggml_tensor * a,
+        struct ggml_tensor * sinks) {
+    if (!sinks) {
+        a->src[4] = NULL;
+        return;
+    }
+
+    GGML_ASSERT(a->op == GGML_OP_FLASH_ATTN_EXT);
+    GGML_ASSERT(a->src[4] == NULL);
+    GGML_ASSERT(a->src[0]->ne[2] == sinks->ne[0]);
+    GGML_ASSERT(sinks->type == GGML_TYPE_F32);
+
+    a->src[4] = sinks;
+}
+
 // ggml_flash_attn_back
 
 struct ggml_tensor * ggml_flash_attn_back(
@@ -6872,6 +6947,7 @@ size_t ggml_quantize_chunk(
         case GGML_TYPE_Q5_0:    result = quantize_q5_0(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_Q5_1:    result = quantize_q5_1(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_Q8_0:    result = quantize_q8_0(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_MXFP4:   result = quantize_mxfp4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_Q2_K:    result = quantize_q2_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_Q3_K:    result = quantize_q3_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_Q4_K:    result = quantize_q4_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;

From ea0666cef0f19933274af425e261ac0a4344f25c Mon Sep 17 00:00:00 2001
From: Chenguang Li <757486878@qq.com>
Date: Wed, 6 Aug 2025 14:12:42 +0800
Subject: [PATCH 35/86] CANN: add support for ACL Graph (llama/15065)

* feat(cann): add optional support for ACL Graph execution

This commit adds support for executing ggml computational graphs using
Huawei's ACL graph mode via the USE_CANN_GRAPH flag. The support can be
enabled at compile time using the CMake option:

    -DUSE_CANN_GRAPH=ON

By default, ACL graph execution is **disabled**, and the fallback path
uses node-by-node execution.

Key additions:
- CMake option  to toggle graph mode
- Graph capture and execution logic using
- Tensor property matching to determine whether graph update is required
- Safe fallback and logging if the environment variable LLAMA_SET_ROWS
  is unset or invalid

This prepares the backend for performance improvements in repetitive graph
execution scenarios on Ascend devices.

Signed-off-by: noemotiovon <757486878@qq.com>

* Fix review comments

Signed-off-by: noemotiovon <757486878@qq.com>

* remane USE_CANN_GRAPH to USE_ACL_GRAPH

Signed-off-by: noemotiovon <757486878@qq.com>

* fix typo

Signed-off-by: noemotiovon <757486878@qq.com>

---------

Signed-off-by: noemotiovon <757486878@qq.com>
---
 ggml/src/ggml-cann/CMakeLists.txt |  14 +++
 ggml/src/ggml-cann/common.h       |  36 ++++++
 ggml/src/ggml-cann/ggml-cann.cpp  | 197 +++++++++++++++++++++++++++---
 3 files changed, 228 insertions(+), 19 deletions(-)

diff --git a/ggml/src/ggml-cann/CMakeLists.txt b/ggml/src/ggml-cann/CMakeLists.txt
index 7742b39153f..aee5e7b06e5 100755
--- a/ggml/src/ggml-cann/CMakeLists.txt
+++ b/ggml/src/ggml-cann/CMakeLists.txt
@@ -31,6 +31,13 @@ string(REGEX MATCH "[0-9]+[a-zA-Z]" SOC_TYPE_MAJOR_SN "${SOC_VERSION}")
 set(SOC_TYPE_COMPILE_OPTION "ASCEND_${SOC_TYPE_MAJOR_SN}")
 string(TOUPPER ${SOC_TYPE_COMPILE_OPTION} SOC_TYPE_COMPILE_OPTION)
 message(STATUS "CANN: SOC_VERSION =  ${SOC_VERSION}")
+option(USE_ACL_GRAPH "Enable CANN graph execution (ACL graph mode)" OFF)
+
+if(USE_ACL_GRAPH AND (SOC_TYPE_MAJOR_SN STREQUAL "310P" OR SOC_TYPE_COMPILE_OPTION STREQUAL "ASCEND_310P"))
+    message(FATAL_ERROR
+        "CANN Graph (ACL graph mode) is not supported on 310P devices. "
+        "Please build with -DUSE_ACL_GRAPH=OFF or use a supported SOC.")
+endif()
 
 if (CANN_INSTALL_DIR)
     # Only Support Linux.
@@ -68,6 +75,13 @@ if (CANN_INSTALL_DIR)
 
     target_compile_definitions(ggml-cann PRIVATE "-D${SOC_TYPE_COMPILE_OPTION}")
 
+    if (USE_ACL_GRAPH)
+        target_compile_definitions(ggml-cann PRIVATE USE_ACL_GRAPH)
+        message(STATUS "CANN: USE_ACL_GRAPH is enabled.")
+    else()
+        message(STATUS "CANN: USE_ACL_GRAPH is disabled.")
+    endif()
+
     message(STATUS "CANN: CANN_INCLUDE_DIRS =  ${CANN_INCLUDE_DIRS}")
     message(STATUS "CANN: CANN_LIBRARIES =  ${CANN_LIBRARIES}")
 else()
diff --git a/ggml/src/ggml-cann/common.h b/ggml/src/ggml-cann/common.h
index 8dfe3b061c1..9d294f72b67 100755
--- a/ggml/src/ggml-cann/common.h
+++ b/ggml/src/ggml-cann/common.h
@@ -337,6 +337,29 @@ class cann_task_queue {
     int32_t device_;
 };
 
+#ifdef USE_ACL_GRAPH
+struct ggml_graph_node_properties {
+    void * node_address;
+    ggml_op node_op;
+    int64_t ne[GGML_MAX_DIMS];
+    size_t nb[GGML_MAX_DIMS];
+    void * src_address[GGML_MAX_SRC];
+    int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
+};
+
+struct ggml_cann_graph {
+    ~ggml_cann_graph() {
+        if (graph != nullptr) {
+            aclmdlRIDestroy(graph);
+        }
+    }
+
+    aclmdlRI graph = nullptr;
+
+    std::vector<ggml_graph_node_properties> ggml_graph_properties;
+};
+#endif  // USE_ACL_GRAPH
+
 /**
  * @brief Context for managing CANN backend operations.
  */
@@ -345,8 +368,13 @@ struct ggml_backend_cann_context {
     std::string name;                /**< Name of the device. */
     std::string description;         /**< Description of the device. */
     aclrtEvent copy_event = nullptr; /**< Event for managing copy operations. */
+#ifdef USE_ACL_GRAPH
+    /// Cached CANN ACL graph used for executing the current ggml computation graph.
+    std::unique_ptr<ggml_cann_graph> cann_graph;
+#endif
     cann_task_queue task_queue;
     bool async_mode;
+    bool support_set_rows;
 
     aclrtStream streams[GGML_CANN_MAX_STREAMS] = {nullptr}; /**< Array of streams for the device. */
 
@@ -362,6 +390,14 @@ struct ggml_backend_cann_context {
         async_mode = parse_bool(get_env("GGML_CANN_ASYNC_MODE").value_or(""));
         GGML_LOG_INFO("%s: device %d async operator submission is %s\n", __func__,
             device, async_mode ? "ON" : "OFF");
+
+        support_set_rows = parse_bool(get_env("LLAMA_SET_ROWS").value_or(""));
+        GGML_LOG_INFO("%s: LLAMA_SET_ROWS is %s\n", __func__, support_set_rows ? "ON" : "OFF");
+
+        if (!support_set_rows) {
+            GGML_LOG_INFO("%s: CANN Graph currently only supports execution when LLAMA_SET_ROWS is ON. "
+                    "Falling back to eager mode.\n", __func__);
+        }
     }
 
     /**
diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
index 551c1f7673a..cf575b36750 100755
--- a/ggml/src/ggml-cann/ggml-cann.cpp
+++ b/ggml/src/ggml-cann/ggml-cann.cpp
@@ -2075,6 +2075,160 @@ static void ggml_backend_cann_synchronize(ggml_backend_t backend) {
     ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
 }
 
+#ifdef USE_ACL_GRAPH
+/**
+ * @brief Populate the internal CANN graph node properties from the ggml computation graph.
+ *
+ * This function copies all node attributes (operation type, dimensions, strides, input sources,
+ * and operation parameters) into the cached CANN graph structure for later reuse or comparison.
+ *
+ * @param cann_ctx  The CANN backend context.
+ * @param cgraph    The ggml computational graph.
+ */
+static void set_ggml_graph_node_properties(ggml_backend_cann_context * cann_ctx, ggml_cgraph * cgraph) {
+    for (int node_idx = 0; node_idx < cgraph->n_nodes; node_idx++) {
+        ggml_tensor * node = cgraph->nodes[node_idx];
+        cann_ctx->cann_graph->ggml_graph_properties[node_idx].node_address = node->data;
+        cann_ctx->cann_graph->ggml_graph_properties[node_idx].node_op = node->op;
+
+        for (int dim = 0; dim < GGML_MAX_DIMS; dim++) {
+            cann_ctx->cann_graph->ggml_graph_properties[node_idx].ne[dim] = node->ne[dim];
+            cann_ctx->cann_graph->ggml_graph_properties[node_idx].nb[dim] = node->nb[dim];
+        }
+        for (int src = 0; src < GGML_MAX_SRC; src++) {
+            cann_ctx->cann_graph->ggml_graph_properties[node_idx].src_address[src] =
+                node->src[src] ? node->src[src]->data : nullptr;
+        }
+        memcpy(cann_ctx->cann_graph->ggml_graph_properties[node_idx].op_params, node->op_params, GGML_MAX_OP_PARAMS);
+    }
+}
+
+/**
+ * @brief Check if a ggml tensor node matches a previously captured CANN graph node.
+ *
+ * This function compares all relevant fields (address, op type, shape, source inputs, op params)
+ * to determine whether the current node matches a previously recorded version.
+ *
+ * @param node                  The current ggml tensor node.
+ * @param graph_node_properties The stored properties of a CANN graph node.
+ * @return true if all fields match (excluding GGML_OP_VIEW); false otherwise.
+ */
+static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
+    if (node->data != graph_node_properties->node_address &&
+           node->op != GGML_OP_VIEW) {
+        return false;
+    }
+    if (node->op != graph_node_properties->node_op) {
+        return false;
+    }
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (node->ne[i] != graph_node_properties->ne[i]) {
+            return false;
+        }
+        if (node->nb[i] != graph_node_properties->nb[i]) {
+            return false;
+        }
+    }
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        if (node->src[i] &&
+            node->src[i]->data != graph_node_properties->src_address[i] &&
+            node->op != GGML_OP_VIEW
+        ) {
+            return false;
+        }
+    }
+    if (node->op == GGML_OP_SCALE &&
+        memcmp(graph_node_properties->op_params, node->op_params, GGML_MAX_OP_PARAMS) != 0) {
+        return false;
+    }
+    return true;
+}
+
+/**
+ * @brief Determine if the CANN graph needs to be rebuilt due to graph changes.
+ *
+ * This checks whether the number or properties of ggml graph nodes have changed
+ * compared to the last captured CANN graph. If so, the CANN graph must be re-captured.
+ *
+ * @param cann_ctx  The CANN backend context.
+ * @param cgraph    The current ggml computation graph.
+ * @return true if an update is required; false otherwise.
+ */
+static bool is_cann_graph_update_required(ggml_backend_cann_context * cann_ctx, ggml_cgraph * cgraph) {
+    // The number of nodes is different, so the graph needs to be reconstructed.
+    if (cann_ctx->cann_graph->ggml_graph_properties.size() != (size_t)cgraph->n_nodes) {
+        cann_ctx->cann_graph->ggml_graph_properties.resize(cgraph->n_nodes);
+        return true;
+    }
+
+    // The number of nodes is the same; iterate over each node to check whether they match.
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        bool has_matching_properties = ggml_graph_node_has_matching_properties(
+            cgraph->nodes[i], &cann_ctx->cann_graph->ggml_graph_properties[i]);
+        if(!has_matching_properties) {
+            return true;
+        }
+    }
+    return false;
+}
+#endif  // USE_ACL_GRAPH
+
+/**
+ * @brief Evaluate the computation graph and optionally capture or execute it using CANN graph API.
+ *
+ * If CANN graph execution is enabled and graph capture is required, this function begins
+ * graph capture, runs the graph, ends capture, and stores the captured graph.
+ *
+ * Otherwise, it falls back to op-by-op execution using the CANN compute kernel dispatcher.
+ *
+ * @param cann_ctx                 The CANN backend context.
+ * @param cgraph                   The ggml computation graph.
+ * @param use_cann_graph           Whether to use CANN graph execution.
+ * @param cann_graph_update_required Whether graph capture is needed due to graph changes.
+ */
+static void evaluate_and_capture_cann_graph(ggml_backend_cann_context * cann_ctx, ggml_cgraph * cgraph,
+    bool & use_cann_graph, bool & cann_graph_update_required) {
+#ifdef USE_ACL_GRAPH
+    if (use_cann_graph && cann_graph_update_required) {
+        if (cann_ctx->cann_graph->graph != nullptr) {
+            ACL_CHECK(aclmdlRIDestroy(cann_ctx->cann_graph->graph));
+            cann_ctx->cann_graph->graph = nullptr;
+        }
+        ACL_CHECK(aclmdlRICaptureBegin(cann_ctx->stream(), ACL_MODEL_RI_CAPTURE_MODE_GLOBAL));
+    }
+#endif // USE_ACL_GRAPH
+
+    // Only perform the graph execution if CANN graphs are not enabled, or we are capturing the graph.
+    // With the use of CANN graphs, the execution will be performed by the graph launch.
+    if (!use_cann_graph || cann_graph_update_required) {
+        for (int i = 0; i < cgraph->n_nodes; i++) {
+            ggml_tensor * node = cgraph->nodes[i];
+
+            if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
+                continue;
+            }
+
+            bool ok = ggml_cann_compute_forward(*cann_ctx, node);
+            if (!ok) {
+                GGML_LOG_ERROR("%s: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
+            }
+            GGML_ASSERT(ok);
+        }
+    }
+
+#ifdef USE_ACL_GRAPH
+    if (use_cann_graph && cann_graph_update_required) { // End CANN graph capture
+        ACL_CHECK(aclmdlRICaptureEnd(cann_ctx->stream(), &cann_ctx->cann_graph->graph));
+    }
+
+    if (use_cann_graph) {
+        // Execute graph
+        ACL_CHECK(aclmdlRIExecuteAsync(cann_ctx->cann_graph->graph, cann_ctx->stream()));
+    }
+#endif // USE_ACL_GRAPH
+}
+
+
 /**
  * @brief Computes a computational graph using a CANN backend.
  *
@@ -2091,26 +2245,37 @@ static enum ggml_status ggml_backend_cann_graph_compute(
     ggml_backend_t backend, ggml_cgraph* cgraph) {
     ggml_backend_cann_context* cann_ctx =
         (ggml_backend_cann_context*)backend->context;
-
     ggml_cann_set_device(cann_ctx->device);
-    //release temp buffer create by set tensor.
     release_nz_workspace();
+#ifdef USE_ACL_GRAPH
+    bool use_cann_graph = true;
+    bool cann_graph_update_required = false;
 
-    for (int i = 0; i < cgraph->n_nodes; i++) {
-        ggml_tensor* node = cgraph->nodes[i];
+    // check environment LLAMA_SET_ROWS
+    if (!cann_ctx->support_set_rows) {
+        use_cann_graph = false;
+    }
 
-        if (ggml_is_empty(node) || node->op == GGML_OP_NONE) {
-            continue;
+    if (use_cann_graph) {
+        if (cann_ctx->cann_graph == nullptr) {
+            cann_ctx->cann_graph.reset(new ggml_cann_graph());
+            cann_graph_update_required = true;
         }
 
-        bool ok = ggml_cann_compute_forward(*cann_ctx, node);
-
-        if (!ok) {
-            GGML_LOG_ERROR("%s: error: op not supported %s (%s)\n", __func__,
-                    node->name, ggml_op_name(node->op));
-        }
-        GGML_ASSERT(ok);
+        cann_graph_update_required = is_cann_graph_update_required(cann_ctx, cgraph);
+        set_ggml_graph_node_properties(cann_ctx, cgraph);
     }
+#else
+    bool use_cann_graph = false;
+    bool cann_graph_update_required = false;
+#endif  // USE_ACL_GRAPH
+
+    evaluate_and_capture_cann_graph(
+        cann_ctx,
+        cgraph,
+        use_cann_graph,
+        cann_graph_update_required
+    );
 
     return GGML_STATUS_SUCCESS;
 }
@@ -2226,12 +2391,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 // only support F32 and F16.
                 return false;
             }
-
-            if (!ggml_are_same_shape(op, src) && !ggml_is_contiguous(op)) {
-                // unsupport dst is not contiguous.
-                return false;
-            }
-
             return true;
         } break;
         case GGML_OP_CONT: {

From e6fd6501d9e22351857f985a1f4eaf896409fbc9 Mon Sep 17 00:00:00 2001
From: Diego Devesa <slarengh@gmail.com>
Date: Wed, 6 Aug 2025 05:37:35 -0700
Subject: [PATCH 36/86] ggml : fix fallback to CPU for ununsupported ops
 (llama/15118)

---
 ggml/src/ggml-backend.cpp      |  9 +++++++--
 ggml/src/ggml-cpu/ggml-cpu.cpp | 37 ++++++++++++++++------------------
 ggml/src/ggml-cpu/traits.cpp   |  4 ++--
 ggml/src/ggml-cpu/traits.h     |  2 +-
 4 files changed, 27 insertions(+), 25 deletions(-)

diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
index eaf41e5a6c8..1b9d29e911f 100644
--- a/ggml/src/ggml-backend.cpp
+++ b/ggml/src/ggml-backend.cpp
@@ -1071,6 +1071,11 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                 }
             }
         }
+        // if the node is still unassigned, assign it to the first backend that supports it
+        for (int b = 0; b < sched->n_backends && *cur_backend_id == -1; b++) {
+            ggml_backend_sched_set_if_supported(sched, node, b, cur_backend_id);
+        }
+        GGML_ASSERT(*cur_backend_id != -1);
     }
 
     // pass 5: split graph, find tensors that need to be copied
@@ -1098,7 +1103,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
 
             const int node_backend_id = tensor_backend_id(node);
 
-            assert(node_backend_id != -1); // all nodes should be assigned by now, this can happen if there is no CPU fallback
+            GGML_ASSERT(node_backend_id != -1); // all nodes should be assigned by now, this can happen if there is no CPU fallback
 
             // check if we should start a new split based on the sources of the current node
             bool need_new_split = false;
@@ -1156,7 +1161,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
 
                 size_t src_id = hash_id(src);
                 const int src_backend_id = sched->hv_tensor_backend_ids[src_id];
-                assert(src_backend_id != -1); // all inputs should be assigned by now
+                GGML_ASSERT(src_backend_id != -1); // all inputs should be assigned by now
 
                 if (src->flags & GGML_TENSOR_FLAG_INPUT && sched->n_copies > 1) {
                     if (tensor_id_copy(src_id, src_backend_id, 0) == NULL) {
diff --git a/ggml/src/ggml-cpu/ggml-cpu.cpp b/ggml/src/ggml-cpu/ggml-cpu.cpp
index c9daa4c39e8..e16cdc9d498 100644
--- a/ggml/src/ggml-cpu/ggml-cpu.cpp
+++ b/ggml/src/ggml-cpu/ggml-cpu.cpp
@@ -35,7 +35,7 @@
 
 // ggml-backend interface
 
-std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type() {
+std::vector<ggml_backend_buffer_type_t> & ggml_backend_cpu_get_extra_buffer_types() {
     static std::vector<ggml_backend_buffer_type_t> bufts = []() {
         std::vector<ggml_backend_buffer_type_t> bufts;
 
@@ -57,8 +57,6 @@ std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type
         }
 #endif
 
-        bufts.push_back(NULL);
-
         return bufts;
     }();
 
@@ -66,14 +64,20 @@ std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type
 }
 
 static ggml_backend_buffer_type_t * ggml_backend_cpu_device_get_extra_buffers_type(ggml_backend_dev_t device) {
-    return ggml_backend_cpu_get_extra_buffers_type().data();
+    static std::vector<ggml_backend_buffer_type_t> extra_bufts = [] {
+        std::vector<ggml_backend_buffer_type_t> bufts = ggml_backend_cpu_get_extra_buffer_types();
+        bufts.push_back(nullptr);
+        return bufts;
+    }();
+
+    return extra_bufts.data();
 
     GGML_UNUSED(device);
 }
 
 static bool ggml_backend_cpu_is_extra_buffer_type(ggml_backend_buffer_type_t buft) {
-    for (auto * extra : ggml_backend_cpu_get_extra_buffers_type()) {
-        if (extra && extra == buft) {
+    for (auto * extra : ggml_backend_cpu_get_extra_buffer_types()) {
+        if (extra == buft) {
             return true;
         }
     }
@@ -397,20 +401,13 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
         return true;
     }
 
-    // extra_buffer_op?
-    for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
-        if (extra) {
-            auto buf_extra = (ggml::cpu::extra_buffer_type*) extra->context;
-            if (buf_extra && buf_extra->supports_op(dev, op)) {
-                return true;
-            }
-        }
-    }
-
-    // the other case need host buffer.
-    for (int i = 0; i < GGML_MAX_SRC; i++) {
-        if (op->src[i] && op->src[i]->buffer && !ggml_backend_buft_is_host(op->src[i]->buffer->buft)) {
-            return false;
+    // check extra buffer types
+    // note: only the first sources are checked for extra buffer types to reduce overhead, increase if necessary
+    for (int i = 0; i < 4; i++) {
+        if (op->src[i] && op->src[i]->buffer &&
+            ggml_backend_cpu_is_extra_buffer_type(op->src[i]->buffer->buft)) {
+            auto * buf_extra = (ggml::cpu::extra_buffer_type *) op->src[i]->buffer->buft->context;
+            return buf_extra->supports_op(dev, op);
         }
     }
 
diff --git a/ggml/src/ggml-cpu/traits.cpp b/ggml/src/ggml-cpu/traits.cpp
index 139fa596414..4f32f10255a 100644
--- a/ggml/src/ggml-cpu/traits.cpp
+++ b/ggml/src/ggml-cpu/traits.cpp
@@ -10,7 +10,7 @@ extra_buffer_type::~extra_buffer_type() {}
 }  // namespace ggml::cpu
 
 bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) {
-    for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
+    for (auto extra : ggml_backend_cpu_get_extra_buffer_types()) {
         if (extra && extra->context) {
             auto buf_extra     = (ggml::cpu::extra_buffer_type *) extra->context;
             auto tensor_traits = buf_extra->get_tensor_traits(op);
@@ -23,7 +23,7 @@ bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct
 }
 
 bool ggml_cpu_extra_work_size(int n_threads, const struct ggml_tensor * op, size_t * size) {
-    for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
+    for (auto extra : ggml_backend_cpu_get_extra_buffer_types()) {
         if (extra && extra->context) {
             auto buf_extra     = (ggml::cpu::extra_buffer_type *) extra->context;
             auto tensor_traits = buf_extra->get_tensor_traits(op);
diff --git a/ggml/src/ggml-cpu/traits.h b/ggml/src/ggml-cpu/traits.h
index 99a6186b1d6..f4e0990ddfc 100644
--- a/ggml/src/ggml-cpu/traits.h
+++ b/ggml/src/ggml-cpu/traits.h
@@ -33,6 +33,6 @@ class extra_buffer_type {
 }  // namespace ggml::cpu
 
 // implemented in ggml-cpu.cpp.
-std::vector<ggml_backend_buffer_type_t> & ggml_backend_cpu_get_extra_buffers_type();
+std::vector<ggml_backend_buffer_type_t> & ggml_backend_cpu_get_extra_buffer_types();
 
 #endif

From c4a26c95ad87c0c2adcf0d216686ab78a9643fbf Mon Sep 17 00:00:00 2001
From: lhez <lih@qti.qualcomm.com>
Date: Thu, 7 Aug 2025 04:12:17 +0900
Subject: [PATCH 37/86] opencl: add `swiglu_oai` and `add_id` (llama/15121)

* opencl: add `swiglu-oai`

* opencl: add `add_id`

* opencl: add missing `add_id.cl`
---
 ggml/src/ggml-opencl/CMakeLists.txt    |   1 +
 ggml/src/ggml-opencl/ggml-opencl.cpp   | 112 ++++++++++++++++++++++++-
 ggml/src/ggml-opencl/kernels/add_id.cl |  42 ++++++++++
 ggml/src/ggml-opencl/kernels/glu.cl    |  41 +++++++++
 4 files changed, 194 insertions(+), 2 deletions(-)
 create mode 100644 ggml/src/ggml-opencl/kernels/add_id.cl

diff --git a/ggml/src/ggml-opencl/CMakeLists.txt b/ggml/src/ggml-opencl/CMakeLists.txt
index 3adea836154..d8290faa467 100644
--- a/ggml/src/ggml-opencl/CMakeLists.txt
+++ b/ggml/src/ggml-opencl/CMakeLists.txt
@@ -55,6 +55,7 @@ endfunction()
 
 set(GGML_OPENCL_KERNELS
     add
+    add_id
     argsort
     clamp
     cpy
diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
index bb8b310b983..eea6ad6cab3 100644
--- a/ggml/src/ggml-opencl/ggml-opencl.cpp
+++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
@@ -345,6 +345,7 @@ struct ggml_backend_opencl_context {
     cl_command_queue queue;
 
     cl_program program_add;
+    cl_program program_add_id;
     cl_program program_clamp;
     cl_program program_cpy;
     cl_program program_cvt;
@@ -404,6 +405,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_mul, kernel_mul_row, kernel_mul_f16, kernel_mul_row_f16;
     cl_kernel kernel_div, kernel_div_row, kernel_div_f16, kernel_div_row_f16;
     cl_kernel kernel_sub, kernel_sub_row, kernel_sub_f16, kernel_sub_row_f16;
+    cl_kernel kernel_add_id;
     cl_kernel kernel_scale;
     cl_kernel kernel_silu, kernel_silu_4;
     cl_kernel kernel_gelu, kernel_gelu_4;
@@ -412,7 +414,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_relu;
     cl_kernel kernel_sigmoid_f32, kernel_sigmoid_f16;
     cl_kernel kernel_clamp;
-    cl_kernel kernel_geglu, kernel_reglu, kernel_swiglu, kernel_geglu_erf, kernel_geglu_quick,
+    cl_kernel kernel_geglu, kernel_reglu, kernel_swiglu, kernel_swiglu_oai, kernel_geglu_erf, kernel_geglu_quick,
               kernel_geglu_f16, kernel_reglu_f16, kernel_swiglu_f16, kernel_geglu_erf_f16, kernel_geglu_quick_f16;
     cl_kernel kernel_norm;
     cl_kernel kernel_rms_norm, kernel_rms_norm_mul;
@@ -681,6 +683,22 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // add_id
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "add_id.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("add_id.cl");
+#endif
+        backend_ctx->program_add_id =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_add_id = clCreateKernel(backend_ctx->program_add_id, "kernel_add_id", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
     // clamp
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -787,6 +805,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         CL_CHECK((backend_ctx->kernel_geglu           = clCreateKernel(backend_ctx->program_glu, "kernel_geglu", &err), err));
         CL_CHECK((backend_ctx->kernel_reglu           = clCreateKernel(backend_ctx->program_glu, "kernel_reglu", &err), err));
         CL_CHECK((backend_ctx->kernel_swiglu          = clCreateKernel(backend_ctx->program_glu, "kernel_swiglu", &err), err));
+        CL_CHECK((backend_ctx->kernel_swiglu_oai      = clCreateKernel(backend_ctx->program_glu, "kernel_swiglu_oai", &err), err));
         CL_CHECK((backend_ctx->kernel_geglu_erf       = clCreateKernel(backend_ctx->program_glu, "kernel_geglu_erf", &err), err));
         CL_CHECK((backend_ctx->kernel_geglu_quick     = clCreateKernel(backend_ctx->program_glu, "kernel_geglu_quick", &err), err));
         CL_CHECK((backend_ctx->kernel_geglu_f16       = clCreateKernel(backend_ctx->program_glu, "kernel_geglu_f16", &err), err));
@@ -2467,6 +2486,8 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
             return (op->src[0]->type == op->src[1]->type) &&
                    (op->src[0]->type == op->type) &&
                    (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16);
+        case GGML_OP_ADD_ID:
+            return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
                 case GGML_UNARY_OP_GELU:
@@ -2488,6 +2509,7 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 case GGML_GLU_OP_GEGLU:
                 case GGML_GLU_OP_REGLU:
                 case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_SWIGLU_OAI:
                 case GGML_GLU_OP_GEGLU_ERF:
                 case GGML_GLU_OP_GEGLU_QUICK:
                     return ggml_is_contiguous_1(op->src[0]) && (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16);
@@ -3824,6 +3846,75 @@ static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const
     }
 }
 
+static void ggml_cl_add_id(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    const ggml_tensor * src2 = dst->src[2];
+    GGML_ASSERT(src2);
+    GGML_ASSERT(src2->extra);
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(src2->type == GGML_TYPE_I32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+
+    GGML_ASSERT(ggml_is_contiguous_rows(src0));
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+
+    const cl_ulong nb11 = src1->nb[1];
+
+    const cl_ulong nb21 = src2->nb[1];
+
+    const int ne0 = dst->ne[0];
+    const int ne1 = dst->ne[1];
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extra2 = (ggml_tensor_extra_cl *)src2->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offset2 = extra2->offset + src2->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel = backend_ctx->kernel_add_id;
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra2->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb21));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne0));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne1));
+
+    int nth = MIN(ne00, (int) backend_ctx->get_kernel_workgroup_size(kernel));
+    size_t global_work_size[] = { (size_t)ne01*nth, (size_t)ne02, 1 };
+    size_t local_work_size[] = { (size_t)nth, 1, 1 };
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
 static void ggml_cl_mul(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -7005,6 +7096,9 @@ static void ggml_cl_glu(ggml_backend_t backend, const ggml_tensor * src0, const
                 kernel = backend_ctx->kernel_swiglu_f16;
             }
             break;
+        case GGML_GLU_OP_SWIGLU_OAI:
+            kernel = backend_ctx->kernel_swiglu_oai;
+            break;
         case GGML_GLU_OP_GEGLU_ERF:
             if (dst->type == GGML_TYPE_F32) {
                 kernel = backend_ctx->kernel_geglu_erf;
@@ -7040,7 +7134,10 @@ static void ggml_cl_glu(ggml_backend_t backend, const ggml_tensor * src0, const
 
     const cl_ulong nb1  = dst->nb[1];
 
-    const int swp = ((const int32_t *) dst->op_params)[1];
+    const int   swp   = ggml_get_op_params_i32(dst, 1);
+    const float alpha = ggml_get_op_params_f32(dst, 2);
+    const float limit = ggml_get_op_params_f32(dst, 3);
+
     const int ne00_off = src1 ? 0 : (swp ? ne0 : 0);
     const int ne10_off = src1 ? 0 : (swp ? 0 : ne0);
 
@@ -7057,6 +7154,11 @@ static void ggml_cl_glu(ggml_backend_t backend, const ggml_tensor * src0, const
     CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne00_off));
     CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne10_off));
 
+    if (ggml_get_glu_op(dst) == GGML_GLU_OP_SWIGLU_OAI) {
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float), &limit));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(float), &alpha));
+    }
+
     const size_t nrows = ggml_nrows(src0);
     size_t nth = 512;
     size_t global_work_size[] = {nrows*nth, 1, 1};
@@ -7113,6 +7215,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_add;
             break;
+        case GGML_OP_ADD_ID:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_add_id;
+            break;
         case GGML_OP_MUL:
             if (!any_on_device) {
                 return false;
diff --git a/ggml/src/ggml-opencl/kernels/add_id.cl b/ggml/src/ggml-opencl/kernels/add_id.cl
new file mode 100644
index 00000000000..e9c6d55e6a2
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/add_id.cl
@@ -0,0 +1,42 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// add_id
+//------------------------------------------------------------------------------
+kernel void kernel_add_id(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * src2,
+    ulong         offset2,
+    global char * dst,
+    ulong         offsetd,
+    ulong         nb01,
+    ulong         nb02,
+    ulong         nb11,
+    ulong         nb21,
+    int           ne0,
+    int           ne1
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    src2 = (global char*)((global char*)src2 + offset2);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    int i1 = get_group_id(0);
+    int i2 = get_group_id(1);
+
+    const int i11 = *((global const int *) (src2 + i1*sizeof(int) + i2*nb21));
+
+    const size_t nb1 = ne0 * sizeof(float);
+    const size_t nb2 = ne1 * nb1;
+
+    global float * dst_row  = (global float *)((global char *)dst  + i1*nb1 + i2*nb2);
+    global float * src0_row = (global float *)((global char *)src0 + i1*nb01 + i2*nb02);
+    global float * src1_row = (global float *)((global char *)src1 + i11*nb11);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        dst_row[i0] = src0_row[i0] + src1_row[i0];
+    }
+}
diff --git a/ggml/src/ggml-opencl/kernels/glu.cl b/ggml/src/ggml-opencl/kernels/glu.cl
index 7cca16e6a9e..059a4bbf1ba 100644
--- a/ggml/src/ggml-opencl/kernels/glu.cl
+++ b/ggml/src/ggml-opencl/kernels/glu.cl
@@ -202,6 +202,47 @@ kernel void kernel_swiglu_f16(
     }
 }
 
+//------------------------------------------------------------------------------
+// swiglu_oai
+//------------------------------------------------------------------------------
+kernel void kernel_swiglu_oai(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * dst,
+    ulong         offsetd,
+    ulong         nb01,
+    ulong         nb11,
+    int           ne0,
+    ulong         nb1,
+    int           ne00_off,
+    int           ne10_off,
+    float         limit,
+    float         alpha
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global float * src0_row = (global float *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global float * src1_row = (global float *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global float * dst_row  = (global float *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        float x0 = src0_row[i0];
+        float x1 = src1_row[i0];
+
+        x0 = min(x0, limit);
+        x1 = max(min(x1, limit), -limit);
+
+        float out_glu = x0 / (1.0f + exp(-x0 * alpha));
+        out_glu = out_glu * (1.0f + x1);
+
+        dst_row[i0] = out_glu;
+    }
+}
+
 //------------------------------------------------------------------------------
 // geglu_erf
 //------------------------------------------------------------------------------

From bbb0da7a30dc1a5f1d2e49c6ad56a4443c6a588d Mon Sep 17 00:00:00 2001
From: rmatif <kingrealriadh@gmail.com>
Date: Wed, 6 Aug 2025 23:17:51 +0200
Subject: [PATCH 38/86] fix profiling crash (llama/15072)

---
 ggml/src/ggml-opencl/ggml-opencl.cpp | 1 +
 1 file changed, 1 insertion(+)

diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
index eea6ad6cab3..4f765ab5309 100644
--- a/ggml/src/ggml-opencl/ggml-opencl.cpp
+++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
@@ -602,6 +602,7 @@ struct ggml_backend_opencl_context {
         if (ref_count == 0) {
 #ifdef GGML_OPENCL_PROFILING
             write_profiling_info();
+            profiling_info.clear();
 #endif
         }
     }

From 75dbb4d76f1b13833cbe12fdea28ffb4619325b8 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= <johannesg@5d6.de>
Date: Thu, 7 Aug 2025 10:53:21 +0200
Subject: [PATCH 39/86] CUDA: GEMM for FP32/FP16/BF16 and ne11 <= 16
 (llama/15131)

* CUDA: GEMM for FP32/FP16/BF16 and ne11 <= 16
---
 ggml/src/ggml-cuda/common.cuh        |  12 +-
 ggml/src/ggml-cuda/fattn-mma-f16.cuh |  12 +-
 ggml/src/ggml-cuda/fattn.cu          |   4 +-
 ggml/src/ggml-cuda/ggml-cuda.cu      |  31 +-
 ggml/src/ggml-cuda/mma.cuh           | 110 ++++--
 ggml/src/ggml-cuda/mmf.cu            | 431 ++++++++++++++++++++++
 ggml/src/ggml-cuda/mmf.cuh           |   5 +
 ggml/src/ggml-cuda/mmq.cu            |   2 +-
 ggml/src/ggml-cuda/mmq.cuh           | 264 +++++++-------
 ggml/src/ggml-cuda/mmvf.cu           | 510 +++++++++++++++++++++++++++
 ggml/src/ggml-cuda/mmvf.cuh          |  11 +
 ggml/src/ggml-cuda/vendors/hip.h     |   1 +
 ggml/src/ggml-cuda/vendors/musa.h    |   3 +-
 13 files changed, 1219 insertions(+), 177 deletions(-)
 create mode 100644 ggml/src/ggml-cuda/mmf.cu
 create mode 100644 ggml/src/ggml-cuda/mmf.cuh
 create mode 100644 ggml/src/ggml-cuda/mmvf.cu
 create mode 100644 ggml/src/ggml-cuda/mmvf.cuh

diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
index 8f27255476d..2e5d48797fa 100644
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -233,9 +233,13 @@ typedef float2 dfloat2;
 #endif // defined(GGML_USE_HIP) && defined(CDNA) && !defined(GGML_HIP_NO_MMQ_MFMA)
 
 #if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
-#define NEW_MMA_AVAILABLE
+#define TURING_MMA_AVAILABLE
 #endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
 
+#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#define AMPERE_MMA_AVAILABLE
+#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+
 #if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #define CP_ASYNC_AVAILABLE
 #endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
@@ -303,10 +307,14 @@ static bool amd_mfma_available(const int cc) {
 }
 
 // Volta technically had FP16 tensor cores but they work very differently compared to Turing and later.
-static bool new_mma_available(const int cc) {
+static bool turing_mma_available(const int cc) {
     return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_TURING;
 }
 
+static bool ampere_mma_available(const int cc) {
+    return cc < GGML_CUDA_CC_OFFSET_AMD && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_AMPERE;
+}
+
 static bool cp_async_available(const int cc) {
     return cc < GGML_CUDA_CC_OFFSET_AMD && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_AMPERE;
 }
diff --git a/ggml/src/ggml-cuda/fattn-mma-f16.cuh b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
index e7570f9d3b8..37125384417 100644
--- a/ggml/src/ggml-cuda/fattn-mma-f16.cuh
+++ b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
@@ -418,7 +418,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         float        * const __restrict__ KQ_max,
         float        * const __restrict__ KQ_rowsum,
         const int kb0) {
-#ifdef NEW_MMA_AVAILABLE
+#ifdef TURING_MMA_AVAILABLE
     typedef fattn_mma_f16_config<DKQ, DV> c;
 
 #ifdef CP_ASYNC_AVAILABLE
@@ -776,7 +776,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
     GGML_UNUSED(VKQ_C); GGML_UNUSED(KQ_max); GGML_UNUSED(KQ_rowsum);
     GGML_UNUSED(kb0); GGML_UNUSED(tile_Q);
     NO_DEVICE_CODE;
-#endif // NEW_MMA_AVAILABLE
+#endif // TURING_MMA_AVAILABLE
 }
 
 template<int DKQ, int DV, int ncols1, int ncols2, int nwarps, int ntiles, bool use_logit_softcap, bool mla, bool needs_fixup, bool is_fixup>
@@ -800,7 +800,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         const int jt,
         const int kb0_start,
         const int kb0_stop) {
-#ifdef NEW_MMA_AVAILABLE
+#ifdef TURING_MMA_AVAILABLE
     //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
 
     typedef fattn_mma_f16_config<DKQ, DV> c;
@@ -1196,7 +1196,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
     GGML_UNUSED(stride_Q2); GGML_UNUSED(stride_K); GGML_UNUSED(stride_V); GGML_UNUSED(stride_mask);
     GGML_UNUSED(jt); GGML_UNUSED(kb0_start); GGML_UNUSED(kb0_stop);
     NO_DEVICE_CODE;
-#endif // NEW_MMA_AVAILABLE
+#endif // TURING_MMA_AVAILABLE
 }
 
 template<int DKQ, int DV, int ncols1, int ncols2, int nwarps, int ntiles, bool use_logit_softcap, bool mla>
@@ -1223,7 +1223,7 @@ static __global__ void flash_attn_ext_f16(
                             const int32_t nb21, const int32_t nb22, const int64_t nb23,
                             const int32_t ne31, const int32_t ne32, const int32_t ne33,
                             const int32_t nb31, const int32_t nb32, const int64_t nb33) {
-#if defined(FLASH_ATTN_AVAILABLE) && defined(NEW_MMA_AVAILABLE)
+#if defined(FLASH_ATTN_AVAILABLE) && defined(TURING_MMA_AVAILABLE)
 
     // Skip unused kernel variants for faster compilation:
     if (use_logit_softcap && !(DKQ == 128 || DKQ == 256)) {
@@ -1354,7 +1354,7 @@ static __global__ void flash_attn_ext_f16(
     GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
     GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33);
     NO_DEVICE_CODE;
-#endif // defined(FLASH_ATTN_AVAILABLE) && defined(NEW_MMA_AVAILABLE)
+#endif // defined(FLASH_ATTN_AVAILABLE) && defined(TURING_MMA_AVAILABLE)
 }
 
 template <int DKQ, int DV, int ncols1, int ncols2>
diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu
index 656e04a4735..8ddd0415b7f 100644
--- a/ggml/src/ggml-cuda/fattn.cu
+++ b/ggml/src/ggml-cuda/fattn.cu
@@ -327,7 +327,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
     const bool gqa_opt_applies = ((Q->ne[2] / K->ne[2]) % 2 == 0) && mask; // The mma-based kernels have GQA-specific optimizations
     const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16;
     const bool mma_faster_for_rtx4000 = Q->ne[3] > 1 || (Q->ne[2] > 4*K->ne[2] && K->ne[1] >= 8192);
-    const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies && !mma_needs_data_conversion &&
+    const bool mma_faster_for_bs1 = turing_mma_available(cc) && gqa_opt_applies && !mma_needs_data_conversion &&
         (cc < GGML_CUDA_CC_ADA_LOVELACE || mma_faster_for_rtx4000);
     const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % (2*warp_size) == 0;
     if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1) {
@@ -340,7 +340,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
     }
 
     // The MMA implementation needs Turing or newer, use the old WMMA code for Volta:
-    if (fp16_mma_available(cc) && !new_mma_available(cc)) {
+    if (fp16_mma_available(cc) && !turing_mma_available(cc)) {
         ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
         return;
     }
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index 60e481b95af..ec7ab255188 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -22,8 +22,9 @@
 #include "ggml-cuda/fattn.cuh"
 #include "ggml-cuda/getrows.cuh"
 #include "ggml-cuda/im2col.cuh"
+#include "ggml-cuda/mmf.cuh"
 #include "ggml-cuda/mmq.cuh"
-#include "ggml-cuda/mmv.cuh"
+#include "ggml-cuda/mmvf.cuh"
 #include "ggml-cuda/mmvq.cuh"
 #include "ggml-cuda/norm.cuh"
 #include "ggml-cuda/opt-step-adamw.cuh"
@@ -2008,7 +2009,9 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     const bool bad_padding_clear = ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE
         && ggml_nbytes(src0) != ggml_backend_buffer_get_alloc_size(src0->buffer, src0) && src0->view_src;
 
-    bool use_mul_mat_vec   = (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16)
+    bool use_mul_mat_vec_f = (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
+    bool use_mul_mat_f     = !ggml_is_quantized(src0->type)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
     bool use_mul_mat_vec_q = ggml_is_quantized(src0->type) && !bad_padding_clear
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
@@ -2028,14 +2031,18 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
             }
 
             const int cc            = ggml_cuda_info().devices[id].cc;
+            const int warp_size     = ggml_cuda_info().devices[id].warp_size;
             use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-            use_mul_mat_vec         = use_mul_mat_vec           && ggml_cuda_should_use_mmv(src0->type, cc, src0->ne, src1->ne[1]);
+            use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src1->ne[1]);
+            use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src1->ne[1]);
             any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
         }
     } else {
         const int cc            = ggml_cuda_info().devices[ctx.device].cc;
+        const int warp_size     = ggml_cuda_info().devices[ctx.device].warp_size;
         use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-        use_mul_mat_vec         = use_mul_mat_vec           && ggml_cuda_should_use_mmv(src0->type, cc, src0->ne, src1->ne[1]);
+        use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src1->ne[1]);
+        use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src1->ne[1]);
         any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
     }
 
@@ -2048,15 +2055,17 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
 
     //TODO update for generic tensor parallelism
-    const int cc                     = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    const int cc                 = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
     bool use_batched_cublas_f16  = src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16);
     bool use_batched_cublas_bf16 = src0->type == GGML_TYPE_BF16 && bf16_mma_hardware_available(cc);
     bool use_batched_cublas_f32  = src0->type == GGML_TYPE_F32;
 
-    if (!split && use_mul_mat_vec) {
+    if (!split && use_mul_mat_vec_f) {
         // the custom F16 vector kernel can be used over batched cuBLAS GEMM
         // but this is only faster for GPUs without tensor cores or with a thin src0 matrix (particularly KQV in attention)
-        ggml_cuda_mul_mat_vec(ctx, src0, src1, nullptr, dst);
+        ggml_cuda_mul_mat_vec_f(ctx, src0, src1, nullptr, dst);
+    } else if (!split && use_mul_mat_f) {
+        ggml_cuda_mul_mat_f(ctx, src0, src1, nullptr, dst);
     } else if (!split && use_mul_mat_vec_q) {
         ggml_cuda_mul_mat_vec_q(ctx, src0, src1, nullptr, dst);
     } else if (!split && use_mul_mat_q) {
@@ -2065,8 +2074,8 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
         && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
         // general KQ + KQV multi-batch without FlashAttention
         ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
-    } else if (use_mul_mat_vec) {
-        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec, nullptr);
+    } else if (use_mul_mat_vec_f) {
+        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec_f, nullptr);
     } else if (use_mul_mat_vec_q) {
         ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, quantize_row_q8_1_cuda);
     } else if (use_mul_mat_q) {
@@ -2094,7 +2103,7 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
             if (ggml_is_quantized(src0->type)) {
                 ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst);
             } else {
-                ggml_cuda_mul_mat_vec(ctx, src0, src1, ids, dst);
+                ggml_cuda_mul_mat_vec_f(ctx, src0, src1, ids, dst);
             }
             return;
         }
@@ -3516,7 +3525,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
 #endif // FLASH_ATTN_AVAILABLE
             if (op->src[1]->ne[0] != op->src[2]->ne[0]) {
                 const int cc = ggml_cuda_info().devices[dev_ctx->device].cc;
-                if (!new_mma_available(cc)) {
+                if (!turing_mma_available(cc)) {
                     return false;
                 }
                 const int gqa_ratio = op->src[0]->ne[2] / op->src[1]->ne[2];
diff --git a/ggml/src/ggml-cuda/mma.cuh b/ggml/src/ggml-cuda/mma.cuh
index a86365c6a06..83ee16b27d0 100644
--- a/ggml/src/ggml-cuda/mma.cuh
+++ b/ggml/src/ggml-cuda/mma.cuh
@@ -23,13 +23,13 @@
 static __device__ __forceinline__ int ggml_cuda_movmatrix(const int x) {
     int ret = 0;
 
-#ifdef NEW_MMA_AVAILABLE
+#ifdef TURING_MMA_AVAILABLE
     asm("movmatrix.sync.aligned.m8n8.trans.b16 %0, %1;"
         : "=r"(ret) : "r"(x));
 #else
     GGML_UNUSED(x);
     NO_DEVICE_CODE;
-#endif // defined(NEW_MMA_AVAILABLE)
+#endif // defined(TURING_MMA_AVAILABLE)
     return ret;
 }
 
@@ -167,6 +167,38 @@ namespace ggml_cuda_mma {
         }
     };
 
+    template <int I_, int J_>
+    struct tile<I_, J_, nv_bfloat162> {
+        static constexpr int I  = I_;
+        static constexpr int J  = J_;
+        static constexpr int ne = I * J / WARP_SIZE;
+        nv_bfloat162 x[ne] = {{0.0f, 0.0f}};
+
+        static __device__ __forceinline__ int get_i(const int l) {
+            if constexpr (I == 8 && J == 8) {
+                return threadIdx.x / 4;
+            } else if constexpr (I == 16 && J == 4) {
+                return l * 8 + threadIdx.x / 4;
+            } else if constexpr (I == 16 && J == 8) {
+                return (l % 2) * 8 + threadIdx.x / 4;
+            } else {
+                static_assert(I == -1 && J == -1, "template specialization not implemented");
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr (I == 8 && J == 8) {
+                return l * 4 + threadIdx.x % 4;
+            } else if constexpr (I == 16 && J == 4) {
+                return threadIdx.x % 4;
+            } else if constexpr (I == 16 && J == 8) {
+                return (l / 2) * 4 + threadIdx.x % 4;
+            } else {
+                static_assert(I == -1 && J == -1, "template specialization not implemented");
+            }
+        }
+    };
+
     template <int I, int J>
     static __device__ __forceinline__ tile<I, J/2, half2> get_half2(const tile<I, J, float> & tile_float) {
         tile<I, J/2, half2> ret;
@@ -209,7 +241,7 @@ namespace ggml_cuda_mma {
     template <typename T>
     static __device__ __forceinline__ void load_ldmatrix(
             tile<8, 8, T> & t, const T * __restrict__ xs0, const int stride) {
-#ifdef NEW_MMA_AVAILABLE
+#ifdef TURING_MMA_AVAILABLE
         int * xi = (int *) t.x;
         const int * xs = (const int *) xs0 + (threadIdx.x % t.I) * stride + ((threadIdx.x / t.I) * (t.J / 2)) % t.J;
         asm volatile("ldmatrix.sync.aligned.m8n8.x2.b16 {%0, %1}, [%2];"
@@ -217,13 +249,13 @@ namespace ggml_cuda_mma {
             : "l"(xs));
 #else
         load_generic(t, xs0, stride);
-#endif // NEW_MMA_AVAILABLE
+#endif // TURING_MMA_AVAILABLE
     }
 
     template <typename T>
     static __device__ __forceinline__ void load_ldmatrix(
             tile<16, 4, T> & t, const T * __restrict__ xs0, const int stride) {
-#ifdef NEW_MMA_AVAILABLE
+#ifdef TURING_MMA_AVAILABLE
         int * xi = (int *) t.x;
         const int * xs = (const int *) xs0 + (threadIdx.x % t.I) * stride;
         asm volatile("ldmatrix.sync.aligned.m8n8.x2.b16 {%0, %1}, [%2];"
@@ -232,13 +264,13 @@ namespace ggml_cuda_mma {
 #else
         load_generic(xs0, stride);
         GGML_UNUSED(t);
-#endif // NEW_MMA_AVAILABLE
+#endif // TURING_MMA_AVAILABLE
     }
 
     template <typename T>
     static __device__ __forceinline__ void load_ldmatrix(
             tile<16, 8, T> & t, const T * __restrict__ xs0, const int stride) {
-#if defined(NEW_MMA_AVAILABLE)
+#if defined(TURING_MMA_AVAILABLE)
         int * xi = (int * ) t.x;
         const int * xs = (const int *) xs0 + (threadIdx.x % t.I) * stride + (threadIdx.x / t.I) * (t.J / 2);
         asm volatile("ldmatrix.sync.aligned.m8n8.x4.b16 {%0, %1, %2, %3}, [%4];"
@@ -246,13 +278,13 @@ namespace ggml_cuda_mma {
             : "l"(xs));
 #else
         load_generic(t, xs0, stride);
-#endif // NEW_MMA_AVAILABLE
+#endif // TURING_MMA_AVAILABLE
     }
 
     template <typename T>
     static __device__ __forceinline__ void load_ldmatrix_trans(
             tile<16, 8, T> & t, const T * __restrict__ xs0, const int stride) {
-#ifdef NEW_MMA_AVAILABLE
+#ifdef TURING_MMA_AVAILABLE
         int * xi = (int * ) t.x;
         const int * xs = (const int *) xs0 + (threadIdx.x % t.I) * stride + (threadIdx.x / t.I) * (t.J / 2);
         asm volatile("ldmatrix.sync.aligned.m8n8.x4.trans.b16 {%0, %1, %2, %3}, [%4];"
@@ -263,12 +295,12 @@ namespace ggml_cuda_mma {
         GGML_UNUSED(xs0);
         GGML_UNUSED(stride);
         NO_DEVICE_CODE;
-#endif // NEW_MMA_AVAILABLE
+#endif // TURING_MMA_AVAILABLE
     }
 
     static __device__ __forceinline__ void mma(
             tile<16, 8, int> & D, const tile<16, 4, int> & A, const tile<8, 4, int> & B) {
-#ifdef NEW_MMA_AVAILABLE
+#ifdef TURING_MMA_AVAILABLE
 #if __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
         asm("mma.sync.aligned.m16n8k16.row.col.s32.s8.s8.s32 {%0, %1, %2, %3}, {%4, %5}, {%6}, {%0, %1, %2, %3};"
             : "+r"(D.x[0]), "+r"(D.x[1]), "+r"(D.x[2]), "+r"(D.x[3])
@@ -287,12 +319,12 @@ namespace ggml_cuda_mma {
         GGML_UNUSED(A);
         GGML_UNUSED(B);
         NO_DEVICE_CODE;
-#endif // NEW_MMA_AVAILABLE
+#endif // TURING_MMA_AVAILABLE
     }
 
     static __device__ __forceinline__ void mma(
             tile<16, 8, int> & D, const tile<16, 8, int> & A, const tile<8, 8, int> & B) {
-#ifdef NEW_MMA_AVAILABLE
+#ifdef TURING_MMA_AVAILABLE
 #if __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
         asm("mma.sync.aligned.m16n8k32.row.col.s32.s8.s8.s32 {%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%0, %1, %2, %3};"
             : "+r"(D.x[0]), "+r"(D.x[1]), "+r"(D.x[2]), "+r"(D.x[3])
@@ -317,12 +349,12 @@ namespace ggml_cuda_mma {
         GGML_UNUSED(A);
         GGML_UNUSED(B);
         NO_DEVICE_CODE;
-#endif // NEW_MMA_AVAILABLE
+#endif // TURING_MMA_AVAILABLE
     }
 
     static __device__ __forceinline__ void mma(
             tile<16, 4, half2> & D, const tile<16, 8, half2> & A, const tile<8, 8, half2> & B) {
-#ifdef NEW_MMA_AVAILABLE
+#ifdef TURING_MMA_AVAILABLE
         const int * Axi = (const int *) A.x;
         const int * Bxi = (const int *) B.x;
         int       * Dxi = (int       *) D.x;
@@ -344,12 +376,12 @@ namespace ggml_cuda_mma {
         GGML_UNUSED(A);
         GGML_UNUSED(B);
         NO_DEVICE_CODE;
-#endif // NEW_MMA_AVAILABLE
+#endif // TURING_MMA_AVAILABLE
     }
 
     static __device__ __forceinline__ void mma(
             tile<16, 8, half2> & D, const tile<16, 8, half2> & A, const tile<16, 8, half2> & B) {
-#ifdef NEW_MMA_AVAILABLE
+#ifdef TURING_MMA_AVAILABLE
         const int * Axi = (const int *) A.x;
         const int * Bxi = (const int *) B.x;
         int       * Dxi = (int       *) D.x;
@@ -380,12 +412,29 @@ namespace ggml_cuda_mma {
         GGML_UNUSED(A);
         GGML_UNUSED(B);
         NO_DEVICE_CODE;
-#endif // NEW_MMA_AVAILABLE
+#endif // TURING_MMA_AVAILABLE
+    }
+
+    static __device__ __forceinline__ void mma(
+            tile<16, 8, float> & D, const tile<16, 8, float> & A, const tile<8, 8, float> & B) {
+#ifdef AMPERE_MMA_AVAILABLE
+        const int * Axi = (const int *) A.x;
+        const int * Bxi = (const int *) B.x;
+        int       * Dxi = (int       *) D.x;
+        asm("mma.sync.aligned.m16n8k8.row.col.f32.tf32.tf32.f32 {%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%0, %1, %2, %3};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3])
+            : "r"(Axi[0]), "r"(Axi[1]), "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[0]), "r"(Bxi[1]));
+#else
+        GGML_UNUSED(D);
+        GGML_UNUSED(A);
+        GGML_UNUSED(B);
+        NO_DEVICE_CODE;
+#endif // AMPERE_MMA_AVAILABLE
     }
 
     static __device__ __forceinline__ void mma(
             tile<16, 8, float> & D, const tile<16, 8, half2> & A, const tile<8, 8, half2> & B) {
-#ifdef NEW_MMA_AVAILABLE
+#ifdef TURING_MMA_AVAILABLE
         const int * Axi = (const int *) A.x;
         const int * Bxi = (const int *) B.x;
         int       * Dxi = (int       *) D.x;
@@ -407,12 +456,29 @@ namespace ggml_cuda_mma {
         GGML_UNUSED(A);
         GGML_UNUSED(B);
         NO_DEVICE_CODE;
-#endif // NEW_MMA_AVAILABLE
+#endif // TURING_MMA_AVAILABLE
+    }
+
+    static __device__ __forceinline__ void mma(
+            tile<16, 8, float> & D, const tile<16, 8, nv_bfloat162> & A, const tile<8, 8, nv_bfloat162> & B) {
+#ifdef AMPERE_MMA_AVAILABLE
+        const int * Axi = (const int *) A.x;
+        const int * Bxi = (const int *) B.x;
+        int       * Dxi = (int       *) D.x;
+        asm("mma.sync.aligned.m16n8k16.row.col.f32.bf16.bf16.f32 {%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%0, %1, %2, %3};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3])
+            : "r"(Axi[0]), "r"(Axi[1]), "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[0]), "r"(Bxi[1]));
+#else
+        GGML_UNUSED(D);
+        GGML_UNUSED(A);
+        GGML_UNUSED(B);
+        NO_DEVICE_CODE;
+#endif // AMPERE_MMA_AVAILABLE
     }
 
     static __device__ __forceinline__ void mma(
             tile<16, 16, float> & D, const tile<16, 8, half2> & A, const tile<16, 8, half2> & B) {
-#ifdef NEW_MMA_AVAILABLE
+#ifdef TURING_MMA_AVAILABLE
         const int * Axi = (const int *) A.x;
         const int * Bxi = (const int *) B.x;
         int       * Dxi = (int       *) D.x;
@@ -443,7 +509,7 @@ namespace ggml_cuda_mma {
         GGML_UNUSED(A);
         GGML_UNUSED(B);
         NO_DEVICE_CODE;
-#endif // NEW_MMA_AVAILABLE
+#endif // TURING_MMA_AVAILABLE
     }
 
     static __device__ __forceinline__ void mma(
diff --git a/ggml/src/ggml-cuda/mmf.cu b/ggml/src/ggml-cuda/mmf.cu
new file mode 100644
index 00000000000..1437367e871
--- /dev/null
+++ b/ggml/src/ggml-cuda/mmf.cu
@@ -0,0 +1,431 @@
+#include "ggml.h"
+#include "common.cuh"
+#include "mma.cuh"
+#include "mmf.cuh"
+
+using namespace ggml_cuda_mma;
+
+#define MMF_ROWS_PER_BLOCK 32
+
+template <typename T, int rows_per_block, int cols_per_block, int nwarps>
+__launch_bounds__(ggml_cuda_get_physical_warp_size()*nwarps, 1)
+static __global__ void mul_mat_f(
+        const T * __restrict__ x, const float * __restrict__ y, const int32_t * __restrict__ ids, float * __restrict__ dst,
+        const int ncols, const int nchannels_y, const int stride_row, const int stride_col_y, const int stride_col_dst,
+        const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    typedef tile<16, 8, T>     tile_A;
+    typedef tile< 8, 8, T>     tile_B;
+    typedef tile<16, 8, float> tile_C;
+
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+    constexpr int tile_k_padded = warp_size + 4;
+    constexpr int ntA = rows_per_block / tile_A::I;
+    constexpr int ntB = (cols_per_block + tile_B::I - 1) / tile_B::I;
+
+    const int row0        = blockIdx.x * rows_per_block;
+    const int channel_dst = blockIdx.y;
+    const int channel_x   = channel_dst / channel_ratio;
+    const int channel_y   = channel_dst;
+    const int sample_dst  = blockIdx.z;
+    const int sample_x    = sample_dst / sample_ratio;
+    const int sample_y    = sample_dst;
+
+    x   += int64_t(sample_x)  *stride_sample_x   + channel_x  *stride_channel_x   + row0*stride_row ;
+    y   += int64_t(sample_y)  *stride_sample_y   + channel_y  *stride_channel_y;
+    dst += int64_t(sample_dst)*stride_sample_dst + channel_dst*stride_channel_dst;
+
+    const float2 * y2 = (const float2 *) y;
+
+    extern __shared__ char data_mmv[];
+
+    tile_C C[ntA][ntB];
+
+    T * tile_xy = (T *) data_mmv + threadIdx.y*(tile_A::I * tile_k_padded);
+
+    for (int col = threadIdx.y*warp_size + threadIdx.x; col < ncols; col += nwarps*warp_size) {
+        tile_A A[ntA][warp_size / tile_A::J];
+#pragma unroll
+        for (int itA = 0; itA < ntA; ++itA) {
+#pragma unroll
+            for (int i = 0; i < tile_A::I; ++i) {
+                tile_xy[i*tile_k_padded + threadIdx.x] = x[(itA*tile_A::I + i)*stride_row  + col];
+            }
+#pragma unroll
+            for (int k0 = 0; k0 < warp_size; k0 += tile_A::J) {
+                load_ldmatrix(A[itA][k0/tile_A::J], tile_xy + k0, tile_k_padded);
+            }
+        }
+
+#pragma unroll
+        for (int itB = 0; itB < ntB; ++itB) {
+            if constexpr (std::is_same_v<T, float>) {
+#pragma unroll
+                for (int j0 = 0; j0 < tile_B::I; ++j0) {
+                    const int j = j0 + itB*tile_B::I;
+
+                    tile_xy[j0*tile_k_padded + threadIdx.x] = j < cols_per_block ? y[j*stride_col_y + col] : 0.0f;
+                }
+            } else if constexpr (std::is_same_v<T, half2> || std::is_same_v<T, nv_bfloat162>) {
+#pragma unroll
+                for (int j0 = 0; j0 < tile_B::I; ++j0) {
+                    const int j = j0 + itB*tile_B::I;
+
+                    const float2 tmp = j < cols_per_block ? y2[j*stride_col_y + col] : make_float2(0.0f, 0.0f);
+                    tile_xy[j0*tile_k_padded + threadIdx.x] = {tmp.x, tmp.y};
+                }
+            } else {
+                static_assert(std::is_same_v<T, void>, "unsupported type");
+            }
+#pragma unroll
+            for (int k0 = 0; k0 < warp_size; k0 += tile_B::J) {
+                tile_B B;
+                load_ldmatrix(B, tile_xy + k0, tile_k_padded);
+#pragma unroll
+                for (int itA = 0; itA < ntA; ++itA) {
+                    mma(C[itA][itB], A[itA][k0/tile_B::J], B);
+                }
+            }
+        }
+    }
+
+    float * buf_iw = (float *) data_mmv;
+    constexpr int kiw = nwarps*rows_per_block + 4;
+
+    if (nwarps > 1) {
+        __syncthreads();
+    }
+#pragma unroll
+    for (int itB = 0; itB < ntB; ++itB) {
+#pragma unroll
+        for (int itA = 0; itA < ntA; ++itA) {
+#pragma unroll
+            for (int l = 0; l < tile_C::ne; ++l) {
+                const int i = threadIdx.y*rows_per_block + itA*tile_C::I + tile_C::get_i(l);
+                const int j = itB*tile_C::J + tile_C::get_j(l);
+                buf_iw[j*kiw + i] = C[itA][itB].x[l];
+            }
+        }
+    }
+
+    if (nwarps > 1) {
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j0 = 0; j0 < cols_per_block; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+
+        if (j0 + nwarps > cols_per_block && j >= cols_per_block) {
+            return;
+        }
+
+        float sum = 0.0f;
+        static_assert(rows_per_block == warp_size, "need loop/check");
+#pragma unroll
+        for (int i0 = 0; i0 < nwarps*rows_per_block; i0 += rows_per_block) {
+            const int i = i0 + threadIdx.x;
+
+            sum += buf_iw[j*kiw + i];
+        }
+        dst[j*stride_col_dst + row0 + threadIdx.x] = sum;
+    }
+#else
+    NO_DEVICE_CODE;
+    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(ids); GGML_UNUSED(dst);
+    GGML_UNUSED(ncols); GGML_UNUSED(nchannels_y); GGML_UNUSED(stride_row); GGML_UNUSED(stride_col_y); GGML_UNUSED(stride_col_dst);
+    GGML_UNUSED(channel_ratio); GGML_UNUSED(stride_channel_x); GGML_UNUSED(stride_channel_y); GGML_UNUSED(stride_channel_dst);
+    GGML_UNUSED(sample_ratio); GGML_UNUSED(stride_sample_x); GGML_UNUSED(stride_sample_y); GGML_UNUSED(stride_sample_dst);
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+}
+
+template <typename T, int cols_per_block>
+static void mul_mat_f_cuda(
+        const T * x, const float * y, const int32_t * ids, float * dst,
+        const int64_t ncols_x, const int64_t nrows_x,
+        const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
+        const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
+        const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
+        const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
+        cudaStream_t stream) {
+    typedef tile<16, 8, T>     tile_A;
+    typedef tile< 8, 8, T>     tile_B;
+    typedef tile<16, 8, float> tile_C;
+
+    GGML_ASSERT(!ids && "mul_mat_id not implemented");
+
+    GGML_ASSERT(ncols_x      % 2 == 0);
+    GGML_ASSERT(stride_row   % 2 == 0);
+    GGML_ASSERT(stride_col_y % 2 == 0);
+    GGML_ASSERT(ids || nchannels_dst % nchannels_x == 0);
+    GGML_ASSERT(       nsamples_dst  % nsamples_x  == 0);
+    const int64_t channel_ratio = nchannels_dst / nchannels_x;
+    const int64_t sample_ratio  = nsamples_dst  / nsamples_x;
+
+    const int device = ggml_cuda_get_device();
+    const int warp_size = ggml_cuda_info().devices[device].warp_size;
+
+    int64_t nwarps_best     = 1;
+    int64_t niter_best      = (ncols_x + warp_size*2 - 1) / (warp_size*2);
+    int64_t max_block_size  = 256;
+    for (int64_t nwarps = 2; nwarps <= max_block_size/warp_size; nwarps++) {
+        const int64_t niter = (ncols_x + nwarps*warp_size*2 - 1) / (nwarps*warp_size*2);
+        if (niter < niter_best) {
+            niter_best  = niter;
+            nwarps_best = nwarps;
+        }
+    }
+
+    constexpr int rows_per_block = MMF_ROWS_PER_BLOCK;
+    const int nbytes_shared_iter = nwarps_best * tile_A::I * (warp_size + 4) * 4;
+    const int nbytes_shared_combine = GGML_PAD(cols_per_block, tile_B::I) * (nwarps_best*rows_per_block + 4) * 4;
+    const int nbytes_shared = std::max(nbytes_shared_iter, nbytes_shared_combine);
+    const dim3 block_nums(nrows_x/rows_per_block, nchannels_dst, nsamples_dst);
+    const dim3 block_dims(warp_size, nwarps_best, 1);
+    switch (nwarps_best) {
+        case 1: {
+            mul_mat_f<T, rows_per_block, cols_per_block, 1><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, y, ids, dst, ncols_x, nchannels_y, stride_row, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+        } break;
+        case 2: {
+            mul_mat_f<T, rows_per_block, cols_per_block, 2><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, y, ids, dst, ncols_x, nchannels_y, stride_row, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+        } break;
+        case 3: {
+            mul_mat_f<T, rows_per_block, cols_per_block, 3><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, y, ids, dst, ncols_x, nchannels_y, stride_row, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+        } break;
+        case 4: {
+            mul_mat_f<T, rows_per_block, cols_per_block, 4><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, y, ids, dst, ncols_x, nchannels_y, stride_row, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+        } break;
+        case 5: {
+            mul_mat_f<T, rows_per_block, cols_per_block, 5><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, y, ids, dst, ncols_x, nchannels_y, stride_row, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+        } break;
+        case 6: {
+            mul_mat_f<T, rows_per_block, cols_per_block, 6><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, y, ids, dst, ncols_x, nchannels_y, stride_row, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+        } break;
+        case 7: {
+            mul_mat_f<T, rows_per_block, cols_per_block, 7><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, y, ids, dst, ncols_x, nchannels_y, stride_row, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+        } break;
+        case 8: {
+            mul_mat_f<T, rows_per_block, cols_per_block, 8><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, y, ids, dst, ncols_x, nchannels_y, stride_row, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+        } break;
+        default: {
+            GGML_ABORT("fatal error");
+        } break;
+    }
+}
+
+template <typename T>
+static void mul_mat_f_switch_cols_per_block(
+        const T * x, const float * y, const int32_t * ids, float * dst,
+        const int64_t ncols_x, const int64_t nrows_x, const int64_t ncols_dst,
+        const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
+        const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
+        const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
+        const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
+        cudaStream_t stream) {
+    switch (ncols_dst) {
+        case  1: {
+            mul_mat_f_cuda<T,  1>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x,                nsamples_dst,  stride_sample_x,  stride_sample_y,  stride_sample_dst,  stream);
+        } break;
+        case  2: {
+            mul_mat_f_cuda<T,  2>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x,                nsamples_dst,  stride_sample_x,  stride_sample_y,  stride_sample_dst,  stream);
+        } break;
+        case  3: {
+            mul_mat_f_cuda<T,  3>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x,                nsamples_dst,  stride_sample_x,  stride_sample_y,  stride_sample_dst,  stream);
+        } break;
+        case  4: {
+            mul_mat_f_cuda<T,  4>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x,                nsamples_dst,  stride_sample_x,  stride_sample_y,  stride_sample_dst,  stream);
+        } break;
+        case  5: {
+            mul_mat_f_cuda<T,  5>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x,                nsamples_dst,  stride_sample_x,  stride_sample_y,  stride_sample_dst,  stream);
+        } break;
+        case  6: {
+            mul_mat_f_cuda<T,  6>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x,                nsamples_dst,  stride_sample_x,  stride_sample_y,  stride_sample_dst,  stream);
+        } break;
+        case  7: {
+            mul_mat_f_cuda<T,  7>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x,                nsamples_dst,  stride_sample_x,  stride_sample_y,  stride_sample_dst,  stream);
+        } break;
+        case  8: {
+            mul_mat_f_cuda<T,  8>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x,                nsamples_dst,  stride_sample_x,  stride_sample_y,  stride_sample_dst,  stream);
+        } break;
+        case  9: {
+            mul_mat_f_cuda<T,  9>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x,                nsamples_dst,  stride_sample_x,  stride_sample_y,  stride_sample_dst,  stream);
+        } break;
+        case 10: {
+            mul_mat_f_cuda<T, 10>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x,                nsamples_dst,  stride_sample_x,  stride_sample_y,  stride_sample_dst,  stream);
+        } break;
+        case 11: {
+            mul_mat_f_cuda<T, 11>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x,                nsamples_dst,  stride_sample_x,  stride_sample_y,  stride_sample_dst,  stream);
+        } break;
+        case 12: {
+            mul_mat_f_cuda<T, 12>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x,                nsamples_dst,  stride_sample_x,  stride_sample_y,  stride_sample_dst,  stream);
+        } break;
+        case 13: {
+            mul_mat_f_cuda<T, 13>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x,                nsamples_dst,  stride_sample_x,  stride_sample_y,  stride_sample_dst,  stream);
+        } break;
+        case 14: {
+            mul_mat_f_cuda<T, 14>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x,                nsamples_dst,  stride_sample_x,  stride_sample_y,  stride_sample_dst,  stream);
+        } break;
+        case 15: {
+            mul_mat_f_cuda<T, 15>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x,                nsamples_dst,  stride_sample_x,  stride_sample_y,  stride_sample_dst,  stream);
+        } break;
+        case 16: {
+            mul_mat_f_cuda<T, 16>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x,                nsamples_dst,  stride_sample_x,  stride_sample_y,  stride_sample_dst,  stream);
+        } break;
+        default: {
+            GGML_ABORT("fatal error");
+        } break;
+    }
+}
+
+void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
+    GGML_ASSERT(        src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(!ids ||  ids->type == GGML_TYPE_I32);
+    GGML_ASSERT(         dst->type == GGML_TYPE_F32);
+
+    GGML_TENSOR_BINARY_OP_LOCALS;
+
+    const size_t ts_src0 = ggml_type_size(src0->type);
+    const size_t ts_src1 = ggml_type_size(src1->type);
+    const size_t ts_dst  = ggml_type_size(dst->type);
+
+    GGML_ASSERT(ne13 == ne3);
+
+    GGML_ASSERT(        nb00       == ts_src0);
+    GGML_ASSERT(        nb10       == ts_src1);
+    GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));
+    GGML_ASSERT(        nb0        == ts_dst);
+
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
+
+    const float   * src1_d =       (const float   *) src1->data;
+    const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
+    float         *  dst_d =       (float         *)  dst->data;
+
+    const int64_t s01 = src0->nb[1] / ts_src0;
+    const int64_t s11 = src1->nb[1] / ts_src1;
+    const int64_t s1  =  dst->nb[1] / ts_dst;
+    const int64_t s02 = src0->nb[2] / ts_src0;
+    const int64_t s12 = src1->nb[2] / ts_src1;
+    const int64_t s2  =  dst->nb[2] / ts_dst;
+    const int64_t s03 = src0->nb[3] / ts_src0;
+    const int64_t s13 = src1->nb[3] / ts_src1;
+    const int64_t s3  =  dst->nb[3] / ts_dst;
+
+    // For MUL_MAT_ID the memory layout is different than for MUL_MAT:
+    const int64_t ncols_dst          = ids ? ne2  : ne1;
+    const int64_t nchannels_y        = ids ? ne11 : ne12;
+    const int64_t nchannels_dst      = ids ? ne1  : ne2;
+    const int64_t stride_channel_dst = ids ? s1   : s2;
+    const int64_t stride_channel_y   = ids ? s11  : s12;
+
+    GGML_ASSERT(!ids || ncols_dst == 1);
+
+    switch (src0->type) {
+        case GGML_TYPE_F32: {
+            const float * src0_d = (const float *) src0->data;
+            constexpr int vals_per_T = 1;
+            mul_mat_f_switch_cols_per_block(
+                src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, s11/vals_per_T, s1,
+                ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst,
+                ne03,              ne3,           s03/vals_per_T, s13,              s3,                 ctx.stream());
+        } break;
+        case GGML_TYPE_F16: {
+            const half2 * src0_d = (const half2 *) src0->data;
+            constexpr int vals_per_T = 2;
+            mul_mat_f_switch_cols_per_block(
+                src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, s11/vals_per_T, s1,
+                ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst,
+                ne03,              ne3,           s03/vals_per_T, s13,              s3,                 ctx.stream());
+        } break;
+        case GGML_TYPE_BF16: {
+            const nv_bfloat162 * src0_d = (const nv_bfloat162 *) src0->data;
+            constexpr int vals_per_T = 2;
+            mul_mat_f_switch_cols_per_block(
+                src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, s11/vals_per_T, s1,
+                ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst,
+                ne03,              ne3,           s03/vals_per_T, s13,              s3,                 ctx.stream());
+        } break;
+        default:
+            GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));
+    }
+}
+
+bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * src0_ne, int64_t ne11) {
+    if (src0_ne[0] % (warp_size * (4/ggml_type_size(type))) != 0) {
+        return false;
+    }
+    if (src0_ne[1] % MMF_ROWS_PER_BLOCK != 0) {
+        return false;
+    }
+    if (ne11 > 16) {
+        return false;
+    }
+    switch (type) {
+        case GGML_TYPE_F32:
+            return ampere_mma_available(cc);
+        case GGML_TYPE_F16:
+            return turing_mma_available(cc);
+        case GGML_TYPE_BF16:
+            return ampere_mma_available(cc);
+        default:
+            return false;
+    }
+}
diff --git a/ggml/src/ggml-cuda/mmf.cuh b/ggml/src/ggml-cuda/mmf.cuh
new file mode 100644
index 00000000000..785f9f211c3
--- /dev/null
+++ b/ggml/src/ggml-cuda/mmf.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
+
+bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * scr0_ne, int64_t ne11);
diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cu
index 8954a383104..384ee7615f7 100644
--- a/ggml/src/ggml-cuda/mmq.cu
+++ b/ggml/src/ggml-cuda/mmq.cu
@@ -310,7 +310,7 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
         return false;
     }
 
-    if (new_mma_available(cc)) {
+    if (turing_mma_available(cc)) {
         return true;
     }
 
diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh
index 1634725c20a..96129bd831f 100644
--- a/ggml/src/ggml-cuda/mmq.cuh
+++ b/ggml/src/ggml-cuda/mmq.cuh
@@ -92,7 +92,7 @@ struct tile_x_sizes {
 };
 
 static int get_mmq_x_max_host(const int cc) {
-    return (amd_mfma_available(cc) || new_mma_available(cc)) ? 128 :
+    return (amd_mfma_available(cc) || turing_mma_available(cc)) ? 128 :
         GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA ?
 #ifdef GGML_CUDA_FORCE_MMQ
             128                     : 64;
@@ -102,9 +102,9 @@ static int get_mmq_x_max_host(const int cc) {
 }
 
 static constexpr __device__ int get_mmq_x_max_device() {
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     return 128;
-#else // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#else // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
 #if defined(GGML_USE_HIP)
     return 64;
@@ -121,7 +121,7 @@ static constexpr __device__ int get_mmq_x_max_device() {
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
 
 #endif // defined(GGML_USE_HIP)
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 }
 
 static int get_mmq_y_host(const int cc) {
@@ -233,7 +233,7 @@ static constexpr __host__ __device__ int mmq_get_mma_tile_x_k(ggml_type type) {
 static int mmq_get_granularity_host(const int mmq_x, const int cc) {
     if (amd_mfma_available(cc)) {
         return mmq_x >= 128 ? 32 : 16;
-    } else if (new_mma_available(cc) && mmq_x >= 48) {
+    } else if (turing_mma_available(cc) && mmq_x >= 48) {
         return 16;
     } else {
         return 8;
@@ -244,7 +244,7 @@ static int mmq_get_granularity_host(const int mmq_x, const int cc) {
 static constexpr __device__ int mmq_get_granularity_device(const int mmq_x) {
     return mmq_x >= 128 ? 32 : 16;
 }
-#elif defined(NEW_MMA_AVAILABLE)
+#elif defined(TURING_MMA_AVAILABLE)
 static constexpr __device__ int mmq_get_granularity_device(const int mmq_x) {
     return mmq_x >= 48 ? 16 : 8;
 }
@@ -279,14 +279,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + 2*MMQ_TILE_NE_K);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_0, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_0);
     constexpr int nrows = warp_size / threads_per_row;
@@ -305,12 +305,12 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         const block_q4_0 * bxi = (const block_q4_0 *) x + kbx0 + i*stride + kbx;
         const int qs0 = get_int_b2(bxi->qs, kqsx);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI4_0) + kqsx + 0]     = __vsubss4((qs0 >> 0) & 0x0F0F0F0F, 0x08080808);
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI4_0) + kqsx + QI4_0] = __vsubss4((qs0 >> 4) & 0x0F0F0F0F, 0x08080808);
 #else
         x_qs[i*(MMQ_TILE_NE_K + 1) + txi] = qs0;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 
     constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI4_0;
@@ -327,11 +327,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         const block_q4_0 * bxi = (const block_q4_0 *) x + kbx0 + i*stride + kbxd;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q8_0           + kbxd] = bxi->d;
 #else
         x_df[i*(MMQ_TILE_NE_K/QI4_0) + i/QI4_0 + kbxd] = bxi->d;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 }
 
@@ -382,14 +382,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_1, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     half2 * x_dm = (half2 *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_1);
     constexpr int nrows = warp_size / threads_per_row;
@@ -408,12 +408,12 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         const block_q4_1 * bxi = (const block_q4_1 *) x + kbx0 + i*stride + kbx;
         const int qs0 = get_int_b4(bxi->qs, kqsx);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI4_1) + kqsx + 0]     = (qs0 >> 0) & 0x0F0F0F0F;
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI4_1) + kqsx + QI4_1] = (qs0 >> 4) & 0x0F0F0F0F;
 #else
         x_qs[i*(MMQ_TILE_NE_K + 1) + txi] = qs0;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 
     constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI4_1;
@@ -430,11 +430,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         const block_q4_1 * bxi = (const block_q4_1 *) x + kbx0 + i*stride + kbxd;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_dm[i*MMQ_MMA_TILE_X_K_Q8_1           + kbxd] = bxi->dm;
 #else
         x_dm[i*(MMQ_TILE_NE_K/QI4_1) + i/QI4_1 + kbxd] = bxi->dm;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 }
 
@@ -485,14 +485,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_0, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR5_0);
     constexpr int nrows = warp_size / threads_per_row;
@@ -527,13 +527,13 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         qs1    |= (qh <<  9)   & 0x10000000;  // 19 -> 28
         qs1     = __vsubss4(qs1, 0x10101010); // subtract 16
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI5_0) + kqsx + 0]     = qs0;
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI5_0) + kqsx + QI5_0] = qs1;
 #else
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_0) + kqsx + 0]     = qs0;
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_0) + kqsx + QI5_0] = qs1;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 
     constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI5_0;
@@ -550,11 +550,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         const block_q5_0 * bxi = (const block_q5_0 *) x + kbx0 + i*stride + kbxd;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q8_0           + kbxd] = bxi->d;
 #else
         x_df[i*(MMQ_TILE_NE_K/QI5_0) + i/QI5_0 + kbxd] = bxi->d;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 }
 
@@ -563,14 +563,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_1, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     half2 * x_dm = (half2 *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR5_1);
     constexpr int nrows = warp_size / threads_per_row;
@@ -603,13 +603,13 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         qs1    |= (qh <<  2) & 0x00100000; // 18 -> 20
         qs1    |= (qh <<  9) & 0x10000000; // 19 -> 28
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI5_1) + kqsx + 0]     = qs0;
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI5_1) + kqsx + QI5_1] = qs1;
 #else
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_1) + kqsx + 0]     = qs0;
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_1) + kqsx + QI5_1] = qs1;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 
     constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI5_1;
@@ -626,11 +626,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         const block_q5_1 * bxi = (const block_q5_1 *) x + kbx0 + i*stride + kbxd;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_dm[i*MMQ_MMA_TILE_X_K_Q8_1           + kbxd] = bxi->dm;
 #else
         x_dm[i*(MMQ_TILE_NE_K/QI5_1) + i/QI5_1 + kbxd] = bxi->dm;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 }
 
@@ -639,14 +639,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_tile + 2*MMQ_TILE_NE_K);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
     // MMQ_ITER_K / (4 * QR8_0) == 64 required. but NV has only 32 threads per warp
     constexpr int threads_per_row = 32;
@@ -665,13 +665,13 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbx;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 0             + txi] = get_int_b2(bxi[0].qs,                   kqsx);
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + MMQ_TILE_NE_K + txi] = get_int_b2(bxi[MMQ_TILE_NE_K/QI8_0].qs, kqsx);
 #else
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + 0             + txi] = get_int_b2(bxi[0].qs,                   kqsx);
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + MMQ_TILE_NE_K + txi] = get_int_b2(bxi[MMQ_TILE_NE_K/QI8_0].qs, kqsx);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 
     constexpr int blocks_per_tile_x_row = 2*MMQ_TILE_NE_K / QI8_0;
@@ -688,11 +688,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbxd;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q8_0                 + kbxd] = bxi->d;
 #else
         x_df[i*(2*MMQ_TILE_NE_K/QI8_0) + i/(QI8_0/2) + kbxd] = bxi->d;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 }
 
@@ -701,14 +701,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_MXFP4, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR_MXFP4);
     constexpr int nrows = warp_size / threads_per_row;
@@ -730,13 +730,13 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         const int2 v = get_int_from_table_16(aux_q4, kvalues_mxfp4);
         const int k0 = kbx * (2 * QI_MXFP4) + kqsx;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + k0 + 0]        = v.x;
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + k0 + QI_MXFP4] = v.y;
 #else
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0]        = v.x;
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + QI_MXFP4] = v.y;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 
     constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI_MXFP4;
@@ -753,11 +753,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         const block_mxfp4 * bxi = (const block_mxfp4 *) x + kbx0 + i*stride + kbxd;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q8_1                 + kbxd] = ggml_cuda_e8m0_to_fp32(bxi->e)*0.5f;
 #else
         x_df[i*(MMQ_TILE_NE_K/QI_MXFP4) + i/QI_MXFP4 + kbxd] = ggml_cuda_e8m0_to_fp32(bxi->e)*0.5f;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 }
 
@@ -1178,7 +1178,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma(
             }
         }
     }
-#elif defined(NEW_MMA_AVAILABLE)
+#elif defined(TURING_MMA_AVAILABLE)
 
     typedef tile<16, 4, int> tile_A;
     typedef tile<16, 8, int> tile_A_8;
@@ -1264,14 +1264,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
     constexpr int nwarps = mmq_get_nwarps_device();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q2_K, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     half2 * x_dm = (half2 *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR2_K);
     constexpr int nrows = ggml_cuda_get_physical_warp_size() / threads_per_row;
@@ -1295,11 +1295,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
             const int x_qs_k = (x_ql_0 >> (2*l)) & 0x03030303;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q2_K + k] = x_qs_k;
 #else
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + k] = x_qs_k;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         }
 
         const int sc_m = bxi->scales[kqsx];
@@ -1310,11 +1310,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         const half2 x_dm_ik = make_half2(bxi_dmf.x*(sc_m & 0x0F), bxi_dmf.y*(sc_m >> 4));
 #endif // FAST_FP16_AVAILABLE
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_dm[i*MMQ_MMA_TILE_X_K_Q2_K + kqsx] = x_dm_ik;
 #else
         x_dm[i*(MMQ_TILE_NE_K + 1)   + kqsx] = x_dm_ik;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 }
 
@@ -1452,7 +1452,7 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
             }
         }
     }
-#elif defined(NEW_MMA_AVAILABLE)
+#elif defined(TURING_MMA_AVAILABLE)
 
     typedef tile<16, 4, int> tile_A;
     typedef tile<16, 8, int> tile_A_8;
@@ -1582,7 +1582,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
 #else
@@ -1590,7 +1590,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
     int   * x_sc = (int   *) (x_df + txs.dm);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR3_K);
     constexpr int nrows = warp_size / threads_per_row;
@@ -1618,11 +1618,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
             const int x_qs_k = __vsubss4(x_ql_k | x_qh_k, 0x04040404);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + k] = x_qs_k;
 #else
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + k] = x_qs_k;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         }
     }
 
@@ -1649,7 +1649,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         const int sc = __vsubss4(sc_low | sc_high, 0x20202020);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         const int8_t * sc8 = (const int8_t *) &sc;
         const float d = bxi->d;
 
@@ -1659,10 +1659,10 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         }
 #else
         x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = sc;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 
-#if !(defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE))
+#if !(defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE))
 #pragma unroll
     for (int i0 = 0; i0 < mmq_y; i0 += nwarps*warp_size) {
         int i = (i0 + threadIdx.y*warp_size + threadIdx.x) % mmq_y;
@@ -1675,7 +1675,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         x_df[i] = bxi->d;
     }
-#endif // !(defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE))
+#endif // !(defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE))
 }
 
 template <int mmq_x, int mmq_y>
@@ -1728,7 +1728,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K);
 #else
@@ -1736,7 +1736,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     int   * x_qs = (int   *)  x_tile;
     half2 * x_dm = (half2 *) (x_qs + txs.qs);
     int   * x_sc = (int   *) (x_dm + txs.dm);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_K);
     constexpr int nrows = warp_size / threads_per_row;
@@ -1753,15 +1753,15 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride;
         const int qs0 = get_int_b4(bxi->qs, txi);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 16*(txi/8) + txi % 8 + 0] = (qs0 >> 0) & 0x0F0F0F0F;
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 16*(txi/8) + txi % 8 + 8] = (qs0 >> 4) & 0x0F0F0F0F;
 #else
         x_qs[i*(MMQ_TILE_NE_K + 1) + txi] = qs0;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     constexpr int rows_per_warp = warp_size / 2;
 #pragma unroll
     for (int i0 = 0; i0 < mmq_y; i0 += nwarps*rows_per_warp) {
@@ -1829,7 +1829,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = scales8;
     }
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 }
 
 template <int mmq_x, int mmq_y>
@@ -1872,7 +1872,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     half2 * x_dm = (half2 *) (x_qs + MMQ_TILE_NE_K*2);
 #else
@@ -1880,7 +1880,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     int   * x_qs = (int   *)  x_tile;
     half2 * x_dm = (half2 *) (x_qs + txs.qs);
     int   * x_sc = (int   *) (x_dm + txs.dm);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR5_K);
     constexpr int nrows = warp_size / threads_per_row;
@@ -1908,16 +1908,16 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         const int kq0 = ky - ky % (QI5_K/2) + txi % (QI5_K/4) + 0;
         const int kq1 = ky - ky % (QI5_K/2) + txi % (QI5_K/4) + QI5_K/4;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kq0] = ql0 | qh0;
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kq1] = ql1 | qh1;
 #else
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq0] = ql0 | qh0;
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq1] = ql1 | qh1;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     constexpr int rows_per_warp = warp_size / 2;
 #pragma unroll
     for (int i0 = 0; i0 < mmq_y; i0 += nwarps*rows_per_warp) {
@@ -1986,7 +1986,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = scales8;
     }
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 }
 
 template <int mmq_x, int mmq_y>
@@ -2029,7 +2029,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
     int   * x_sc = (int   *) (x_df + MMQ_TILE_NE_K/QI6_K);
@@ -2038,7 +2038,7 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
     int   * x_sc = (int   *) (x_df + txs.dm);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR6_K);
     constexpr int nrows = warp_size / threads_per_row;
@@ -2065,13 +2065,13 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         const int kq0 = 2*txi - txi % (QI6_K/2) + 0;
         const int kq1 = 2*txi - txi % (QI6_K/2) + QI6_K/2;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_qs[i*MMQ_MMA_TILE_X_K_Q6_K + kq0] = __vsubss4(ql0 | qh0, 0x20202020);
         x_qs[i*MMQ_MMA_TILE_X_K_Q6_K + kq1] = __vsubss4(ql1 | qh1, 0x20202020);
 #else
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq0] = __vsubss4(ql0 | qh0, 0x20202020);
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq1] = __vsubss4(ql1 | qh1, 0x20202020);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 
 #pragma unroll
@@ -2084,11 +2084,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q6_K]           = bxi->d;
 #else
         x_df[i*(MMQ_TILE_NE_K/QI6_K) + i/QI6_K] = bxi->d;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 
     constexpr int rows_per_warp = warp_size / 4;
@@ -2102,11 +2102,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride + (threadIdx.x % (MMQ_TILE_NE_K/8)) / 4;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_sc[i*MMQ_MMA_TILE_X_K_Q6_K + threadIdx.x%4] = get_int_b2(bxi->scales, threadIdx.x % (MMQ_TILE_NE_K/8));
 #else
         x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + threadIdx.x%(MMQ_TILE_NE_K/8)] = get_int_b2(bxi->scales, threadIdx.x%(QI6_K/8));
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 }
 
@@ -2199,7 +2199,7 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma(
             }
         }
     }
-#elif defined(NEW_MMA_AVAILABLE)
+#elif defined(TURING_MMA_AVAILABLE)
 
     typedef tile<16, 4, int> tile_A;
     typedef tile< 8, 4, int> tile_B;
@@ -2311,14 +2311,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ4_NL, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_NL);
     constexpr int nrows = warp_size / threads_per_row;
@@ -2340,13 +2340,13 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         const int2 v = get_int_from_table_16(aux_q4, kvalues_iq4nl);
         const int k0 = kbx * (2 * QI4_NL) + kqsx;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + 0]      = v.x;
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + QI4_NL] = v.y;
 #else
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0]      = v.x;
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + QI4_NL] = v.y;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 
     constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI4_NL;
@@ -2363,11 +2363,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
         const block_iq4_nl * bxi = (const block_iq4_nl *) x + kbx0 + i*stride + kbxd;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q8_0             + kbxd] = __half2float(bxi->d);
 #else
         x_df[i*(MMQ_TILE_NE_K/QI4_NL) + i/QI4_NL + kbxd] = __half2float(bxi->d);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 }
 
@@ -2376,14 +2376,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ2_XXS, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
     constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_XXS)) / 2;
     constexpr int nrows = warp_size / threads_per_row;
@@ -2414,22 +2414,22 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
             const int signs1 = __vcmpne4(((signs_packed & 0x30) << 3) | ((signs_packed & 0xC0) << 17), 0x00000000);
             const int grid1 = __vsub4(grid_pos[1] ^ signs1, signs1);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 0)] = grid0;
             x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 1)] = grid1;
 #else
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid0;
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid1;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         }
 
         const int ls = aux32 >> 28;
         const float d = bxi->d;
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q8_0   + kqsx] = (ls*d + d/2)/4;
 #else
         x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = (ls*d + d/2)/4;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 }
 
@@ -2438,14 +2438,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
 #else
     constexpr tile_x_sizes txs = MMQ_DP4A_TXS_Q8_0_16;
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
     constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_XS)) / 2;
     constexpr int nrows = warp_size / threads_per_row;
@@ -2472,24 +2472,24 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
             const int grid_l = __vsub4(grid_pos[0] ^ signs[0], signs[0]);
             const int grid_h = __vsub4(grid_pos[1] ^ signs[1], signs[1]);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 0)] = grid_l;
             x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 1)] = grid_h;
 #else
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l;
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         }
 
         const int ls = bxi->scales[kqsx];
         const float d = bxi->d;
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q3_K                   + 2*kqsx+0] = ((ls &  0x0F)*d + d/2)/4;
         x_df[i*MMQ_MMA_TILE_X_K_Q3_K                   + 2*kqsx+1] = ((ls >>    4)*d + d/2)/4;
 #else
         x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+0] = ((ls &  0x0F)*d + d/2)/4;
         x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+1] = ((ls >>    4)*d + d/2)/4;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 }
 
@@ -2498,14 +2498,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ2_S, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
     constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_S)) / 2;
     constexpr int nrows = warp_size / threads_per_row;
@@ -2539,24 +2539,24 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
             const int grid_l = __vsub4(grid_pos[0] ^ signs0, signs0);
             const int grid_h = __vsub4(grid_pos[1] ^ signs1, signs1);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 0)] = grid_l;
             x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 1)] = grid_h;
 #else
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l;
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         }
 
         const int ls = bxi->scales[kqsx];
         const float d = bxi->d;
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q3_K                   + 2*kqsx+0] = ((ls &  0x0F)*d + d/2)/4;
         x_df[i*MMQ_MMA_TILE_X_K_Q3_K                   + 2*kqsx+1] = ((ls >>    4)*d + d/2)/4;
 #else
         x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+0] = ((ls &  0x0F)*d + d/2)/4;
         x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+1] = ((ls >>    4)*d + d/2)/4;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 }
 
@@ -2565,14 +2565,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_XXS, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
     constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR3_XXS)) / 2;
     constexpr int nrows = warp_size / threads_per_row;
@@ -2601,22 +2601,22 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
             const int grid_l = __vsub4(grid_pos.x ^ signs[0], signs[0]);
             const int grid_h = __vsub4(grid_pos.y ^ signs[1], signs[1]);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 0)] = grid_l;
             x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 1)] = grid_h;
 #else
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l;
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         }
 
         const int ls = aux32 >> 28;
         const float d = bxi->d;
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q8_0     + kqsx] = (ls*d + d/2)/2;
 #else
         x_df[i*(MMQ_TILE_NE_K/4) + i/4   + kqsx] = (ls*d + d/2)/2;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 }
 
@@ -2625,14 +2625,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_S, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
     constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR3_S)) / 2;
     constexpr int nrows = warp_size / threads_per_row;
@@ -2668,22 +2668,22 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
             const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0);
             const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l+0)] = grid_l;
             x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l+1)] = grid_h;
 #else
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+0)] = grid_l;
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+1)] = grid_h;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         }
 
         const int ls = 1 + 2*((bxi->scales[kqsx/2] >> (((2*kqsx) << 1) & 0x04)) & 0x0F);
         const float d = bxi->d;
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q8_0     + kqsx] = ls*d;
 #else
         x_df[i*(MMQ_TILE_NE_K/4) + i/4   + kqsx] = ls*d;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 }
 
@@ -2692,14 +2692,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     half2 * x_ds = (half2 *) (x_qs + MMQ_TILE_NE_K*2);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_S, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     half2 * x_ds = (half2 *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR1_S);
     constexpr int nrows = warp_size / threads_per_row;
@@ -2727,23 +2727,23 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
             const int grid0 = (grid >> 0) & 0x0F0F0F0F;
             const int grid1 = (grid >> 4) & 0x0F0F0F0F;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 8*kqsx + (2*l+0)] = grid0;
             x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 8*kqsx + (2*l+1)] = grid1;
 #else
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+0)] = grid0;
             x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+1)] = grid1;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         }
 
         const float  d1q   = __half2float(bxi->d) * (((qh >> 11) & 0x0E) + 1);
         const float  delta = -1.0f + IQ1S_DELTA - (qh & 0x8000) * (2.0f*IQ1S_DELTA/0x8000);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_ds[i*MMQ_MMA_TILE_X_K_Q8_1     + kqsx] = make_half2(d1q, d1q*delta);
 #else
         x_ds[i*(MMQ_TILE_NE_K/4) + i/4   + kqsx] = make_half2(d1q, d1q*delta);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 }
 
@@ -2752,14 +2752,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     constexpr int nwarps = mmq_get_nwarps_device();
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
 #else
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ4_XS, mmq_y);
     int   * x_qs = (int   *)  x_tile;
     float * x_df = (float *) (x_qs + txs.qs);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
     constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_XS);
     constexpr int nrows = warp_size / threads_per_row;
@@ -2779,13 +2779,13 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         const int2 v = get_int_from_table_16(aux_q4, kvalues_iq4nl);
         const int k0 = 8 * (kqsx / 4) + kqsx % 4;
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + 0] = v.x;
         x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + 4] = v.y;
 #else
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0] = v.x;
         x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 4] = v.y;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 
     constexpr int rows_per_warp = warp_size / 8;
@@ -2804,11 +2804,11 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
         const int ls = ((bxi->scales_l[(threadIdx.x % 8)/2] >> (4*(threadIdx.x % 2))) & 0x0F)
             | (((bxi->scales_h >> (2*(threadIdx.x % 8))) & 0x03) << 4);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
         x_df[i*MMQ_MMA_TILE_X_K_Q8_0   + threadIdx.x % 8] = d * (ls - 32);
 #else
         x_df[i*(MMQ_TILE_NE_K/4) + i/4 + threadIdx.x % 8] = d * (ls - 32);
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     }
 }
 
@@ -2859,9 +2859,9 @@ static __device__ __forceinline__ void mmq_write_back_mma(
     constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
 
     const int i0 = (threadIdx.y / ntx) * (ntx*tile_C::I);
-#if defined(NEW_MMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
+#if defined(TURING_MMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
     static_assert(nwarps*tile_C::I == mmq_y, "nwarps*tile_C::I != mmq_y");
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
 #pragma unroll
     for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
@@ -3061,13 +3061,13 @@ static __device__ __forceinline__ void mul_mat_q_process_tile(
     int * tile_y = data_mul_mat_q + mmq_x;
     int * tile_x = tile_y + GGML_PAD(mmq_x*MMQ_TILE_Y_K, nwarps*warp_size);
 
-#if defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
     constexpr vec_dot_mmq_t    vec_dot    = mmq_type_traits<mmq_x, mmq_y, need_check, type>::vec_dot_mma;
     constexpr mmq_write_back_t write_back = mmq_write_back_mma<type, mmq_x, mmq_y, need_check>;
 #else
     constexpr vec_dot_mmq_t    vec_dot    = mmq_type_traits<mmq_x, mmq_y, need_check, type>::vec_dot_dp4a;
     constexpr mmq_write_back_t write_back = mmq_write_back_dp4a<mmq_x, mmq_y, need_check>;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(NEW_MMA_AVAILABLE)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
     constexpr int blocks_per_iter = MMQ_ITER_K / qk;
 
@@ -3534,7 +3534,7 @@ static size_t mmq_get_nbytes_shared(const int mmq_x, const int mmq_y, const int
     const tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(type, mmq_y);
     const int mmq_tile_x_k = mmq_get_mma_tile_x_k(type);
     const size_t nbs_ids = mmq_x*sizeof(int);
-    const size_t nbs_x = (new_mma_available(cc) || amd_mfma_available(cc)) ? mmq_y*mmq_tile_x_k*sizeof(int) : txs.qs*sizeof(int) + txs.dm*sizeof(half2) + txs.sc*sizeof(int);
+    const size_t nbs_x = (turing_mma_available(cc) || amd_mfma_available(cc)) ? mmq_y*mmq_tile_x_k*sizeof(int) : txs.qs*sizeof(int) + txs.dm*sizeof(half2) + txs.sc*sizeof(int);
     const size_t nbs_y = mmq_x*sizeof(block_q8_1_mmq);
     return nbs_ids + nbs_x + GGML_PAD(nbs_y, nwarps*warp_size*sizeof(int));
 }
diff --git a/ggml/src/ggml-cuda/mmvf.cu b/ggml/src/ggml-cuda/mmvf.cu
new file mode 100644
index 00000000000..1ad4bc75ba6
--- /dev/null
+++ b/ggml/src/ggml-cuda/mmvf.cu
@@ -0,0 +1,510 @@
+#include "ggml.h"
+#include "common.cuh"
+#include "mmvf.cuh"
+
+template <typename T, typename type_acc, int ncols_dst, int block_size>
+static __global__ void mul_mat_vec_f(
+        const T * __restrict__ x, const float * __restrict__ y, const int32_t * __restrict__ ids, float * __restrict__ dst,
+        const int ncols2, const int nchannels_y, const int stride_row, const int stride_col_y2, const int stride_col_dst,
+        const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
+    const int row         = blockIdx.x;
+    const int channel_dst = blockIdx.y;
+    const int channel_x   = ids ? ids[channel_dst]          : channel_dst / channel_ratio;
+    const int channel_y   = ids ? channel_dst % nchannels_y : channel_dst;
+    const int sample_dst  = blockIdx.z;
+    const int sample_x    = sample_dst / sample_ratio;
+    const int sample_y    = sample_dst;
+    const int tid         = threadIdx.x;
+
+    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
+
+    x   += int64_t(sample_x)  *stride_sample_x   + channel_x  *stride_channel_x   + row*stride_row;
+    y   += int64_t(sample_y)  *stride_sample_y   + channel_y  *stride_channel_y;
+    dst += int64_t(sample_dst)*stride_sample_dst + channel_dst*stride_channel_dst;
+
+    const float2 * y2 = (const float2 *) y;
+
+    extern __shared__ char data_mmv[];
+    float * buf_iw = (float *) data_mmv;
+
+    if (block_size > warp_size) {
+        if (tid < warp_size) {
+            buf_iw[tid] = 0.0f;
+        }
+        __syncthreads();
+    }
+
+    float sumf[ncols_dst] = {0.0f};
+
+    if constexpr (std::is_same_v<T, float>) {
+        const float2 * x2 = (const float2 *) x;
+
+        for (int col2 = tid; col2 < ncols2; col2 += block_size) {
+            const float2 tmpx = x2[col2];
+
+#pragma unroll
+            for (int j = 0; j < ncols_dst; ++j) {
+                const float2 tmpy = y2[j*stride_col_y2 + col2];
+                sumf[j] += tmpx.x*tmpy.x;
+                sumf[j] += tmpx.y*tmpy.y;
+            }
+        }
+    } else if constexpr (std::is_same_v<T, half>) {
+        const half2 * x2 = (const half2 *) x;
+
+        if (std::is_same_v<type_acc, float>) {
+            for (int col2 = tid; col2 < ncols2; col2 += block_size) {
+                const float2 tmpx = __half22float2(x2[col2]);
+
+#pragma unroll
+                for (int j = 0; j < ncols_dst; ++j) {
+                    const float2 tmpy = y2[j*stride_col_y2 + col2];
+                    sumf[j] += tmpx.x * tmpy.x;
+                    sumf[j] += tmpx.y * tmpy.y;
+                }
+            }
+        } else {
+#ifdef FP16_AVAILABLE
+            half2 sumh2[ncols_dst] = {{0.0f, 0.0f}};
+
+            for (int col2 = tid; col2 < ncols2; col2 += block_size) {
+                const half2 tmpx = x2[col2];
+
+#pragma unroll
+                for (int j = 0; j < ncols_dst; ++j) {
+                    const float2 tmpy = y2[j*stride_col_y2 + col2];
+                    sumh2[j] += tmpx * make_half2(tmpy.x, tmpy.y);
+                }
+            }
+
+#pragma unroll
+            for (int j = 0; j < ncols_dst; ++j) {
+                sumf[j] = __low2float(sumh2[j]) + __high2float(sumh2[j]);
+            }
+#else
+            NO_DEVICE_CODE;
+#endif // FP16_AVAILABLE
+        }
+    } else if constexpr (std::is_same_v<T, nv_bfloat16>) {
+        const int * x2 = (const int *) x;
+        for (int col2 = tid; col2 < ncols2; col2 += block_size) {
+            const int tmpx = x2[col2];
+#pragma unroll
+            for (int j = 0; j < ncols_dst; ++j) {
+                const float2 tmpy = y2[j*stride_col_y2 + col2];
+                sumf[j] += float(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[0]) * tmpy.x;
+                sumf[j] += float(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[1]) * tmpy.y;
+            }
+        }
+    } else {
+        static_assert(std::is_same_v<T, void>, "unsupported type");
+    }
+
+#pragma unroll
+    for (int j = 0; j < ncols_dst; ++j) {
+        sumf[j] = warp_reduce_sum<warp_size>(sumf[j]);
+
+        if (block_size > warp_size) {
+            buf_iw[tid/warp_size] = sumf[j];
+            __syncthreads();
+            if (tid < warp_size) {
+                sumf[j] = buf_iw[tid];
+                sumf[j] = warp_reduce_sum<warp_size>(sumf[j]);
+            }
+            if (j < ncols_dst) {
+                __syncthreads();
+            }
+        }
+    }
+
+    if (tid >= ncols_dst) {
+        return;
+    }
+
+    dst[tid*stride_col_dst + row] = sumf[tid];
+}
+
+template <typename T, typename type_acc, int ncols_dst>
+static void launch_mul_mat_vec_f_cuda(
+        const T * x, const float * y, const int32_t * ids, float * dst,
+        const int64_t ncols, const int64_t nrows,
+        const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
+        const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
+        const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
+        const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
+        cudaStream_t stream) {
+    GGML_ASSERT(ncols        % 2 == 0);
+    GGML_ASSERT(stride_row   % 2 == 0);
+    GGML_ASSERT(stride_col_y % 2 == 0);
+    GGML_ASSERT(ids || nchannels_dst % nchannels_x == 0);
+    GGML_ASSERT(       nsamples_dst  % nsamples_x  == 0);
+    const int64_t channel_ratio = nchannels_dst / nchannels_x;
+    const int64_t sample_ratio  = nsamples_dst  / nsamples_x;
+
+    const int device = ggml_cuda_get_device();
+    const int warp_size = ggml_cuda_info().devices[device].warp_size;
+
+    int64_t block_size_best = warp_size;
+    int64_t niter_best      = (ncols + 2*warp_size - 1) / (2*warp_size);
+    int64_t max_block_size  = 256;
+    if(ggml_cuda_info().devices[device].cc > GGML_CUDA_CC_OFFSET_AMD && ggml_cuda_info().devices[device].cc < GGML_CUDA_CC_RDNA1) {
+        max_block_size = 128;
+    }
+    for (int64_t block_size = 2*warp_size; block_size <= max_block_size; block_size += warp_size) {
+        const int64_t niter = (ncols + 2*block_size - 1) / (2*block_size);
+        if (niter < niter_best) {
+            niter_best      = niter;
+            block_size_best = block_size;
+        }
+    }
+
+    const int nbytes_shared = warp_size*sizeof(float);
+    const dim3 block_nums(nrows, nchannels_dst, nsamples_dst);
+    const dim3 block_dims(block_size_best, 1, 1);
+    switch (block_size_best) {
+        case   32: {
+            mul_mat_vec_f<T, type_acc, ncols_dst,  32><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+        } break;
+        case   64: {
+            mul_mat_vec_f<T, type_acc, ncols_dst,  64><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+        } break;
+        case   96: {
+            mul_mat_vec_f<T, type_acc, ncols_dst,  96><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+        } break;
+        case  128: {
+            mul_mat_vec_f<T, type_acc, ncols_dst, 128><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+        } break;
+        case  160: {
+            mul_mat_vec_f<T, type_acc, ncols_dst, 160><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+        } break;
+        case  192: {
+            mul_mat_vec_f<T, type_acc, ncols_dst, 192><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+        } break;
+        case  224: {
+            mul_mat_vec_f<T, type_acc, ncols_dst, 224><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+        } break;
+        case  256: {
+            mul_mat_vec_f<T, type_acc, ncols_dst, 256><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+        } break;
+        default: {
+            GGML_ABORT("fatal error");
+        } break;
+    }
+}
+
+template <typename T, typename type_acc>
+static void mul_mat_vec_f_cuda_switch_ncols_dst(
+        const T * x, const float * y, const int32_t * ids, float * dst,
+        const int64_t ncols, const int64_t nrows, const int64_t ncols_dst,
+        const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
+        const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
+        const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
+        const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
+        cudaStream_t stream) {
+    switch (ncols_dst) {
+        case 1:
+            launch_mul_mat_vec_f_cuda<T, type_acc, 1>
+                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+            break;
+        case 2:
+            launch_mul_mat_vec_f_cuda<T, type_acc, 2>
+                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+            break;
+        case 3:
+            launch_mul_mat_vec_f_cuda<T, type_acc, 3>
+                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+            break;
+        case 4:
+            launch_mul_mat_vec_f_cuda<T, type_acc, 4>
+                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+            break;
+        case 5:
+            launch_mul_mat_vec_f_cuda<T, type_acc, 5>
+                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+            break;
+        case 6:
+            launch_mul_mat_vec_f_cuda<T, type_acc, 6>
+                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+            break;
+        case 7:
+            launch_mul_mat_vec_f_cuda<T, type_acc, 7>
+                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+            break;
+        case 8:
+            launch_mul_mat_vec_f_cuda<T, type_acc, 8>
+                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
+}
+
+template<typename T>
+static void mul_mat_vec_f_cuda(
+        const T * x, const float * y, const int32_t * ids, float * dst,
+        const int64_t ncols, const int64_t nrows, const int64_t ncols_dst,
+        const int64_t stride_row, const int64_t stride_col_y, const int stride_col_dst,
+        const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
+        const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
+        const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
+        enum ggml_prec prec, cudaStream_t stream) {
+    if constexpr(std::is_same_v<T, half>) {
+        if (prec == GGML_PREC_DEFAULT) {
+            mul_mat_vec_f_cuda_switch_ncols_dst<T, half>
+                (x, y, ids, dst, ncols, nrows, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+            return;
+        }
+    }
+    mul_mat_vec_f_cuda_switch_ncols_dst<T, float>
+        (x, y, ids, dst, ncols, nrows, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+         nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+         stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+}
+
+void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
+    GGML_ASSERT(        src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(!ids ||  ids->type == GGML_TYPE_I32);
+    GGML_ASSERT(         dst->type == GGML_TYPE_F32);
+
+    GGML_TENSOR_BINARY_OP_LOCALS;
+
+    const size_t ts_src0 = ggml_type_size(src0->type);
+    const size_t ts_src1 = ggml_type_size(src1->type);
+    const size_t ts_dst  = ggml_type_size(dst->type);
+
+    GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for  batch size 1.
+    GGML_ASSERT(ne13 == ne3);
+
+    GGML_ASSERT(        nb00       == ts_src0);
+    GGML_ASSERT(        nb10       == ts_src1);
+    GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));
+    GGML_ASSERT(        nb0        == ts_dst);
+
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
+
+    const float   * src1_d =       (const float   *) src1->data;
+    const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
+    float         *  dst_d =       (float         *)  dst->data;
+
+    const int64_t s01 = src0->nb[1] / ts_src0;
+    const int64_t s11 = src1->nb[1] / ts_src1;
+    const int64_t s1  =  dst->nb[1] / ts_dst;
+    const int64_t s02 = src0->nb[2] / ts_src0;
+    const int64_t s12 = src1->nb[2] / ts_src1;
+    const int64_t s2  =  dst->nb[2] / ts_dst;
+    const int64_t s03 = src0->nb[3] / ts_src0;
+    const int64_t s13 = src1->nb[3] / ts_src1;
+    const int64_t s3  =  dst->nb[3] / ts_dst;
+
+    // For MUL_MAT_ID the memory layout is different than for MUL_MAT:
+    const int64_t ncols_dst          = ids ? ne2  : ne1;
+    const int64_t nchannels_y        = ids ? ne11 : ne12;
+    const int64_t nchannels_dst      = ids ? ne1  : ne2;
+    const int64_t stride_channel_dst = ids ? s1   : s2;
+    const int64_t stride_channel_y   = ids ? s11  : s12;
+
+    GGML_ASSERT(!ids || ncols_dst == 1);
+
+    switch (src0->type) {
+        case GGML_TYPE_F32: {
+            const float * src0_d = (const float *) src0->data;
+            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
+                ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
+                ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
+        } break;
+        case GGML_TYPE_F16: {
+            const half * src0_d = (const half *) src0->data;
+            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
+                ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
+                ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
+        } break;
+        case GGML_TYPE_BF16: {
+            const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0->data;
+            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
+                ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
+                ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
+        } break;
+        default:
+            GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));
+    }
+}
+
+void ggml_cuda_op_mul_mat_vec_f(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream) {
+
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne0  =  dst->ne[0];
+    const int64_t row_diff = row_high - row_low;
+
+    const int id = ggml_cuda_get_device();
+    const int cc = ggml_cuda_info().devices[id].cc;
+    const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
+
+
+    // ggml_cuda_op provides single, contiguous matrices
+    const int64_t stride_row         = ne00;
+    const int64_t stride_col_y       = ne10;
+    const int64_t stride_col_dst     = id == ctx.device ? ne0 : row_diff; // main device has larger memory buffer
+    const int64_t nchannels_x        = 1;
+    const int64_t nchannels_y        = 1;
+    const int64_t nchannels_dst      = 1;
+    const int64_t stride_channel_x   = 0;
+    const int64_t stride_channel_y   = 0;
+    const int64_t stride_channel_dst = 0;
+    const int64_t nsamples_x         = 1;
+    const int64_t nsamples_dst       = 1;
+    const int64_t stride_sample_x    = 0;
+    const int64_t stride_sample_y    = 0;
+    const int64_t stride_sample_dst  = 0;
+
+    switch (src0->type) {
+        case GGML_TYPE_F32: {
+            const float * src0_d = (const float *) src0_dd_i;
+            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
+                nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
+        } break;
+        case GGML_TYPE_F16: {
+            const half * src0_d = (const half *) src0_dd_i;
+            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
+                nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
+        } break;
+        case GGML_TYPE_BF16: {
+            const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0_dd_i;
+            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
+                nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
+        } break;
+        default:
+            GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));
+    }
+
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_ddq_i);
+    GGML_UNUSED(src1_ncols);
+    GGML_UNUSED(src1_padded_row_size);
+}
+
+bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, const int64_t * src0_ne, int64_t ne11) {
+    if (src0_ne[0] % 2 != 0) {
+        return false;
+    }
+    switch (type) {
+        case GGML_TYPE_F32:
+            if (GGML_CUDA_CC_IS_NVIDIA(cc)) {
+                if (ampere_mma_available(cc)) {
+                    return ne11 <= 3;
+                }
+                if (cc >= GGML_CUDA_CC_TURING) {
+                    return ne11 <= 4;
+                }
+                return ne11 <= 3;
+            } else if (GGML_CUDA_CC_IS_AMD(cc)) {
+                if (fp32_mma_hardware_available(cc)) {
+                    return ne11 <= 3;
+                }
+                return ne11 <= 8;
+            }
+            return ne11 <= 8;
+        case GGML_TYPE_F16:
+            if (GGML_CUDA_CC_IS_NVIDIA(cc)) {
+                const bool src0_small = (src0_ne[1] <= 512 || src0_ne[2]*src0_ne[3] == 1);
+                if (ampere_mma_available(cc)) {
+                    return src0_small && ne11 == 1;
+                }
+                if (cc >= GGML_CUDA_CC_ADA_LOVELACE) {
+                    return src0_small && ne11 <= 4;
+                }
+                if (fp16_mma_hardware_available(cc)) {
+                    return src0_small && ne11 <= 3;
+                }
+                return ne11 <= 8;
+            } else if (GGML_CUDA_CC_IS_AMD(cc)) {
+                if (fp16_mma_hardware_available(cc)) {
+                    if (GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
+                        return ne11 <= 5;
+                    }
+                    return ne11 <= 2;
+                }
+                return ne11 <= 8;
+            }
+            return ne11 <= 8;
+        case GGML_TYPE_BF16:
+            if (GGML_CUDA_CC_IS_NVIDIA(cc)) {
+                const bool src0_small = (src0_ne[1] <= 512 || src0_ne[2]*src0_ne[3] == 1);
+                if (ampere_mma_available(cc)) {
+                    return src0_small && ne11 == 1;
+                }
+                if (cc >= GGML_CUDA_CC_ADA_LOVELACE) {
+                    return src0_small && ne11 <= 4;
+                }
+                if (bf16_mma_hardware_available(cc)) {
+                    return src0_small && ne11 <= 3;
+                }
+                return ne11 <= 8;
+            } else if (GGML_CUDA_CC_IS_AMD(cc)) {
+                if (bf16_mma_hardware_available(cc)) {
+                    return ne11 <= 3;
+                }
+                return ne11 <= 8;
+            }
+            return ne11 <= 8;
+        default:
+            return false;
+    }
+}
diff --git a/ggml/src/ggml-cuda/mmvf.cuh b/ggml/src/ggml-cuda/mmvf.cuh
new file mode 100644
index 00000000000..1da460992e7
--- /dev/null
+++ b/ggml/src/ggml-cuda/mmvf.cuh
@@ -0,0 +1,11 @@
+#include "common.cuh"
+
+void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
+
+void ggml_cuda_op_mul_mat_vec_f(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream);
+
+bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, const int64_t * src0_ne, int64_t ne11);
diff --git a/ggml/src/ggml-cuda/vendors/hip.h b/ggml/src/ggml-cuda/vendors/hip.h
index 8b172e60f4b..c31f3192322 100644
--- a/ggml/src/ggml-cuda/vendors/hip.h
+++ b/ggml/src/ggml-cuda/vendors/hip.h
@@ -200,6 +200,7 @@
 #endif
 
 typedef hip_bfloat16 nv_bfloat16;
+typedef short2 nv_bfloat162; // FIXME there is no 2x BF16 type being defined in bfloat16.h, ad-hoc compilation fix
 
 typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
 typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
diff --git a/ggml/src/ggml-cuda/vendors/musa.h b/ggml/src/ggml-cuda/vendors/musa.h
index 19896320244..8c55a2e4e56 100644
--- a/ggml/src/ggml-cuda/vendors/musa.h
+++ b/ggml/src/ggml-cuda/vendors/musa.h
@@ -137,4 +137,5 @@
 #define cudaStreamEndCapture musaStreamEndCapture
 #define cudaOccupancyMaxActiveBlocksPerMultiprocessor musaOccupancyMaxActiveBlocksPerMultiprocessor
 
-typedef mt_bfloat16 nv_bfloat16;
+typedef __mt_bfloat16 nv_bfloat16;
+typedef __mt_bfloat162 nv_bfloat162;

From b8ac343022c1d47b31135bdbad991c3871d8a713 Mon Sep 17 00:00:00 2001
From: Christian Kastner <ckk@kvr.at>
Date: Thu, 7 Aug 2025 13:45:41 +0200
Subject: [PATCH 40/86] ggml: Skip backend library linking code when
 GGML_BACKEND_DL=ON (llama/15094)

Any available libraries are found and loaded dynamically at runtime.
---
 ggml/cmake/ggml-config.cmake.in | 82 +++++++++++++++++----------------
 1 file changed, 42 insertions(+), 40 deletions(-)

diff --git a/ggml/cmake/ggml-config.cmake.in b/ggml/cmake/ggml-config.cmake.in
index 65a75a2f1c8..91c9d5cd343 100644
--- a/ggml/cmake/ggml-config.cmake.in
+++ b/ggml/cmake/ggml-config.cmake.in
@@ -106,7 +106,7 @@ if(NOT TARGET ggml::ggml)
 
     find_library(GGML_LIBRARY ggml
         REQUIRED
-        HINTS ${GGML_LIB_DIR} ${GGML_BACKEND_DIR}
+        HINTS ${GGML_LIB_DIR}
         NO_CMAKE_FIND_ROOT_PATH)
 
     add_library(ggml::ggml UNKNOWN IMPORTED)
@@ -125,54 +125,56 @@ if(NOT TARGET ggml::ggml)
             IMPORTED_LOCATION "${GGML_BASE_LIBRARY}")
 
     set(_ggml_all_targets "")
-    foreach(_ggml_backend ${GGML_AVAILABLE_BACKENDS})
-        string(REPLACE "-" "_" _ggml_backend_pfx "${_ggml_backend}")
-        string(TOUPPER "${_ggml_backend_pfx}" _ggml_backend_pfx)
-
-        find_library(${_ggml_backend_pfx}_LIBRARY ${_ggml_backend}
-            REQUIRED
-            HINTS ${GGML_LIB_DIR}
-            NO_CMAKE_FIND_ROOT_PATH)
-
-        message(STATUS "Found ${${_ggml_backend_pfx}_LIBRARY}")
-
-        add_library(ggml::${_ggml_backend} UNKNOWN IMPORTED)
-        set_target_properties(ggml::${_ggml_backend}
-            PROPERTIES
-                INTERFACE_INCLUDE_DIRECTORIES "${GGML_INCLUDE_DIR}"
-                IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
-                IMPORTED_LOCATION "${${_ggml_backend_pfx}_LIBRARY}"
-                INTERFACE_COMPILE_FEATURES c_std_90
-                POSITION_INDEPENDENT_CODE ON)
-
-        string(REGEX MATCH "^ggml-cpu" is_cpu_variant "${_ggml_backend}")
-        if(is_cpu_variant)
-            list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES "ggml::ggml-base")
-            set_target_properties(ggml::${_ggml_backend}
-            PROPERTIES
-                INTERFACE_LINK_LIBRARIES "${GGML_CPU_INTERFACE_LINK_LIBRARIES}")
+    if (NOT GGML_BACKEND_DL)
+        foreach(_ggml_backend ${GGML_AVAILABLE_BACKENDS})
+            string(REPLACE "-" "_" _ggml_backend_pfx "${_ggml_backend}")
+            string(TOUPPER "${_ggml_backend_pfx}" _ggml_backend_pfx)
 
-            if(GGML_CPU_INTERFACE_LINK_OPTIONS)
-                set_target_properties(ggml::${_ggml_backend}
-                    PROPERTIES
-                        INTERFACE_LINK_OPTIONS "${GGML_CPU_INTERFACE_LINK_OPTIONS}")
-            endif()
+            find_library(${_ggml_backend_pfx}_LIBRARY ${_ggml_backend}
+                REQUIRED
+                HINTS ${GGML_LIB_DIR}
+                NO_CMAKE_FIND_ROOT_PATH)
+
+            message(STATUS "Found ${${_ggml_backend_pfx}_LIBRARY}")
 
-        else()
-            list(APPEND ${_ggml_backend_pfx}_INTERFACE_LINK_LIBRARIES "ggml::ggml-base")
+            add_library(ggml::${_ggml_backend} UNKNOWN IMPORTED)
             set_target_properties(ggml::${_ggml_backend}
                 PROPERTIES
-                    INTERFACE_LINK_LIBRARIES "${${_ggml_backend_pfx}_INTERFACE_LINK_LIBRARIES}")
+                    INTERFACE_INCLUDE_DIRECTORIES "${GGML_INCLUDE_DIR}"
+                    IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
+                    IMPORTED_LOCATION "${${_ggml_backend_pfx}_LIBRARY}"
+                    INTERFACE_COMPILE_FEATURES c_std_90
+                    POSITION_INDEPENDENT_CODE ON)
+
+            string(REGEX MATCH "^ggml-cpu" is_cpu_variant "${_ggml_backend}")
+            if(is_cpu_variant)
+                list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES "ggml::ggml-base")
+                set_target_properties(ggml::${_ggml_backend}
+                PROPERTIES
+                    INTERFACE_LINK_LIBRARIES "${GGML_CPU_INTERFACE_LINK_LIBRARIES}")
+
+                if(GGML_CPU_INTERFACE_LINK_OPTIONS)
+                    set_target_properties(ggml::${_ggml_backend}
+                        PROPERTIES
+                            INTERFACE_LINK_OPTIONS "${GGML_CPU_INTERFACE_LINK_OPTIONS}")
+                endif()
 
-            if(${_ggml_backend_pfx}_INTERFACE_LINK_OPTIONS)
+            else()
+                list(APPEND ${_ggml_backend_pfx}_INTERFACE_LINK_LIBRARIES "ggml::ggml-base")
                 set_target_properties(ggml::${_ggml_backend}
                     PROPERTIES
-                        INTERFACE_LINK_OPTIONS "${${_ggml_backend_pfx}_INTERFACE_LINK_OPTIONS}")
+                        INTERFACE_LINK_LIBRARIES "${${_ggml_backend_pfx}_INTERFACE_LINK_LIBRARIES}")
+
+                if(${_ggml_backend_pfx}_INTERFACE_LINK_OPTIONS)
+                    set_target_properties(ggml::${_ggml_backend}
+                        PROPERTIES
+                            INTERFACE_LINK_OPTIONS "${${_ggml_backend_pfx}_INTERFACE_LINK_OPTIONS}")
+                endif()
             endif()
-        endif()
 
-        list(APPEND _ggml_all_targets ggml::${_ggml_backend})
-    endforeach()
+            list(APPEND _ggml_all_targets ggml::${_ggml_backend})
+        endforeach()
+    endif()
 
     list(APPEND GGML_INTERFACE_LINK_LIBRARIES ggml::ggml-base "${_ggml_all_targets}")
     set_target_properties(ggml::ggml

From 6642ea839d497f9592ab6cc80f8eef1a73d8ff11 Mon Sep 17 00:00:00 2001
From: uvos <carl@uvos.xyz>
Date: Thu, 7 Aug 2025 16:44:14 +0200
Subject: [PATCH 41/86] HIP: add cmake option to enable compiler output of
 kernel resource usage metrics (llama/15103)

---
 ggml/CMakeLists.txt              | 1 +
 ggml/src/ggml-hip/CMakeLists.txt | 4 ++++
 2 files changed, 5 insertions(+)

diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt
index 7c4666b7edb..1fb7abeaf08 100644
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -176,6 +176,7 @@ option(GGML_HIP_NO_VMM                      "ggml: do not try to use HIP VMM"
 option(GGML_HIP_ROCWMMA_FATTN               "ggml: enable rocWMMA for FlashAttention"         OFF)
 option(GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12   "ggml: enable rocWMMA FlashAttention on GFX12"    OFF)
 option(GGML_HIP_MMQ_MFMA                    "ggml: enable MFMA MMA for CDNA in MMQ"           ON)
+option(GGML_HIP_EXPORT_METRICS              "ggml: enable kernel perf metrics output"         OFF)
 option(GGML_MUSA_GRAPHS                     "ggml: use MUSA graph, experimental, unstable"    OFF)
 option(GGML_MUSA_MUDNN_COPY                 "ggml: enable muDNN for accelerated copy"         OFF)
 option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
diff --git a/ggml/src/ggml-hip/CMakeLists.txt b/ggml/src/ggml-hip/CMakeLists.txt
index e92ec7faa33..852de973460 100644
--- a/ggml/src/ggml-hip/CMakeLists.txt
+++ b/ggml/src/ggml-hip/CMakeLists.txt
@@ -121,6 +121,10 @@ if (GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12 OR ${hip_VERSION} VERSION_GREATER_EQUAL 7
     add_compile_definitions(GGML_HIP_ROCWMMA_FATTN_GFX12)
 endif()
 
+if (GGML_HIP_EXPORT_METRICS)
+    set(CMAKE_HIP_FLAGS "${CMAKE_HIP_FLAGS} -Rpass-analysis=kernel-resource-usage --save-temps")
+endif()
+
 if (NOT GGML_CUDA_FA)
     add_compile_definitions(GGML_CUDA_NO_FA)
 endif()

From ba6238a268d7fda15648583a00584ff48b79b7c3 Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Thu, 7 Aug 2025 15:07:11 -0500
Subject: [PATCH 42/86] vulkan: Add env var to disable host visible vidmem
 (llama/15109)

---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 165933a7294..f1cb90e3b36 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -534,6 +534,7 @@ struct vk_device_struct {
     ggml_backend_buffer_type buffer_type;
 
     bool disable_fusion;
+    bool disable_host_visible_vidmem;
 
 #ifdef GGML_VULKAN_MEMORY_DEBUG
     std::unique_ptr<vk_memory_logger> memory_logger;
@@ -1804,6 +1805,8 @@ static vk_buffer ggml_vk_create_buffer_device(vk_device& device, size_t size) {
         } else if (device->uma) {
             // Fall back to host memory type
             buf = ggml_vk_create_buffer(device, size, vk::MemoryPropertyFlagBits::eDeviceLocal, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
+        } else if (device->disable_host_visible_vidmem) {
+            buf = ggml_vk_create_buffer(device, size, vk::MemoryPropertyFlagBits::eDeviceLocal, vk::MemoryPropertyFlagBits::eDeviceLocal);
         } else {
             // use rebar if available, otherwise fallback to device only visible memory
             buf = ggml_vk_create_buffer(device, size, vk::MemoryPropertyFlagBits::eDeviceLocal | vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent, vk::MemoryPropertyFlagBits::eDeviceLocal);
@@ -3265,6 +3268,9 @@ static vk_device ggml_vk_get_device(size_t idx) {
         const char* GGML_VK_PREFER_HOST_MEMORY = getenv("GGML_VK_PREFER_HOST_MEMORY");
         device->prefer_host_memory = GGML_VK_PREFER_HOST_MEMORY != nullptr;
 
+        const char* GGML_VK_DISABLE_HOST_VISIBLE_VIDMEM = getenv("GGML_VK_DISABLE_HOST_VISIBLE_VIDMEM");
+        device->disable_host_visible_vidmem = GGML_VK_DISABLE_HOST_VISIBLE_VIDMEM != nullptr;
+
         bool fp16_storage = false;
         bool fp16_compute = false;
         bool maintenance4_support = false;

From 244c4949aa2156ba0657c5ad6d645997ebfdcb60 Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Thu, 7 Aug 2025 15:44:20 -0500
Subject: [PATCH 43/86] vulkan: support fattn sinks (llama/15126)

---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp          | 58 +++++++++++++------
 .../vulkan-shaders/flash_attn.comp            | 21 +++++++
 .../vulkan-shaders/flash_attn_base.comp       | 13 ++++-
 .../vulkan-shaders/flash_attn_cm1.comp        | 21 +++++++
 .../vulkan-shaders/flash_attn_cm2.comp        | 28 +++++++++
 .../flash_attn_split_k_reduce.comp            | 27 ++++++++-
 6 files changed, 148 insertions(+), 20 deletions(-)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index f1cb90e3b36..b1cbbc9866c 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -2286,14 +2286,14 @@ static void ggml_vk_load_shaders(vk_device& device) {
     };
 
 #define CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, HSK, HSV, HEAD_SIZES) \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][0][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f16acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,false), 1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][0][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f16acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,false), fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,false)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][0][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f32acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,false), 1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][0][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f32acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,false), fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,false)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][1][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f16acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,true),   1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][1][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f16acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,true),   fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,true)[1],  true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][1][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f32acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,true),   1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][1][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f32acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,true),   fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,true)[1],  true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][0][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f16acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,false), 1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][0][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f16acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,false), fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,false)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][0][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f32acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,false), 1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][0][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f32acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,false), fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,false)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][1][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f16acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,true),   1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][1][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f16acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,true),   fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,true)[1],  true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][1][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f32acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,true),   1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][1][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f32acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,true),   fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,true)[1],  true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
 
 #define CREATE_FA(TYPE, NAMELC, FAPATH, SUFFIX) \
         CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 64, 64, 64) \
@@ -2910,7 +2910,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_MXFP4],   "get_rows_mxfp4_f32",   get_rows_mxfp4_f32_len,   get_rows_mxfp4_f32_data,   "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_matmul_split_k_reduce, "split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256 * 4, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_flash_attn_split_k_reduce, "fa_split_k_reduce", fa_split_k_reduce_len, fa_split_k_reduce_data, "main", 2, 4 * sizeof(uint32_t), {1, device->subgroup_size, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_flash_attn_split_k_reduce, "fa_split_k_reduce", fa_split_k_reduce_len, fa_split_k_reduce_data, "main", 3, 5 * sizeof(uint32_t), {1, device->subgroup_size, 1}, {device->subgroup_size}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1, "quantize_q8_1", quantize_q8_1_len, quantize_q8_1_data, "main", 2, 1 * sizeof(uint32_t), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1);
 
     for (uint32_t i = 0; i < p021_max_gqa_ratio; ++i) {
@@ -6507,11 +6507,14 @@ static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, co
     return supported;
 }
 
-static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * q, const ggml_tensor * k, const ggml_tensor * v, const ggml_tensor * mask, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * q, const ggml_tensor * k, const ggml_tensor * v, const ggml_tensor * mask, const ggml_tensor * sinks, ggml_tensor * dst, bool dryrun = false) {
     VK_LOG_DEBUG("ggml_vk_flash_attn((" << q << ", name=" << q->name << ", type=" << q->type << ", ne0=" << q->ne[0] << ", ne1=" << q->ne[1] << ", ne2=" << q->ne[2] << ", ne3=" << q->ne[3] << ", nb0=" << q->nb[0] << ", nb1=" << q->nb[1] << ", nb2=" << q->nb[2] << ", nb3=" << q->nb[3];
     std::cerr << "), (" << k << ", name=" << k->name << ", type=" << k->type << ", ne0=" << k->ne[0] << ", ne1=" << k->ne[1] << ", ne2=" << k->ne[2] << ", ne3=" << k->ne[3] << ", nb0=" << k->nb[0] << ", nb1=" << k->nb[1] << ", nb2=" << k->nb[2] << ", nb3=" << k->nb[3];
     std::cerr << "), (" << v << ", name=" << v->name << ", type=" << v->type << ", ne0=" << v->ne[0] << ", ne1=" << v->ne[1] << ", ne2=" << v->ne[2] << ", ne3=" << v->ne[3] << ", nb0=" << v->nb[0] << ", nb1=" << v->nb[1] << ", nb2=" << v->nb[2] << ", nb3=" << v->nb[3];
     std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
+    if (sinks) {
+        std::cerr << "), (" << sinks << ", name=" << sinks->name << ", type=" << sinks->type << ", ne0=" << sinks->ne[0] << ", ne1=" << sinks->ne[1] << ", ne2=" << sinks->ne[2] << ", ne3=" << sinks->ne[3] << ", nb0=" << sinks->nb[0] << ", nb1=" << sinks->nb[1] << ", nb2=" << sinks->nb[2] << ", nb3=" << sinks->nb[3];
+    }
     std::cerr << "), " << (dryrun ? "dryrun" : "") << ")");
 
     GGML_TENSOR_LOCALS(int64_t, neq, q,   ne)
@@ -6710,10 +6713,10 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
-    vk_buffer d_Q = nullptr, d_K = nullptr, d_V = nullptr, d_D = nullptr, d_M = nullptr;
-    size_t q_buf_offset = 0, k_buf_offset = 0, v_buf_offset = 0, d_buf_offset = 0, m_buf_offset = 0;
+    vk_buffer d_Q = nullptr, d_K = nullptr, d_V = nullptr, d_D = nullptr, d_M = nullptr, d_S = nullptr;
+    size_t q_buf_offset = 0, k_buf_offset = 0, v_buf_offset = 0, d_buf_offset = 0, m_buf_offset = 0, s_buf_offset = 0;
 
-    bool Q_uma = false, K_uma = false, V_uma = false, D_uma = false, M_uma = false;
+    bool Q_uma = false, K_uma = false, V_uma = false, D_uma = false, M_uma = false, S_uma = false;
 
     if (ctx->device->uma) {
         ggml_vk_host_get(ctx->device, q->data, d_Q, q_buf_offset);
@@ -6728,6 +6731,10 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
             ggml_vk_host_get(ctx->device, mask->data, d_M, m_buf_offset);
             M_uma = d_M != nullptr;
         }
+        if (sinks) {
+            ggml_vk_host_get(ctx->device, sinks->data, d_S, s_buf_offset);
+            S_uma = d_S != nullptr;
+        }
     }
 
 
@@ -6763,7 +6770,17 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         }
     }
 
-    uint32_t mask_n_head_log2 = ((mask != nullptr) << 16) | n_head_log2;
+    if (!S_uma) {
+        d_S = d_Q;
+        s_buf_offset = q_buf_offset;
+        if (sinks) {
+            ggml_backend_vk_buffer_context * s_buf_ctx = (ggml_backend_vk_buffer_context*)sinks->buffer->context;
+            d_S = s_buf_ctx->dev_buffer;
+            s_buf_offset = vk_tensor_offset(sinks) + sinks->view_offs;
+        }
+    }
+
+    uint32_t mask_n_head_log2 = ((sinks != nullptr) << 24) | ((mask != nullptr) << 16) | n_head_log2;
 
     const vk_flash_attn_push_constants pc = { N, KV,
                                               (uint32_t)ne1, (uint32_t)ne2, (uint32_t)ne3,
@@ -6787,6 +6804,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                         vk_subbuffer{d_K, k_buf_offset, VK_WHOLE_SIZE},
                                         vk_subbuffer{d_V, v_buf_offset, VK_WHOLE_SIZE},
                                         vk_subbuffer{d_M, m_buf_offset, VK_WHOLE_SIZE},
+                                        vk_subbuffer{d_S, s_buf_offset, VK_WHOLE_SIZE},
                                         vk_subbuffer{ctx->prealloc_split_k, 0, VK_WHOLE_SIZE},
                                     },
                                     // We only use split_k when group query attention is enabled, which means
@@ -6796,10 +6814,11 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                     pc, { workgroups_x * pipeline->wg_denoms[0], workgroups_y, workgroups_z });
 
         ggml_vk_sync_buffers(subctx);
-        const std::array<uint32_t, 4> pc2 = { HSV, (uint32_t)ne1, (uint32_t)ne3, split_k };
+        const std::array<uint32_t, 5> pc2 = { HSV, (uint32_t)ne1, (uint32_t)ne3, split_k, (sinks != nullptr) };
         ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_flash_attn_split_k_reduce,
                                     {
                                         vk_subbuffer{ctx->prealloc_split_k, 0, VK_WHOLE_SIZE},
+                                        vk_subbuffer{d_S, s_buf_offset, VK_WHOLE_SIZE},
                                         vk_subbuffer{d_D, d_buf_offset, VK_WHOLE_SIZE},
                                     },
                                     pc2, { (uint32_t)ne1, HSV, (uint32_t)ne3 });
@@ -6810,6 +6829,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                         vk_subbuffer{d_K, k_buf_offset, VK_WHOLE_SIZE},
                                         vk_subbuffer{d_V, v_buf_offset, VK_WHOLE_SIZE},
                                         vk_subbuffer{d_M, m_buf_offset, VK_WHOLE_SIZE},
+                                        vk_subbuffer{d_S, s_buf_offset, VK_WHOLE_SIZE},
                                         vk_subbuffer{d_D, d_buf_offset, VK_WHOLE_SIZE},
                                     },
                                     pc, { workgroups_x, workgroups_y, workgroups_z });
@@ -9874,7 +9894,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         break;
 
     case GGML_OP_FLASH_ATTN_EXT:
-        ggml_vk_flash_attn(ctx, compute_ctx, src0, src1, src2, src3, node, dryrun);
+        ggml_vk_flash_attn(ctx, compute_ctx, src0, src1, src2, src3, node->src[4], node, dryrun);
 
         break;
 
@@ -10951,8 +10971,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 if (head_sizes == FA_HEAD_SIZE_UNSUPPORTED) {
                     return false;
                 }
-                // TODO: support attention sinks [TAG_ATTN_SINKS]
-                if (op->src[4]) {
+                if (op->src[4] && op->src[4]->type != GGML_TYPE_F32) {
                     return false;
                 }
                 if (op->src[0]->type != GGML_TYPE_F32) {
@@ -11547,6 +11566,9 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
     if (tensor->op == GGML_OP_FLASH_ATTN_EXT) {
         const float * params = (const float *)tensor->op_params;
         tensor_clone = ggml_flash_attn_ext(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], src_clone[3], params[0], params[1], params[2]);
+        if (src_clone[4]) {
+            ggml_flash_attn_ext_add_sinks(tensor_clone, src_clone[4]);
+        }
     } else if (tensor->op == GGML_OP_MUL_MAT) {
         tensor_clone = ggml_mul_mat(ggml_ctx, src_clone[0], src_clone[1]);
     } else if (tensor->op == GGML_OP_MUL_MAT_ID) {
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp
index 45c6e7736ac..d40848e15fe 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp
@@ -305,6 +305,27 @@ void main() {
         return;
     }
 
+    if ((p.mask_n_head_log2 & SINK_ENABLE_BIT) != 0) {
+        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+            float sink = perElemOpGetSink(r, 0u, ACC_TYPE(0), iq2);
+
+            float ms = 1.0f;
+            float vs = 1.0f;
+
+            if (sink > Mf[r]) {
+                ms = exp(Mf[r] - sink);
+
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+                    Of[r][d] *= ms;
+                }
+            } else {
+                vs = exp(sink - Mf[r]);
+            }
+
+            Lf[r] = Lf[r]*ms + vs;
+        }
+    }
+
     float Lfrcp[Br];
     [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
         Lfrcp[r] = 1.0 / Lf[r];
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp
index 7defe72b403..b57c9dcfc4e 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp
@@ -50,10 +50,13 @@ layout (push_constant) uniform parameter {
     uint32_t k_num;
 } p;
 
+#define SINK_ENABLE_BIT (1<<24)
 #define MASK_ENABLE_BIT (1<<16)
 #define N_LOG2_MASK 0xFFFF
 
-layout (binding = 4) writeonly buffer O {D_TYPE data_o[];};
+layout (binding = 4) readonly buffer S {float data_s[];};
+
+layout (binding = 5) writeonly buffer O {D_TYPE data_o[];};
 
 #if defined(A_TYPE_PACKED16)
 #define BINDING_IDX_K 0
@@ -111,6 +114,14 @@ ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const i
     return ACC_TYPE(pow(base, ACC_TYPE(exph)));
 }
 
+// Load the sink value, indexed by Q's dimension 2.
+ACC_TYPE perElemOpGetSink(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t iq2)
+{
+    const uint32_t h = iq2 + (r % p.gqa_ratio);
+
+    return ACC_TYPE(data_s[h]);
+}
+
 uint32_t i, N, KV, split_k_index, Tr, start_j, end_j,
          iq2, iq3, rk2, rk3, rv2, rv3, ik2, ik3, iv2, iv3,
          q_stride, k_stride, v_stride, m_stride;
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
index 486735fe8b0..230e815f22c 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
@@ -329,6 +329,27 @@ void main() {
         return;
     }
 
+    if ((p.mask_n_head_log2 & SINK_ENABLE_BIT) != 0) {
+        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+            float sink = perElemOpGetSink(r, 0u, ACC_TYPE(0), iq2);
+
+            float ms = 1.0f;
+            float vs = 1.0f;
+
+            if (sink > Mf[r]) {
+                ms = exp(Mf[r] - sink);
+
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+                    Of[r][d] *= ACC_TYPE(ms);
+                }
+            } else {
+                vs = exp(sink - Mf[r]);
+            }
+
+            Lf[r] = Lf[r]*ms + vs;
+        }
+    }
+
     float Lfrcp[rows_per_thread];
     [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
         Lfrcp[r] = 1.0 / Lf[r];
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
index 274f48fcabd..b0564ca0bfc 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
@@ -248,6 +248,34 @@ void main() {
     // resize L by using smear/reduce
     coopMatReduceNV(Ldiag, L, gl_CooperativeMatrixReduceRowNV, smearReduce);
 
+    if ((p.mask_n_head_log2 & SINK_ENABLE_BIT) != 0) {
+        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> S;
+        coopMatPerElementNV(S, S, perElemOpGetSink, iq2);
+
+        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> Mr;
+
+        // resize M by using smear/reduce
+        coopMatReduceNV(Mr, M, gl_CooperativeMatrixReduceRowNV, smearReduce);
+
+        // O, Ldiag, Mr all have the same type so all element locations match
+        [[unroll]] for (uint32_t i = 0; i < Ldiag.length(); ++i) {
+            ACC_TYPE sink = S[i];
+
+            ACC_TYPE ms = ACC_TYPE(1.0f);
+            ACC_TYPE vs = ACC_TYPE(1.0f);
+
+            if (sink > Mr[i]) {
+                ms = exp(Mr[i] - sink);
+
+                O[i] *= ms;
+            } else {
+                vs = exp(sink - Mr[i]);
+            }
+
+            Ldiag[i] = Ldiag[i]*ms + vs;
+        }
+    }
+
     [[unroll]]
     for (int k = 0; k < Ldiag.length(); ++k) {
         Ldiag[k] = ACC_TYPE(1.0) / Ldiag[k];
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp
index 0a17a9df23f..76ef4b6dfb5 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp
@@ -7,13 +7,15 @@ layout(constant_id = 0) const uint BLOCK_SIZE = 32;
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {float data_a[];};
-layout (binding = 1) writeonly buffer D {float data_d[];};
+layout (binding = 1) readonly buffer B {float data_s[];};
+layout (binding = 2) writeonly buffer D {float data_d[];};
 
 layout (push_constant) uniform parameter {
     uint D;
     uint N;
     uint ne3;
     uint k_num;
+    uint sinks;
 } p;
 
 shared float tmpsh[BLOCK_SIZE];
@@ -73,6 +75,22 @@ void main() {
     }
     L = tmpsh[0];
 
+    float sink;
+    if (p.sinks != 0) {
+        sink = data_s[n];
+
+        float ms = 1.0f;
+        float vs = 1.0f;
+
+        if (sink > m_max) {
+            ms = exp(m_max - sink);
+        } else {
+            vs = exp(sink - m_max);
+        }
+
+        L = L*ms + vs;
+    }
+
     L = 1.0 / L;
 
     // D dimension is split across workgroups in the y dimension
@@ -85,6 +103,13 @@ void main() {
             float m = data_a[m_offset + k * lm_stride];
             O += exp(m - m_max) * data_a[o_offset];
         }
+        if (p.sinks != 0) {
+            if (sink > m_max) {
+                float ms = 1.0f;
+                ms = exp(m_max - sink);
+                O *= ms;
+            }
+        }
         O *= L;
         data_d[iq3 * D * N + D * n + d] = O;
     }

From 648591012035ddc52b2f0f9eae11c9bd5beb78db Mon Sep 17 00:00:00 2001
From: lhez <lih@qti.qualcomm.com>
Date: Fri, 8 Aug 2025 13:47:03 +0900
Subject: [PATCH 44/86] opencl: support sink in `soft_max` (attn sinks)
 (llama/15152)

---
 ggml/src/ggml-opencl/ggml-opencl.cpp          | 49 +++++++++++--------
 ggml/src/ggml-opencl/kernels/softmax_4_f16.cl | 12 ++++-
 ggml/src/ggml-opencl/kernels/softmax_4_f32.cl | 12 ++++-
 ggml/src/ggml-opencl/kernels/softmax_f16.cl   | 12 ++++-
 ggml/src/ggml-opencl/kernels/softmax_f32.cl   | 12 ++++-
 5 files changed, 68 insertions(+), 29 deletions(-)

diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
index 4f765ab5309..b32d5da3075 100644
--- a/ggml/src/ggml-opencl/ggml-opencl.cpp
+++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
@@ -2520,8 +2520,6 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
         case GGML_OP_CLAMP:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_SOFT_MAX:
-            // TODO: support attention sinks [TAG_ATTN_SINKS]
-            return op->src[2] == nullptr;
         case GGML_OP_NORM:
         case GGML_OP_RMS_NORM:
             return true;
@@ -6594,17 +6592,24 @@ static void ggml_cl_soft_max(ggml_backend_t backend, const ggml_tensor * src0, c
         GGML_ASSERT(src1->extra);
     }
 
+    const ggml_tensor * src2 = dst->src[2];
+    if (src2) {
+        GGML_ASSERT(src2->extra);
+    }
+
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
     ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
     ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
 
     ggml_tensor_extra_cl * extra1 = src1 ? (ggml_tensor_extra_cl *)src1->extra : nullptr;
+    ggml_tensor_extra_cl * extra2 = src2 ? (ggml_tensor_extra_cl *)src2->extra : nullptr;
 
     cl_ulong offset0 = extra0->offset + src0->view_offs;
     cl_ulong offsetd = extrad->offset + dst->view_offs;
 
     cl_ulong offset1 = extra1 ? extra1->offset + src1->view_offs : offset0;
+    cl_ulong offset2 = extra2 ? extra2->offset + src2->view_offs : offset0;
 
     const int ne00 = src0->ne[0];
     const int ne01 = src0->ne[1];
@@ -6672,25 +6677,27 @@ static void ggml_cl_soft_max(ggml_backend_t backend, const ggml_tensor * src0, c
     CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
     CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   extra1 ? &extra1->data_device : &extra0->data_device));
     CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
-    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
-    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
-    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
-    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb01));
-    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb02));
-    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb03));
-    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
-    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne13));
-    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb11));
-    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb12));
-    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb13));
-    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb1));
-    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb2));
-    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb3));
-    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(float),    &scale));
-    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(float),    &max_bias));
-    CL_CHECK(clSetKernelArg(kernel, 20, sizeof(float),    &m0));
-    CL_CHECK(clSetKernelArg(kernel, 21, sizeof(float),    &m1));
-    CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &n_head_log2));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   extra2 ? &extra2->data_device : &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne13));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb12));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb13));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb3));
+    CL_CHECK(clSetKernelArg(kernel, 20, sizeof(float),    &scale));
+    CL_CHECK(clSetKernelArg(kernel, 21, sizeof(float),    &max_bias));
+    CL_CHECK(clSetKernelArg(kernel, 22, sizeof(float),    &m0));
+    CL_CHECK(clSetKernelArg(kernel, 23, sizeof(float),    &m1));
+    CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int),      &n_head_log2));
 
     size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
     size_t local_work_size[] = {(size_t)nth, 1, 1};
diff --git a/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl b/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl
index a6d8ede6701..571d16507c6 100644
--- a/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl
+++ b/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl
@@ -26,6 +26,8 @@ kernel void kernel_soft_max_4_f16(
         ulong offset0,
         global char * src1,
         ulong offset1,
+        global char * src2,
+        ulong offset2,
         global char * dst,
         ulong offsetd,
         int ne00,
@@ -48,6 +50,7 @@ kernel void kernel_soft_max_4_f16(
 ) {
     src0 = src0 + offset0;
     src1 = src1 + offset1;
+    src2 = src2 + offset2;
     dst  = dst  + offsetd;
 
     int i03 = get_group_id(2);
@@ -60,6 +63,7 @@ kernel void kernel_soft_max_4_f16(
 
     global float4 * psrc4 = (global float4 *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
     global half4  * pmask = src1 != src0 ? (global half4 *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0;
+    global float  * psrc2 = src2 != src0 ? (global float *)(src2) : 0;
     global float4 * pdst4 = (global float4 *)(dst  + i01*nb1 + i02*nb2 + i03*nb3);
 
     float slope = 1.0f;
@@ -75,7 +79,7 @@ kernel void kernel_soft_max_4_f16(
     }
 
     // parallel max
-    float4 lmax4 = -INFINITY;
+    float4 lmax4 = psrc2 ? psrc2[i02] : -INFINITY;
     for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
         lmax4 = fmax(lmax4, psrc4[i00]*scale + slope*(pmask ? convert_float4(pmask[i00]) : 0.0f));
     }
@@ -92,7 +96,11 @@ kernel void kernel_soft_max_4_f16(
     }
     float lsum = lsum4.s0 + lsum4.s1 + lsum4.s2 + lsum4.s3;
 
-    const float sum = sub_group_reduce_add(lsum);
+    float sum = sub_group_reduce_add(lsum);
+
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max);
+    }
 
     for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
         pdst4[i00] /= sum;
diff --git a/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl b/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl
index 35b5573b46a..1f944b2201d 100644
--- a/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl
+++ b/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl
@@ -26,6 +26,8 @@ kernel void kernel_soft_max_4(
         ulong offset0,
         global char * src1,
         ulong offset1,
+        global char * src2,
+        ulong offset2,
         global char * dst,
         ulong offsetd,
         int ne00,
@@ -48,6 +50,7 @@ kernel void kernel_soft_max_4(
 ) {
     src0 = src0 + offset0;
     src1 = src1 + offset1;
+    src2 = src2 + offset2;
     dst  = dst  + offsetd;
 
     int i03 = get_group_id(2);
@@ -60,6 +63,7 @@ kernel void kernel_soft_max_4(
 
     global float4 * psrc4 = (global float4 *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
     global float4 * pmask = src1 != src0 ? (global float4 *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0;
+    global float  * psrc2 = src2 != src0 ? (global float  *)(src2) : 0;
     global float4 * pdst4 = (global float4 *)(dst  + i01*nb1 + i02*nb2 + i03*nb3);
 
     float slope = 1.0f;
@@ -75,7 +79,7 @@ kernel void kernel_soft_max_4(
     }
 
     // parallel max
-    float4 lmax4 = -INFINITY;
+    float4 lmax4 = psrc2 ? psrc2[i02] : -INFINITY;
     for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
         lmax4 = fmax(lmax4, psrc4[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f));
     }
@@ -92,7 +96,11 @@ kernel void kernel_soft_max_4(
     }
     float lsum = lsum4.s0 + lsum4.s1 + lsum4.s2 + lsum4.s3;
 
-    const float sum = sub_group_reduce_add(lsum);
+    float sum = sub_group_reduce_add(lsum);
+
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max);
+    }
 
     for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
         pdst4[i00] /= sum;
diff --git a/ggml/src/ggml-opencl/kernels/softmax_f16.cl b/ggml/src/ggml-opencl/kernels/softmax_f16.cl
index 9d292b57465..4baa6c28e4f 100644
--- a/ggml/src/ggml-opencl/kernels/softmax_f16.cl
+++ b/ggml/src/ggml-opencl/kernels/softmax_f16.cl
@@ -26,6 +26,8 @@ kernel void kernel_soft_max_f16(
         ulong offset0,
         global char * src1,
         ulong offset1,
+        global char * src2,
+        ulong offset2,
         global char * dst,
         ulong offsetd,
         int ne00,
@@ -48,6 +50,7 @@ kernel void kernel_soft_max_f16(
 ) {
     src0 = src0 + offset0;
     src1 = src1 + offset1;
+    src2 = src2 + offset2;
     dst  = dst  + offsetd;
 
     int i03 = get_group_id(2);
@@ -60,6 +63,7 @@ kernel void kernel_soft_max_f16(
 
     global float * psrc0 = (global float *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
     global half  * pmask = src1 != src0 ? (global half *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0;
+    global float * psrc2 = src2 != src0 ? (global float *)(src2) : 0;
     global float * pdst  = (global float *)(dst  + i01*nb1 + i02*nb2 + i03*nb3);
 
     float slope = 1.0f;
@@ -75,7 +79,7 @@ kernel void kernel_soft_max_f16(
     }
 
     // parallel max
-    float lmax = -INFINITY;
+    float lmax = psrc2 ? psrc2[i02] : -INFINITY;
     for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
         lmax = fmax(lmax, psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f));
     }
@@ -91,7 +95,11 @@ kernel void kernel_soft_max_f16(
         pdst[i00] = exp_psrc0;
     }
 
-    const float sum = sub_group_reduce_add(lsum);
+    float sum = sub_group_reduce_add(lsum);
+
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max);
+    }
 
     for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
         pdst[i00] /= sum;
diff --git a/ggml/src/ggml-opencl/kernels/softmax_f32.cl b/ggml/src/ggml-opencl/kernels/softmax_f32.cl
index 7c53dfbe5a2..d503190b476 100644
--- a/ggml/src/ggml-opencl/kernels/softmax_f32.cl
+++ b/ggml/src/ggml-opencl/kernels/softmax_f32.cl
@@ -26,6 +26,8 @@ kernel void kernel_soft_max(
         ulong offset0,
         global char * src1,
         ulong offset1,
+        global char * src2,
+        ulong offset2,
         global char * dst,
         ulong offsetd,
         int ne00,
@@ -48,6 +50,7 @@ kernel void kernel_soft_max(
 ) {
     src0 = src0 + offset0;
     src1 = src1 + offset1;
+    src2 = src2 + offset2;
     dst  = dst  + offsetd;
 
     int i03 = get_group_id(2);
@@ -60,6 +63,7 @@ kernel void kernel_soft_max(
 
     global float * psrc0 = (global float *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
     global float * pmask = src1 != src0 ? (global float *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0;
+    global float * psrc2 = src2 != src0 ? (global float *)(src2) : 0;
     global float * pdst  = (global float *)(dst  + i01*nb1 + i02*nb2 + i03*nb3);
 
     float slope = 1.0f;
@@ -75,7 +79,7 @@ kernel void kernel_soft_max(
     }
 
     // parallel max
-    float lmax = -INFINITY;
+    float lmax = psrc2 ? psrc2[i02] : -INFINITY;
     for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
         lmax = fmax(lmax, psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f));
     }
@@ -91,7 +95,11 @@ kernel void kernel_soft_max(
         pdst[i00] = exp_psrc0;
     }
 
-    const float sum = sub_group_reduce_add(lsum);
+    float sum = sub_group_reduce_add(lsum);
+
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max);
+    }
 
     for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
         pdst[i00] /= sum;

From 42e295cca244ebd1433c29e3b24ffdb63bc9aee6 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= <johannesg@5d6.de>
Date: Fri, 8 Aug 2025 08:19:58 +0200
Subject: [PATCH 45/86] CUDA: attention sinks for mma FlashAttention
 (llama/15157)

---
 ggml/src/ggml-cuda/fattn-mma-f16.cuh | 87 ++++++++++++++++++++++------
 ggml/src/ggml-cuda/fattn.cu          |  2 +-
 ggml/src/ggml-cuda/ggml-cuda.cu      |  3 +-
 3 files changed, 73 insertions(+), 19 deletions(-)

diff --git a/ggml/src/ggml-cuda/fattn-mma-f16.cuh b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
index 37125384417..39731baaeb7 100644
--- a/ggml/src/ggml-cuda/fattn-mma-f16.cuh
+++ b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
@@ -785,6 +785,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         const half2  * const __restrict__ K_h2,
         const half2  * const __restrict__ V_h2,
         const half2  * const __restrict__ mask_h2,
+        const float  * const __restrict__ sinks_f,
         float2       * const __restrict__ dstk,
         float2       * const __restrict__ dstk_fixup,
         const float scale,
@@ -957,6 +958,52 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         }
     }
 
+    // If attention sinks are used, potentially re-scale if KQ_max is small.
+    // Also add the sink as a value to KQ_rowsum, this is done after synchonization of KQ_rowsum
+    //     so it's being done unconditionally for every thread.
+    if (!is_fixup && (np == 1 || threadIdx.y % np == 0) && sinks_f) {
+        float KQ_max_scale[cols_per_thread];
+#pragma unroll
+        for (int col = 0; col < cols_per_thread; ++col) {
+            static_assert(ntiles == 1 || ntiles == 2, "ntiles > 2 not implemented");
+            const int jc = ntiles == 1 ? 2*tile_C_VKQ::get_j(col/2) + col % 2 : tile_C_VKQ_16::get_i(col);
+            const float sink = sinks_f[jc % ncols2];
+
+            const float KQ_max_new = fmaxf(KQ_max[col], sink);
+            const float KQ_max_diff = KQ_max[col] - KQ_max_new;
+            KQ_max_scale[col] = expf(KQ_max_diff);
+            KQ_max[col] = KQ_max_new;
+
+            *((uint32_t *) &KQ_max_scale[col]) *= KQ_max_diff >= SOFTMAX_FTZ_THRESHOLD;
+
+            const float KQ_max_add = expf(sink - KQ_max_new);
+            KQ_rowsum[col] = KQ_max_scale[col]*KQ_rowsum[col] + KQ_max_add;
+        }
+
+        if (ntiles == 1) {
+            const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[0], KQ_max_scale[1]);
+#pragma unroll
+            for (int i = 0; i < DV/tile_C_VKQ::I; ++i) {
+#pragma unroll
+                for (int l = 0; l < tile_C_VKQ::ne; ++l) {
+                    VKQ_C[i].x[l] *= KQ_max_scale_h2;
+                }
+            }
+        } else {
+#pragma unroll
+            for (int col = 0; col < cols_per_thread; ++col) {
+                const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[col], KQ_max_scale[col]);
+#pragma unroll
+                for (int i = 0; i < DV/tile_C_VKQ_16::J; ++i) {
+#pragma unroll
+                    for (int l0 = 0; l0 < tile_C_VKQ_16::ne; l0 += 2) {
+                        VKQ_C_16[i*ntiles/2 + col/2].x[l0 + col % 2] *= KQ_max_scale_h2;
+                    }
+                }
+            }
+        }
+    }
+
     // Combine VKQ accumulator values if np > 1.
     // It's also faster to do small writes to shared memory, then large write to VRAM than to do small writes to VRAM.
     // So also write VKQ accumulators to shared memory in column-major format if np == 1.
@@ -1271,18 +1318,21 @@ static __global__ void flash_attn_ext_f16(
 
     while (kbc < kbc_stop && kb0_stop == iter_k) {
         const int sequence = kbc / (iter_k*iter_j*(ne02/ncols2));
-        const int head = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence) / (iter_k*iter_j);
-        const int jt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence - iter_k*iter_j*head) / iter_k; // j index of current tile.
+        const int zt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence) / (iter_k*iter_j); // head in units of ncols2
+        const int jt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence - iter_k*iter_j*zt) / iter_k; // j index of current tile.
 
-        const float2 * Q_f2    = (const float2 *) (Q + nb03*sequence + nb02*(head*ncols2));
-        const half2  * K_h2    = (const half2  *) (K + nb13*sequence + nb12*(head*ncols2 / gqa_ratio));
+        const int head0 = zt * ncols2;
+
+        const float2 * Q_f2    = (const float2 *) (Q + nb03*sequence + nb02* head0);
+        const half2  * K_h2    = (const half2  *) (K + nb13*sequence + nb12*(head0 / gqa_ratio));
         const half2  * mask_h2 = ncols2 == 1 && !mask ? nullptr :
             (const half2  *) (mask + nb33*(sequence % ne33) + nb31*jt*ncols1);
-        float2       * dstk    = ((float2 *) dst) + (sequence*ne01*ne02 + head*ncols2) * (DV/2);
+        float2       * dstk    = ((float2 *) dst) + (sequence*ne01*ne02 + head0) * (DV/2);
 
-        const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb23*sequence + nb22*(head*ncols2 / gqa_ratio));
+        const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb23*sequence + nb22*(head0 / gqa_ratio));
+        const float * sinks_f = sinks ? (const float *) sinks + head0 : nullptr;
 
-        const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, head, n_head_log2, m0, m1) : 1.0f;
+        const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, head0, n_head_log2, m0, m1) : 1.0f;
 
         const int kb0_start_kernel = kb0_start * kb_niter;
         int       kb0_stop_kernel  = kb0_stop  * kb_niter;
@@ -1295,12 +1345,12 @@ static __global__ void flash_attn_ext_f16(
         if (kb0_start == 0) {
             constexpr bool needs_fixup = false; // CUDA block is working on an entire tile.
             flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla, needs_fixup, is_fixup>
-                (Q_f2, K_h2, V_h2, mask_h2, dstk, dst_meta, scale, slope, logit_softcap,
+                (Q_f2, K_h2, V_h2, mask_h2, sinks_f, dstk, dst_meta, scale, slope, logit_softcap,
                  ne01, ne02, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel);
         } else {
             constexpr bool needs_fixup = true; // CUDA block is working on the beginning of a tile.
             flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla, needs_fixup, is_fixup>
-                (Q_f2, K_h2, V_h2, mask_h2, dstk, dst_meta, scale, slope, logit_softcap,
+                (Q_f2, K_h2, V_h2, mask_h2, sinks_f, dstk, dst_meta, scale, slope, logit_softcap,
                  ne01, ne02, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel);
         }
 
@@ -1316,18 +1366,21 @@ static __global__ void flash_attn_ext_f16(
     }
 
     const int sequence = kbc / (iter_k*iter_j*(ne02/ncols2));
-    const int head = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence) / (iter_k*iter_j);
-    const int jt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence - iter_k*iter_j*head) / iter_k; // j index of current tile.
+    const int zt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence) / (iter_k*iter_j); // head in units of ncols2
+    const int jt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence - iter_k*iter_j*zt) / iter_k; // j index of current tile.
+
+    const int head0 = zt * ncols2;
 
-    const float2 * Q_f2    = (const float2 *) (Q + nb03*sequence + nb02*(head*ncols2));
-    const half2  * K_h2    = (const half2  *) (K + nb13*sequence + nb12*(head*ncols2 / gqa_ratio));
+    const float2 * Q_f2    = (const float2 *) (Q + nb03*sequence + nb02* head0);
+    const half2  * K_h2    = (const half2  *) (K + nb13*sequence + nb12*(head0 / gqa_ratio));
     const half2  * mask_h2 = ncols2 == 1 && !mask ? nullptr :
         (const half2  *) (mask + nb33*(sequence % ne33) + nb31*jt*ncols1);
-    float2       * dstk    = ((float2 *) dst) + (sequence*ne01*ne02 + head*ncols2) * (DV/2);
+    float2       * dstk    = ((float2 *) dst) + (sequence*ne01*ne02 + head0) * (DV/2);
 
-    const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb23*sequence + nb22*(head*ncols2 / gqa_ratio));
+    const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb23*sequence + nb22*(head0 / gqa_ratio));
+    const float * sinks_f = sinks ? (const float *) sinks + head0 : nullptr;
 
-    const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, head, n_head_log2, m0, m1) : 1.0f;
+    const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, head0, n_head_log2, m0, m1) : 1.0f;
 
     const int kb0_start_kernel = kb0_start * kb_niter;
     int       kb0_stop_kernel  = kb0_stop  * kb_niter;
@@ -1339,7 +1392,7 @@ static __global__ void flash_attn_ext_f16(
     constexpr bool is_fixup = true; // Last index writes its data to fixup buffer to avoid data races with other blocks.
     constexpr bool needs_fixup = false;
     flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla, needs_fixup, is_fixup>
-        (Q_f2, K_h2, V_h2, mask_h2, dstk, dst_meta, scale, slope, logit_softcap,
+        (Q_f2, K_h2, V_h2, mask_h2, sinks_f, dstk, dst_meta, scale, slope, logit_softcap,
          ne01, ne02, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel);
 #else
     GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); GGML_UNUSED(sinks);
diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu
index 8ddd0415b7f..6c1185deac8 100644
--- a/ggml/src/ggml-cuda/fattn.cu
+++ b/ggml/src/ggml-cuda/fattn.cu
@@ -282,7 +282,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
     const enum ggml_prec prec = ggml_flash_attn_ext_get_prec(KQV);
 
     // TODO: currently only vec implementation for sinks is supported [TAG_ATTN_SINKS]
-    if (sinks) {
+    if (sinks && !fp16_mma_available(cc)) {
         if (prec == GGML_PREC_DEFAULT && fast_fp16_available(cc)) {
             ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
         } else {
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index ec7ab255188..19e9c405ea2 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -3532,7 +3532,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 return op->src[1]->ne[0] == 576 && op->src[2]->ne[0] == 512 && op->src[3] && gqa_ratio % 16 == 0;
             }
             // TODO: more general-purpose attention sink support [TAG_ATTN_SINKS]
-            if (op->src[4] && op->src[0]->ne[0] != 64 && op->src[0]->ne[0] != 128) { // currently only sinks for head_size 64 and 128 are supported
+            if (op->src[4] && !fp16_mma_available(ggml_cuda_info().devices[dev_ctx->device].cc)
+                    && op->src[0]->ne[0] != 64 && op->src[0]->ne[0] != 128) {
                 return false;
             }
             if (op->src[0]->ne[0] == 192) {

From 2dc56f88e41b228a40ee67a9cb1ad7bad9a7d830 Mon Sep 17 00:00:00 2001
From: AN Long <aisk@users.noreply.github.com>
Date: Fri, 8 Aug 2025 21:37:22 +0900
Subject: [PATCH 46/86] ggml : fix field name when new ggml_backend
 (llama/14944)

---
 ggml/src/ggml-blas/ggml-blas.cpp     | 8 ++++----
 ggml/src/ggml-cpu/ggml-cpu.cpp       | 8 ++++----
 ggml/src/ggml-cuda/ggml-cuda.cu      | 8 ++++----
 ggml/src/ggml-opencl/ggml-opencl.cpp | 8 ++++----
 ggml/src/ggml-rpc/ggml-rpc.cpp       | 8 ++++----
 ggml/src/ggml-sycl/ggml-sycl.cpp     | 8 ++++----
 ggml/src/ggml-vulkan/ggml-vulkan.cpp | 8 ++++----
 7 files changed, 28 insertions(+), 28 deletions(-)

diff --git a/ggml/src/ggml-blas/ggml-blas.cpp b/ggml/src/ggml-blas/ggml-blas.cpp
index ec158dfac6e..aeac2e57449 100644
--- a/ggml/src/ggml-blas/ggml-blas.cpp
+++ b/ggml/src/ggml-blas/ggml-blas.cpp
@@ -281,10 +281,10 @@ ggml_backend_t ggml_backend_blas_init(void) {
     ggml_backend_blas_context * ctx = new ggml_backend_blas_context;
 
     ggml_backend_t backend = new ggml_backend {
-        /* .guid      = */ ggml_backend_blas_guid(),
-        /* .interface = */ blas_backend_i,
-        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_blas_reg(), 0),
-        /* .context   = */ ctx,
+        /* .guid    = */ ggml_backend_blas_guid(),
+        /* .iface   = */ blas_backend_i,
+        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_blas_reg(), 0),
+        /* .context = */ ctx,
     };
 
 #if defined(OPENBLAS_VERSION) && defined(GGML_USE_OPENMP)
diff --git a/ggml/src/ggml-cpu/ggml-cpu.cpp b/ggml/src/ggml-cpu/ggml-cpu.cpp
index e16cdc9d498..8dacd36714b 100644
--- a/ggml/src/ggml-cpu/ggml-cpu.cpp
+++ b/ggml/src/ggml-cpu/ggml-cpu.cpp
@@ -214,10 +214,10 @@ ggml_backend_t ggml_backend_cpu_init(void) {
     ctx->abort_callback_data = NULL;
 
     ggml_backend_t cpu_backend = new ggml_backend {
-        /* .guid      = */ ggml_backend_cpu_guid(),
-        /* .interface = */ ggml_backend_cpu_i,
-        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
-        /* .context   = */ ctx,
+        /* .guid    = */ ggml_backend_cpu_guid(),
+        /* .iface   = */ ggml_backend_cpu_i,
+        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .context = */ ctx,
     };
 
     if (cpu_backend == NULL) {
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index 19e9c405ea2..d9110491ec7 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -3799,10 +3799,10 @@ ggml_backend_t ggml_backend_cuda_init(int device) {
     }
 
     ggml_backend_t cuda_backend = new ggml_backend {
-        /* .guid      = */ ggml_backend_cuda_guid(),
-        /* .interface = */ ggml_backend_cuda_interface,
-        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cuda_reg(), device),
-        /* .context   = */ ctx,
+        /* .guid    = */ ggml_backend_cuda_guid(),
+        /* .iface   = */ ggml_backend_cuda_interface,
+        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cuda_reg(), device),
+        /* .context = */ ctx,
     };
 
     return cuda_backend;
diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
index b32d5da3075..8ba1e00df7b 100644
--- a/ggml/src/ggml-opencl/ggml-opencl.cpp
+++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
@@ -2624,10 +2624,10 @@ ggml_backend_t ggml_backend_opencl_init(void) {
     ggml_backend_opencl_context *backend_ctx = ggml_cl2_init(dev);
 
     ggml_backend_t backend = new ggml_backend {
-        /* .guid      = */ ggml_backend_opencl_guid(),
-        /* .interface = */ ggml_backend_opencl_i,
-        /* .device    = */ dev,
-        /* .context   = */ backend_ctx
+        /* .guid    = */ ggml_backend_opencl_guid(),
+        /* .iface   = */ ggml_backend_opencl_i,
+        /* .device  = */ dev,
+        /* .context = */ backend_ctx
     };
 
     return backend;
diff --git a/ggml/src/ggml-rpc/ggml-rpc.cpp b/ggml/src/ggml-rpc/ggml-rpc.cpp
index 29bc421d58f..df6ba54076d 100644
--- a/ggml/src/ggml-rpc/ggml-rpc.cpp
+++ b/ggml/src/ggml-rpc/ggml-rpc.cpp
@@ -823,10 +823,10 @@ ggml_backend_t ggml_backend_rpc_init(const char * endpoint) {
     };
 
     ggml_backend_t backend = new ggml_backend {
-        /* .guid      = */ ggml_backend_rpc_guid(),
-        /* .interface = */ ggml_backend_rpc_interface,
-        /* .device    = */ ggml_backend_rpc_add_device(endpoint),
-        /* .context   = */ ctx
+        /* .guid    = */ ggml_backend_rpc_guid(),
+        /* .iface   = */ ggml_backend_rpc_interface,
+        /* .device  = */ ggml_backend_rpc_add_device(endpoint),
+        /* .context = */ ctx
     };
     return backend;
 }
diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
index 6fa27418ce8..3992dad01d5 100644
--- a/ggml/src/ggml-sycl/ggml-sycl.cpp
+++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
@@ -4586,10 +4586,10 @@ ggml_backend_t ggml_backend_sycl_init(int device) {
     };
 
     ggml_backend_t sycl_backend = new ggml_backend {
-        /* .guid      = */ ggml_backend_sycl_guid(),
-        /* .interface = */ ggml_backend_sycl_interface,
-        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_sycl_reg(), device),
-        /* .context   = */ ctx
+        /* .guid    = */ ggml_backend_sycl_guid(),
+        /* .iface   = */ ggml_backend_sycl_interface,
+        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_sycl_reg(), device),
+        /* .context = */ ctx
     };
 
     return sycl_backend;
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index b1cbbc9866c..4070e248baa 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -10767,10 +10767,10 @@ ggml_backend_t ggml_backend_vk_init(size_t dev_num) {
     ggml_vk_init(ctx, dev_num);
 
     ggml_backend_t vk_backend = new ggml_backend {
-        /* .guid      = */ ggml_backend_vk_guid(),
-        /* .interface = */ ggml_backend_vk_interface,
-        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_vk_reg(), dev_num),
-        /* .context   = */ ctx,
+        /* .guid    = */ ggml_backend_vk_guid(),
+        /* .iface   = */ ggml_backend_vk_interface,
+        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_vk_reg(), dev_num),
+        /* .context = */ ctx,
     };
 
     return vk_backend;

From 430cbcc3db2728405a689f0e000a881b148c4a91 Mon Sep 17 00:00:00 2001
From: compilade <git@compilade.net>
Date: Fri, 8 Aug 2025 17:48:26 -0400
Subject: [PATCH 47/86] gguf-py : add Numpy MXFP4 de/quantization support
 (llama/15111)

* gguf-py : add MXFP4 de/quantization support

* ggml-quants : handle zero amax for MXFP4
---
 ggml/src/ggml-quants.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ggml/src/ggml-quants.c b/ggml/src/ggml-quants.c
index a57d2a16d6c..94f6405ca1e 100644
--- a/ggml/src/ggml-quants.c
+++ b/ggml/src/ggml-quants.c
@@ -288,7 +288,7 @@ void quantize_row_mxfp4_ref(const float * GGML_RESTRICT x, block_mxfp4 * GGML_RE
             }
         }
 
-        const uint8_t e = (uint8_t) (floorf(log2f(amax)) - 2 + 127);
+        const uint8_t e = amax > 0.0f ? (uint8_t) (floorf(log2f(amax)) - 2 + 127) : 0;
 
         const float d = GGML_E8M0_TO_FP32_HALF(e);
 

From 46e8145af8431477e9dd46f449a093f876c43792 Mon Sep 17 00:00:00 2001
From: Aman Gupta <amangupta052@gmail.com>
Date: Sat, 9 Aug 2025 20:00:24 +0800
Subject: [PATCH 48/86] CUDA: add attention sinks for tile and wmma
 (llama/15178)

* CUDA: add attention sinks for tile and wmma

* Review: formatting changes + remove syncthreads from tile + remove warp_reduce_max from wmma
---
 ggml/src/ggml-cuda/fattn-tile-f16.cu | 34 ++++++++++++++--
 ggml/src/ggml-cuda/fattn-tile-f32.cu | 36 +++++++++++++++--
 ggml/src/ggml-cuda/fattn-wmma-f16.cu | 58 +++++++++++++++++++++++++---
 ggml/src/ggml-cuda/fattn.cu          | 11 ------
 4 files changed, 115 insertions(+), 24 deletions(-)

diff --git a/ggml/src/ggml-cuda/fattn-tile-f16.cu b/ggml/src/ggml-cuda/fattn-tile-f16.cu
index 0fcfaa32ea4..1e23f8f79c2 100644
--- a/ggml/src/ggml-cuda/fattn-tile-f16.cu
+++ b/ggml/src/ggml-cuda/fattn-tile-f16.cu
@@ -49,10 +49,11 @@ static __global__ void flash_attn_tile_ext_f16(
     const int sequence = blockIdx.z / ne02;
     const int head = blockIdx.z - sequence*ne02;
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float2 * Q_f2  = (const float2 *) (Q    + nb03* sequence         + nb02* head              + nb01*ic0);
-    const half2  * K_h2  = (const half2  *) (K    + nb13* sequence         + nb12*(head / gqa_ratio));
-    const half2  * V_h2  = (const half2  *) (V    + nb13* sequence         + nb12*(head / gqa_ratio)); // K and V have same shape
-    const half   * maskh = (const half   *) (mask + nb33*(sequence % ne33)                           + nb31*ic0);
+    const float2 * Q_f2   = (const float2 *) (Q    + nb03* sequence         + nb02* head              + nb01*ic0);
+    const half2  * K_h2   = (const half2  *) (K    + nb13* sequence         + nb12*(head / gqa_ratio));
+    const half2  * V_h2   = (const half2  *) (V    + nb13* sequence         + nb12*(head / gqa_ratio)); // K and V have same shape
+    const half   * maskh  = (const half   *) (mask  + nb33*(sequence % ne33)                          + nb31*ic0);
+    const float  * sinksf = (const float  *) (sinks);
 
     const int stride_KV2 = nb11 / sizeof(half2);
 
@@ -242,6 +243,31 @@ static __global__ void flash_attn_tile_ext_f16(
         __syncthreads();
     }
 
+    //Attention sink: adjust running max and sum once per head
+    if (sinksf && blockIdx.y == 0) {
+        const half sink = __float2half(sinksf[head]);
+
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            half kqmax_new_j = fmaxf(kqmax[j0/nwarps], sink);
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+
+            const half2 KQ_max_scale = __half2half2(hexp(kqmax[j0/nwarps] - kqmax_new_j));
+            kqmax[j0/nwarps] = kqmax_new_j;
+
+            const half val = hexp(sink - kqmax[j0/nwarps]);
+            kqsum[j0/nwarps] = kqsum[j0/nwarps] * KQ_max_scale;
+            if (threadIdx.x == 0) {
+                kqsum[j0/nwarps].x = __hadd(kqsum[j0/nwarps].x, val);
+            }
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                VKQ[j0/nwarps][i0/WARP_SIZE] *= KQ_max_scale;
+            }
+        }
+    }
+
     float2 * dst2 = (float2 *) dst;
 
 #pragma unroll
diff --git a/ggml/src/ggml-cuda/fattn-tile-f32.cu b/ggml/src/ggml-cuda/fattn-tile-f32.cu
index 23550cbbd97..c58194937d7 100644
--- a/ggml/src/ggml-cuda/fattn-tile-f32.cu
+++ b/ggml/src/ggml-cuda/fattn-tile-f32.cu
@@ -60,10 +60,11 @@ static __global__ void flash_attn_tile_ext_f32(
     const int sequence = blockIdx.z / ne02;
     const int head = blockIdx.z - sequence*ne02;
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float2 * Q_f2  = (const float2 *) (Q    + nb03* sequence         + nb02* head              + nb01*ic0);
-    const half2  * K_h2  = (const half2  *) (K    + nb13* sequence         + nb12*(head / gqa_ratio));
-    const half2  * V_h2  = (const half2  *) (V    + nb13* sequence         + nb12*(head / gqa_ratio)); // K and V have same shape
-    const half   * maskh = (const half   *) (mask + nb33*(sequence % ne33)                           + nb31*ic0);
+    const float2 * Q_f2   = (const float2 *) (Q    + nb03* sequence         + nb02* head              + nb01*ic0);
+    const half2  * K_h2   = (const half2  *) (K    + nb13* sequence         + nb12*(head / gqa_ratio));
+    const half2  * V_h2   = (const half2  *) (V    + nb13* sequence         + nb12*(head / gqa_ratio)); // K and V have same shape
+    const half   * maskh  = (const half   *) (mask  + nb33*(sequence % ne33)                          + nb31*ic0);
+    const float  * sinksf = (const float  *) (sinks);
 
     const int stride_KV2 = nb11 / sizeof(half2);
 
@@ -252,6 +253,33 @@ static __global__ void flash_attn_tile_ext_f32(
         __syncthreads();
     }
 
+
+    //Attention sink: adjust running max and sum once per head
+    if (sinksf && blockIdx.y == 0) {
+        const float sink = sinksf[head];
+
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            float kqmax_new_j = fmaxf(kqmax[j0/nwarps], sink);
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+
+            const float KQ_max_scale = expf(kqmax[j0/nwarps] - kqmax_new_j);
+            kqmax[j0/nwarps] = kqmax_new_j;
+
+            const float val = expf(sink - kqmax[j0/nwarps]);
+            kqsum[j0/nwarps] = kqsum[j0/nwarps] * KQ_max_scale;
+            if (threadIdx.x == 0) {
+                kqsum[j0/nwarps] += val;
+            }
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                VKQ[j0/nwarps][i0/WARP_SIZE].x *= KQ_max_scale;
+                VKQ[j0/nwarps][i0/WARP_SIZE].y *= KQ_max_scale;
+            }
+        }
+    }
+
     float2 * dst2 = (float2 *) dst;
 
 #pragma unroll
diff --git a/ggml/src/ggml-cuda/fattn-wmma-f16.cu b/ggml/src/ggml-cuda/fattn-wmma-f16.cu
index 93d4d810218..fdc4d17da2d 100644
--- a/ggml/src/ggml-cuda/fattn-wmma-f16.cu
+++ b/ggml/src/ggml-cuda/fattn-wmma-f16.cu
@@ -82,11 +82,12 @@ static __global__ void flash_attn_ext_f16(
     const int sequence = blockIdx.z / ne02;
     const int head = blockIdx.z - sequence*ne02;
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float * Q_f   = (const float *) (Q    + nb03* sequence         + nb02* head              + nb01*ic0);
-    const half  * K_h   = (const half  *) (K    + nb13* sequence         + nb12*(head / gqa_ratio));
-    const half  * V_h   = (const half  *) (V    + nb13* sequence         + nb12*(head / gqa_ratio)); // K and V have same shape
-    const half  * maskh = (const half  *) (mask + nb33*(sequence % ne33)                           + nb31*ic0);
-    const half2 * mask2 = (const half2 *)  maskh;
+    const float * Q_f    = (const float *) (Q    + nb03* sequence         + nb02* head              + nb01*ic0);
+    const half  * K_h    = (const half  *) (K    + nb13* sequence         + nb12*(head / gqa_ratio));
+    const half  * V_h    = (const half  *) (V    + nb13* sequence         + nb12*(head / gqa_ratio)); // K and V have same shape
+    const half  * maskh  = (const half  *) (mask + nb33*(sequence % ne33)                           + nb31*ic0);
+    const half2 * mask2  = (const half2 *)  maskh;
+    const float * sinksf = (const float *) sinks;
 
     const int stride_Q  = nb01 / sizeof(float);
     const int stride_KV = nb11 / sizeof(half);
@@ -381,6 +382,53 @@ static __global__ void flash_attn_ext_f16(
         __syncthreads();
     }
 
+    // Apply attention sinks
+    if (sinksf && blockIdx.y == 0) {
+        const float sinkf = sinksf[head];
+        const half  sinkh = __float2half(sinkf);
+
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            if (std::is_same<KQ_acc_t, float>::value) {
+                float kqmax_new = fmaxf(KQ_max_f[j0/nwarps], sinkf);
+
+                const float KQ_max_scale = expf(KQ_max_f[j0/nwarps] - kqmax_new);
+                KQ_max_f[j0/nwarps] = kqmax_new;
+
+                KQ_rowsum_f[j0/nwarps] = KQ_rowsum_f[j0/nwarps] * KQ_max_scale + expf(sinkf - KQ_max_f[j0/nwarps]);
+
+                const half2 scale_h2 = make_half2(KQ_max_scale, KQ_max_scale);
+#pragma unroll
+                for (int i0 = 0; i0 < D/2; i0 += warp_size) {
+                    const int i = i0 + threadIdx.x;
+                    if (i0 + warp_size > D/2 && i >= D/2) break;
+                    VKQ2[j*(D_padded/2) + i] *= scale_h2;
+                }
+            } else {
+                half kqmax_old = __low2half(KQ_max_h2[j0/nwarps]);
+                half kqmax_new = fmaxf(kqmax_old, sinkh);
+                KQ_max_h2[j0/nwarps] = __half2half2(kqmax_new);
+
+                const half  KQ_max_scale_h = hexp(kqmax_old - kqmax_new);
+                const half2 KQ_max_scale   = __half2half2(KQ_max_scale_h);
+
+                KQ_rowsum_h2[j0/nwarps] = KQ_rowsum_h2[j0/nwarps] * KQ_max_scale;
+                const half val = hexp(sinkh - kqmax_new);
+                KQ_rowsum_h2[j0/nwarps].x = __hadd(KQ_rowsum_h2[j0/nwarps].x, val);
+
+#pragma unroll
+                for (int i0 = 0; i0 < D/2; i0 += warp_size) {
+                    const int i = i0 + threadIdx.x;
+                    if (i0 + warp_size > D/2 && i >= D/2) break;
+                    VKQ2[j*(D_padded/2) + i] *= KQ_max_scale;
+                }
+            }
+        }
+
+        __syncthreads();
+    }
 #pragma unroll
     for (int j0 = 0; j0 < ncols; j0 += nwarps) {
         const int j_VKQ = j0 + threadIdx.y;
diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu
index 6c1185deac8..22e90d0e7b3 100644
--- a/ggml/src/ggml-cuda/fattn.cu
+++ b/ggml/src/ggml-cuda/fattn.cu
@@ -274,23 +274,12 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
     const ggml_tensor * K     = dst->src[1];
     const ggml_tensor * V     = dst->src[2];
     const ggml_tensor * mask  = dst->src[3];
-    const ggml_tensor * sinks = dst->src[4];
 
     ggml_cuda_set_device(ctx.device);
     const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
     const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size;
     const enum ggml_prec prec = ggml_flash_attn_ext_get_prec(KQV);
 
-    // TODO: currently only vec implementation for sinks is supported [TAG_ATTN_SINKS]
-    if (sinks && !fp16_mma_available(cc)) {
-        if (prec == GGML_PREC_DEFAULT && fast_fp16_available(cc)) {
-            ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
-        } else {
-            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
-        }
-        return;
-    }
-
 #if defined(GGML_HIP_ROCWMMA_FATTN)
     if (GGML_CUDA_CC_IS_AMD(cc) && fp16_mma_available(cc)) {
         ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);

From b65235d696161e205fbaea423cb10ce211a8bd8e Mon Sep 17 00:00:00 2001
From: David Zhao <90013954+Your-Cheese@users.noreply.github.com>
Date: Sat, 9 Aug 2025 13:29:43 -0500
Subject: [PATCH 49/86] cuda: refactored ssm_scan and use CUB (llama/13291)

* cuda: refactored ssm_scan to use CUB

* fixed compilation error when when not using CUB

* assign L to constant and use size_t instead of int

* deduplicated functions

* change min blocks per mp to 1

* Use cub load and store warp transpose

* suppress clang warning
---
 ggml/src/ggml-cuda/ssm-scan.cu | 222 ++++++++++++++++++++++-----------
 1 file changed, 152 insertions(+), 70 deletions(-)

diff --git a/ggml/src/ggml-cuda/ssm-scan.cu b/ggml/src/ggml-cuda/ssm-scan.cu
index c9184398b42..dc9a7d58d05 100644
--- a/ggml/src/ggml-cuda/ssm-scan.cu
+++ b/ggml/src/ggml-cuda/ssm-scan.cu
@@ -1,87 +1,117 @@
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
+#define USE_CUB
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
+
+#ifdef USE_CUB
+#include <cub/cub.cuh>
+using namespace cub;
+#endif // USE_CUB
+
 #include "ssm-scan.cuh"
 
-template <size_t splitD, size_t N>
-__global__ void __launch_bounds__(splitD, 2)
-    ssm_scan_f32(const float * __restrict__ src0, const float * __restrict__ src1, const float * __restrict__ src2,
-                 const float * __restrict__ src3, const float * __restrict__ src4, const float * __restrict__ src5,
+// We would like to keep pragma unroll for cases where L_template is not 0,
+// so we suppress the clang transformation warning.
+#ifdef __clang__
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wpass-failed"
+#endif // __clang__
+template <size_t splitD, size_t N, size_t L_template>
+__global__ void __launch_bounds__(splitD, 1)
+    ssm_scan_f32(const float *__restrict__ src0, const float *__restrict__ src1, const float *__restrict__ src2,
+                 const float *__restrict__ src3, const float *__restrict__ src4, const float *__restrict__ src5,
                  const int32_t * __restrict__ src6, float * __restrict__ dst,
                  const int src0_nb2, const int src0_nb3, const int src1_nb2, const int src1_nb3,
                  const int src2_nb1, const int src2_nb2, const int src3_nb1,
                  const int src4_nb2, const int src4_nb3, const int src5_nb2, const int src5_nb3,
-                 const int64_t s_off, const int64_t d_inner, const int64_t L) {
-
-    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
-    const int bidx = blockIdx.x;  // split along B (sequences)
-    const int bidy = blockIdx.y;  // split along D (d_inner)
-    const int tid  = threadIdx.x;
-    const int wid  = tid / 32;
-    const int wtid = tid % 32;
-
-    extern __shared__ float smem[];
-    const int               stride_sA  = N + 1;
-    const int               stride_ss0 = N + 1;
-    float *                 smem_A     = smem;
-    float *                 smem_s0    = smem_A + splitD * stride_sA;
-
-    const float * s0_block = (const float *) ((const char *) src0 + src6[bidx] * src0_nb3 + bidy * splitD * src0_nb2);
-    const float * x_block  = (const float *) ((const char *) src1 + (bidx * src1_nb3) + bidy * splitD * sizeof(float));
-    const float * dt_block = (const float *) ((const char *) src2 + (bidx * src2_nb2) + bidy * splitD * sizeof(float));
-    const float * A_block  = (const float *) ((const char *) src3 + bidy * splitD * src3_nb1);
-    const float * B_block  = (const float *) ((const char *) src4 + (bidx * src4_nb3));
-    const float * C_block  = (const float *) ((const char *) src5 + (bidx * src5_nb3));
-    float *       y_block  = (float *) ((char *) dst + (bidx * d_inner * L * sizeof(float)) + bidy * splitD * sizeof(float));
-    float *       s_block  = (float *) ((char *) dst + s_off + bidx * src0_nb3 + bidy * splitD * src0_nb2);
-
-    const int stride_s0 = src0_nb2 / sizeof(float);
-    const int stride_x  = src1_nb2 / sizeof(float);
+                 const int64_t s_off, const int64_t d_inner, const int64_t L_param)
+{
+    const size_t L = L_template == 0 ? L_param : L_template;
+    const float *s0_block = (const float *)((const char *)src0 + src6[blockIdx.x] * src0_nb3 + blockIdx.y * splitD * src0_nb2);
+    const float *x_block = (const float *)((const char *)src1 + (blockIdx.x * src1_nb3) + blockIdx.y * splitD * sizeof(float));
+    const float *dt_block = (const float *)((const char *)src2 + (blockIdx.x * src2_nb2) + blockIdx.y * splitD * sizeof(float));
+    const float *A_block = (const float *)((const char *)src3 + blockIdx.y * splitD * src3_nb1);
+    const float *B_block = (const float *)((const char *)src4 + (blockIdx.x * src4_nb3));
+    const float *C_block = (const float *)((const char *)src5 + (blockIdx.x * src5_nb3));
+    float *y_block = (float *)((char *)dst + (blockIdx.x * d_inner * L * sizeof(float)) + blockIdx.y * splitD * sizeof(float));
+    float *s_block = (float *)((char *)dst + s_off + blockIdx.x * src0_nb3 + blockIdx.y * splitD * src0_nb2);
+
+    const int stride_x = src1_nb2 / sizeof(float);
     const int stride_dt = src2_nb1 / sizeof(float);
-    const int stride_A  = src3_nb1 / sizeof(float);
-    const int stride_B  = src4_nb2 / sizeof(float);
-    const int stride_C  = src5_nb2 / sizeof(float);
-    const int stride_s  = stride_s0;
-    const int stride_y  = d_inner;
+    const int stride_B = src4_nb2 / sizeof(float);
+    const int stride_C = src5_nb2 / sizeof(float);
+    const int stride_y = d_inner;
 
-    // can N not be 16? for example 32?
-    if (N == 16) {
-#pragma unroll
-        for (size_t i = 0; i < splitD / 4; i += 2) {
-            float value = A_block[(wid * warp_size + i) * stride_A + wtid];
-            // todo: bank conflict
-            // I am always confused with how to use the swizzling method to solve
-            // bank conflit. Hoping somebody can tell me.
-            smem_A[(wid * warp_size + i) * stride_sA + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
-        }
+    float regA[N];
+    float regs0[N];
+
+    __shared__ float smemB[N];
+    __shared__ float smemC[N];
+
+#ifdef USE_CUB
+    using BlockLoad = cub::BlockLoad<float, splitD, N, cub::BLOCK_LOAD_WARP_TRANSPOSE>;
+    using BlockStore = cub::BlockStore<float, splitD, N, cub::BLOCK_STORE_WARP_TRANSPOSE>;
+
+    union CubTempStorage {
+        typename BlockLoad::TempStorage load_temp;
+        typename BlockStore::TempStorage store_temp;
+    };
+    __shared__ CubTempStorage cub_temp_storage;
+
+    BlockLoad(cub_temp_storage.load_temp).Load(A_block, regA);
+    BlockLoad(cub_temp_storage.load_temp).Load(s0_block, regs0);
+#else
+    const int stride_s0 = src0_nb2 / sizeof(float);
+    const int stride_A = src3_nb1 / sizeof(float);
 #pragma unroll
-        for (size_t i = 0; i < splitD / 4; i += 2) {
-            float value = s0_block[(wid * warp_size + i) * stride_s0 + wtid];
-            smem_s0[(wid * warp_size + i) * stride_ss0 + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
-        }
+    for (size_t n = 0; n < N; ++n)
+    {
+        regA[n] = A_block[threadIdx.x * stride_A + n];
+        regs0[n] = s0_block[threadIdx.x * stride_s0 + n];
     }
+#endif
 
-    __syncthreads();
+#pragma unroll
+    for (size_t i = 0; i < L; i++)
+    {
+        if (threadIdx.x < N)
+        {
+            smemB[threadIdx.x] = B_block[i * stride_B + threadIdx.x];
+            smemC[threadIdx.x] = C_block[i * stride_C + threadIdx.x];
+        }
+        __syncthreads();
 
-    for (int64_t i = 0; i < L; i++) {
-        float dt_soft_plus = dt_block[i * stride_dt + tid];
-        if (dt_soft_plus <= 20.0f) {
-            dt_soft_plus = log1pf(exp(dt_soft_plus));
+        float dt_soft_plus = dt_block[i * stride_dt + threadIdx.x];
+        if (dt_soft_plus <= 20.0f)
+        {
+            dt_soft_plus = log1pf(expf(dt_soft_plus));
         }
-        float x_dt = x_block[i * stride_x + tid] * dt_soft_plus;
+        float x_dt = x_block[i * stride_x + threadIdx.x] * dt_soft_plus;
+
         float sumf = 0.0f;
 #pragma unroll
-        for (size_t j = 0; j < N; j++) {
-            float state = (smem_s0[tid * stride_ss0 + j] * expf(dt_soft_plus * smem_A[tid * stride_sA + j])) +
-                          (B_block[i * stride_B + j] * x_dt);
-            sumf += state * C_block[i * stride_C + j];
-            if (i == L - 1) {
-                s_block[tid * stride_s + j] = state;
-            } else {
-                smem_s0[tid * stride_ss0 + j] = state;
-            }
+        for (size_t n = 0; n < N; n++)
+        {
+            float state = regs0[n] * expf(dt_soft_plus * regA[n]) + smemB[n] * x_dt;
+            sumf += state * smemC[n];
+            regs0[n] = state;
         }
-        __syncthreads();
-        y_block[i * stride_y + tid] = sumf;
+        y_block[i * stride_y + threadIdx.x] = sumf;
     }
+
+#ifdef USE_CUB
+    BlockStore(cub_temp_storage.store_temp).Store(s_block, regs0);
+#else
+    const int stride_s = stride_s0;
+#pragma unroll
+    for (size_t n = 0; n < N; ++n)
+    {
+        s_block[threadIdx.x * stride_s + n] = regs0[n];
+    }
+#endif
 }
+#ifdef __clang__
+#pragma clang diagnostic pop
+#endif // __clang__
 
 // assumes as many threads as d_state
 template <int splitH, int d_state>
@@ -201,11 +231,11 @@ static void ssm_scan_f32_cuda(const float * src0, const float * src1, const floa
                               const int src5_nb3, const int64_t s_off, const int64_t d_state, const int64_t head_dim,
                               const int64_t n_head, const int64_t n_group, const int64_t n_tok, const int64_t n_seq,
                               cudaStream_t stream) {
+    const int threads = 128;
     // NOTE: if you change conditions here, be sure to update the corresponding supports_op condition!
     if (src3_nb1 == sizeof(float)) {
         // Mamba-2
         if (d_state == 128) {
-            const int threads = 128;
             GGML_ASSERT(d_state % threads == 0);
             // NOTE: can be any power of two between 4 and 64
             const int splitH = 16;
@@ -229,7 +259,6 @@ static void ssm_scan_f32_cuda(const float * src0, const float * src1, const floa
             GGML_ABORT("doesn't support d_state!=(128 or 256).");
         }
     } else {
-        const int threads = 128;
         // Mamba-1
         GGML_ASSERT(n_head % threads == 0);
         GGML_ASSERT(head_dim == 1);
@@ -237,10 +266,63 @@ static void ssm_scan_f32_cuda(const float * src0, const float * src1, const floa
         const dim3 blocks(n_seq, (n_head + threads - 1) / threads, 1);
         const int  smem_size = (threads * (d_state + 1) * 2) * sizeof(float);
         if (d_state == 16) {
-            ssm_scan_f32<128, 16><<<blocks, threads, smem_size, stream>>>(
-                src0, src1, src2, src3, src4, src5, src6, dst,
+            switch (n_tok)
+            {
+            case 1:
+                ssm_scan_f32<threads, 16, 1><<<blocks, threads, smem_size, stream>>>(
+                    src0, src1, src2, src3, src4, src5, src6, dst,
+                src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
+                src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
+                break;
+            case 2:
+                ssm_scan_f32<threads, 16, 2><<<blocks, threads, smem_size, stream>>>(
+                    src0, src1, src2, src3, src4, src5, src6, dst,
+                src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
+                src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
+                break;
+            case 3:
+                ssm_scan_f32<threads, 16, 3><<<blocks, threads, smem_size, stream>>>(
+                    src0, src1, src2, src3, src4, src5, src6, dst,
+                src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
+                src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
+                break;
+            case 4:
+                ssm_scan_f32<threads, 16, 4><<<blocks, threads, smem_size, stream>>>(
+                    src0, src1, src2, src3, src4, src5, src6, dst,
+                src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
+                src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
+                break;
+            case 5:
+                ssm_scan_f32<threads, 16, 5><<<blocks, threads, smem_size, stream>>>(
+                    src0, src1, src2, src3, src4, src5, src6, dst,
                 src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
                 src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
+                break;
+            case 6:
+                ssm_scan_f32<threads, 16, 6><<<blocks, threads, smem_size, stream>>>(
+                    src0, src1, src2, src3, src4, src5, src6, dst,
+                src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
+                src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
+                break;
+            case 7:
+                ssm_scan_f32<threads, 16, 7><<<blocks, threads, smem_size, stream>>>(
+                    src0, src1, src2, src3, src4, src5, src6, dst,
+                src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
+                src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
+                break;
+            case 8:
+                ssm_scan_f32<threads, 16, 8><<<blocks, threads, smem_size, stream>>>(
+                    src0, src1, src2, src3, src4, src5, src6, dst,
+                src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
+                src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
+                break;
+            default:
+                ssm_scan_f32<threads, 16, 0><<<blocks, threads, smem_size, stream>>>(
+                    src0, src1, src2, src3, src4, src5, src6, dst,
+                src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
+                src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
+                break;
+            }
         } else {
             GGML_ABORT("doesn't support d_state!=16.");
         }

From 22294624102ca4e1c543998e0b4dfd25457d83ef Mon Sep 17 00:00:00 2001
From: Charles Xu <charles.xu@arm.com>
Date: Mon, 11 Aug 2025 09:59:26 +0200
Subject: [PATCH 50/86] kleidiai: fix unsigned overflow bug (llama/15150)

* kleidiai: fix unsigned overflow bug

* address review comments
---
 ggml/src/ggml-cpu/kleidiai/kleidiai.cpp | 23 ++++++++++++++++-------
 1 file changed, 16 insertions(+), 7 deletions(-)

diff --git a/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp b/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp
index 3a513a55d76..dff8fa244a1 100644
--- a/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp
+++ b/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp
@@ -259,7 +259,10 @@ class tensor_traits : public ggml::cpu::tensor_traits {
                 const int64_t m_start      = 0;
 
                 const int64_t n_step      = static_cast<int64_t>(kernel->get_n_step());
-                const int64_t num_threads = KAI_MIN(n / n_step, nth);
+                int64_t num_threads       = KAI_MIN(n / n_step, nth);
+                if (num_threads <= 0) {
+                    num_threads = 1;
+                }
 
                 if (ith < num_threads) {
                     const int64_t num_n_per_thread0   = round_down(n / num_threads, n_step);
@@ -309,7 +312,8 @@ class tensor_traits : public ggml::cpu::tensor_traits {
         GGML_ASSERT(kernel);
 
         const int ith = params->ith;
-        const int nth = params->nth;
+        const int nth_raw = params->nth;
+        const int nth = nth_raw > 0 ? nth_raw : 1;
 
         const size_t k = ne00;
         const size_t m = ne11;
@@ -327,9 +331,12 @@ class tensor_traits : public ggml::cpu::tensor_traits {
         const size_t num_n_per_thread = kai_roundup(kai_roundup(n, nth) / nth, n_step);
         const size_t n_start = ith * num_n_per_thread;
 
-        size_t n_to_process = num_n_per_thread;
-        if ((n_start + n_to_process) > n) {
-            n_to_process = n - n_start;
+        size_t n_to_process = 0;
+        if (n_start < n) {
+            n_to_process = num_n_per_thread;
+            if ((n_start + n_to_process) > n) {
+                n_to_process = n - n_start;
+            }
         }
 
         // Calculate number of columns to be processed per thread
@@ -361,8 +368,10 @@ class tensor_traits : public ggml::cpu::tensor_traits {
         const void* lhs_ptr            = (const void*)((const char *)lhs_packed + lhs_packed_offset);
         float *dst_ptr                 = reinterpret_cast<float *>(static_cast<uint8_t *>(dst->data) + dst_offset);
 
-        variant_call<void>(kernel->run_kernel, m, n_to_process, k, QK4_0, lhs_ptr, rhs_ptr, dst_ptr, dst_stride,
-                           sizeof(float), -FLT_MAX, FLT_MAX);
+        if (n_to_process > 0) {
+            variant_call<void>(kernel->run_kernel, m, n_to_process, k, QK4_0, lhs_ptr, rhs_ptr, dst_ptr, dst_stride,
+                               sizeof(float), -FLT_MAX, FLT_MAX);
+        }
 
         return true;
     }

From 77c05f04071059d45fd7cce84b76bc99fda6f275 Mon Sep 17 00:00:00 2001
From: hipudding <huafengchun@gmail.com>
Date: Mon, 11 Aug 2025 22:50:31 +0800
Subject: [PATCH 51/86] CANN: Add broadcast for softmax and FA (llama/15208)

* refactor softmax

* fix fa

* fix mask shape

* format

* add comments

* Remove whitespace
---
 ggml/src/ggml-cann/aclnn_ops.cpp | 539 ++++++++++++-------------------
 ggml/src/ggml-cann/ggml-cann.cpp |  14 +-
 2 files changed, 213 insertions(+), 340 deletions(-)

diff --git a/ggml/src/ggml-cann/aclnn_ops.cpp b/ggml/src/ggml-cann/aclnn_ops.cpp
index 07d6b8b67d4..0b409ce87d2 100755
--- a/ggml/src/ggml-cann/aclnn_ops.cpp
+++ b/ggml/src/ggml-cann/aclnn_ops.cpp
@@ -812,7 +812,7 @@ void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
             ggml_cann_release_resources(ctx, src_trans_tensor);
             return;
         } else {
-            GGML_ABORT("Unsupport dst is not tontiguous.");
+            GGML_ABORT("Unsupport dst is not contiguous.");
         }
     }
     ggml_cann_release_resources(ctx, acl_src, acl_dst);
@@ -1330,160 +1330,196 @@ static void aclnn_pow_tensor_tensor(ggml_backend_cann_context& ctx,
 }
 
 /**
- * @brief   Applies the Alibi (Attention with Linear Biases) mechanism to the
- * @details This function implements the Alibi mechanism, which introduces
- *          learnable biases into the attention scores to simulate relative
- *          position encoding without the need for explicit positional
- *          embeddings.
- *
- * @param ctx          The backend CANN context for executing operations.
- * @param acl_src      The source tensor representing the query or key.
- * @param acl_position The position tensor containing relative positions.
- * @param acl_dst      The destination tensor where the result will be stored.
- * @param n_head       The number of attention heads.
- * @param src_ne       The dimensions of the source tensor.
- * @param src_nb0      The byte size of the first dimension of the source
- tensor.
- * @param max_bias     The maximum bias value used in the Alibi mechanism.
- * @param dst          The destination tensor object for additional metadata.
- *
- * The function performs the following steps:
- * 1. Calculates the logarithm floor of the number of heads to determine the
-      base for bias calculation.
- * 2. Initializes arrays with arithmetic sequences and fills them with bias
-      values.
- * 3. Computes the bias tensor based on the calculated biases and arithmetic
-      sequences.
- * 4. Reshapes the bias tensor to match the dimensions of the input tensors.
- * 5. Multiplies the position tensor by the bias tensor.
- * 6. Adds the result of the multiplication to the source tensor to produce the
-      final output.
+ * @brief Generate a range of values and apply a scalar base exponentiation.
+ *
+ * This function creates an evenly spaced sequence from `start` to `stop` (exclusive),
+ * with step size `step`, stores it in a temporary buffer, and then computes:
+ *
+ * @f[
+ * slope[i] = m^{\left( start + i \cdot step \right)}, \quad 0 \le i < size
+ * @f]
+ *
+ * The results are written to the provided @p slope_buffer.
+ *
+ * @param ctx           CANN backend context for memory allocation and operator execution.
+ * @param slope_buffer  Pointer to the output buffer (float array) for the computed slope values.
+ * @param m             Scalar base for the exponentiation.
+ * @param size          Number of elements in the generated sequence.
+ * @param start         Starting exponent offset.
+ * @param stop          Stopping exponent offset (exclusive).
+ * @param step          Step size for the exponent increment.
  */
-static void aclnn_alibi(ggml_backend_cann_context& ctx, aclTensor* acl_src,
-                        aclTensor* acl_position, aclTensor* acl_dst,
-                        const int n_head, int64_t* src_ne, const size_t src_nb0,
-                        float max_bias, ggml_tensor* dst) {
-    const int64_t ne2_ne3 = src_ne[2] * src_ne[3];
-    GGML_ASSERT(src_nb0 == sizeof(float));
-    GGML_ASSERT(n_head == src_ne[2]);
-
-    const int n_heads_log2_floor = 1u << (uint32_t)floor(log2(n_head));
-
-    float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
-    float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
-
-    // init arange
-    ggml_cann_pool_alloc arange_allocator(ctx.pool(),
-                                          ne2_ne3 * ggml_type_size(dst->type));
-    void* tmp_arange_buffer = arange_allocator.get();
+static void aclnn_get_slope_inner(ggml_backend_cann_context& ctx, void* slope_buffer,
+    float m, int64_t size, float start, float stop, float step){
+    int64_t ne[] = {size};
+    size_t nb[] = {sizeof(float)};
 
-    // arange1: [1, ..., n_heads_log2_floor+1)
-    float start = 1;
-    float stop = n_heads_log2_floor + 1;
-    float step = 1;
-    int64_t n_elements_arange = n_heads_log2_floor;
+    ggml_cann_pool_alloc arange_allocator(ctx.pool(), size * sizeof(float));
+    void* arange_buffer = arange_allocator.get();
 
-    int64_t tmp_arange1_ne[] = {n_heads_log2_floor};
-    size_t tmp_arange1_nb[] = {sizeof(dst->type)};
-    aclTensor* tmp_arange1_tensor = ggml_cann_create_tensor(
-        tmp_arange_buffer, ggml_cann_type_mapping(dst->type),
-        ggml_type_size(dst->type), tmp_arange1_ne, tmp_arange1_nb,
-        GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
+    aclTensor* arange_tensor = ggml_cann_create_tensor(
+        arange_buffer, ACL_FLOAT, sizeof(float), ne, nb, 1);
+    aclnn_arange(ctx, arange_tensor, start, stop, step, size);
 
-    aclnn_arange(ctx, tmp_arange1_tensor, start, stop, step, n_elements_arange);
-
-    aclTensor* tmp_arange2_tensor = nullptr;
-    if (n_heads_log2_floor < ne2_ne3) {
-        // arange2: [1, ..., 2 * (k - n_heads_log2_floor) + 1)
-        start = 1;
-        stop = 2 * (ne2_ne3 - n_heads_log2_floor) + 1;
-        step = 2;
-        n_elements_arange = ne2_ne3 - n_heads_log2_floor;
-        int64_t tmp_arange2_ne[] = {ne2_ne3 - n_heads_log2_floor};
-        size_t tmp_arange2_nb[] = {sizeof(dst->type)};
-
-        aclTensor* tmp_arange2_tensor = ggml_cann_create_tensor(
-            (char*)tmp_arange_buffer +
-                n_heads_log2_floor * ggml_type_size(dst->type),
-            ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type),
-            tmp_arange2_ne, tmp_arange2_nb, GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
-        aclnn_arange(ctx, tmp_arange2_tensor, start, stop, step,
-                     n_elements_arange);
-    }
+    aclTensor* slope_tensor = ggml_cann_create_tensor(
+        slope_buffer, ACL_FLOAT, sizeof(float), ne, nb, 1);
 
-    // init mk_base
-    ggml_cann_pool_alloc mk_base_allocator(ctx.pool(),
-                                           ne2_ne3 * ggml_type_size(dst->type));
-    void* tmp_mk_base_buffer = mk_base_allocator.get();
-    int64_t tmp_mk_base1_ne[] = {n_heads_log2_floor};
-    size_t tmp_mk_base1_nb[] = {sizeof(dst->type)};
-    aclTensor* tmp_mk_base1_tensor = ggml_cann_create_tensor(
-        tmp_mk_base_buffer, ggml_cann_type_mapping(dst->type),
-        ggml_type_size(dst->type), tmp_mk_base1_ne, tmp_mk_base1_nb,
-        GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
+    aclScalar* sc = aclCreateScalar(&m, aclDataType::ACL_FLOAT);
 
-    aclnn_fill_scalar(ctx, m0, tmp_mk_base1_tensor);
-
-    aclTensor* tmp_mk_base2_tensor = nullptr;
-    if (n_heads_log2_floor < ne2_ne3) {
-        int64_t tmp_mk_base2_ne[] = {ne2_ne3 - n_heads_log2_floor};
-        size_t tmp_mk_base2_nb[] = {sizeof(dst->type)};
-        aclTensor* tmp_mk_base2_tensor = ggml_cann_create_tensor(
-            (char*)tmp_mk_base_buffer +
-                n_heads_log2_floor * ggml_type_size(dst->type),
-            ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type),
-            tmp_mk_base2_ne, tmp_mk_base2_nb, GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
-        aclnn_fill_scalar(ctx, m1, tmp_mk_base2_tensor);
-    }
+    GGML_CANN_CALL_ACLNN_OP(ctx, PowScalarTensor, sc, arange_tensor, slope_tensor);
+    ggml_cann_release_resources(ctx, sc, arange_tensor, slope_tensor);
+}
 
-    // init mk
-    int64_t tmp_mk_base_ne[] = {ne2_ne3};
-    size_t tmp_mk_base_nb[] = {sizeof(dst->type)};
-    aclTensor* tmp_mk_base_tensor = ggml_cann_create_tensor(
-        tmp_mk_base_buffer, ggml_cann_type_mapping(dst->type),
-        ggml_type_size(dst->type), tmp_mk_base_ne, tmp_mk_base_nb,
-        GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
-    aclTensor* tmp_arange_tensor = ggml_cann_create_tensor(
-        tmp_arange_buffer, ggml_cann_type_mapping(dst->type),
-        ggml_type_size(dst->type), tmp_mk_base_ne, tmp_mk_base_nb,
-        GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
-    aclnn_pow_tensor_tensor(ctx, tmp_mk_base_tensor, tmp_arange_tensor);
+/**
+ * @brief Compute slope values for multiple attention heads based on ALiBi bias parameters.
+ *
+ * This function generates slope values for each attention head according to the ALiBi
+ * (Attention with Linear Biases) method. It splits the computation into two ranges depending
+ * on whether the head index is less than @p n_head_log2 or not, and uses different base values
+ * (`m0` and `m1`) for the exponentiation.
+ *
+ * @f[
+ * slope[h] =
+ * \begin{cases}
+ * m_0^{(h + 1)}, & h < n\_head\_log2 \\
+ * m_1^{\left( 2 \cdot (h - n\_head\_log2) + 1 \right)}, & h \geq n\_head\_log2
+ * \end{cases}
+ * \quad , \quad \text{if } max\_bias > 0
+ * @f]
+ *
+ * If @p max_bias <= 0, all slope values are set to 1.0.
+ *
+ * @param ctx           CANN backend context for memory allocation and operator execution.
+ * @param n_head        Total number of attention heads.
+ * @param slope_buffer  Pointer to the output buffer (float array) for storing slopes.
+ * @param max_bias      Maximum bias value for slope computation.
+ *
+*/
+static void aclnn_get_slope(ggml_backend_cann_context & ctx, int64_t n_head,
+    void* slope_buffer, float max_bias) {
+    const int n_head_log2 = 1u << (uint32_t) floor(log2(n_head));
+
+    float m0 = powf(2.0f, -(max_bias) / n_head_log2);
+    float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    // const float slope = (max_bias > 0.0f) ?
+    //                          h < n_head_log2 ?
+    //                              powf(m0, h + 1) :
+    //                              powf(m1, 2*(h - n_head_log2) + 1) :
+    //                          1.0f;
+    // arange1
+    float start = 0 + 1;
+    float end   = (n_head_log2 - 1) + 1;
+    float step  = 1;
+    float count = n_head_log2;
+    // end needs to be +1 because aclnn uses a left-closed, right-open interval.
+    aclnn_get_slope_inner(ctx, slope_buffer, m0, count, start, end + 1, step);
+    if (n_head_log2 < n_head) {
+        // arange2
+        start = 2 * (n_head_log2 - n_head_log2) + 1;
+        end   = 2 * ((n_head - 1) - n_head_log2) + 1;
+        step  = 2;
+        count = n_head - n_head_log2;
+        aclnn_get_slope_inner(
+            ctx, (char *) slope_buffer + n_head_log2 * sizeof(float),
+            m1, count, start, end + 1, step);
+    }
+}
 
-    // reshape mk
-    int64_t tmp_mk_ne[] = {1, 1, src_ne[2], src_ne[3]};
-    size_t tmp_mk_nb[GGML_MAX_DIMS];
-    tmp_mk_nb[0] = ggml_type_size(dst->type);
-    for (int i = 1; i < GGML_MAX_DIMS; i++) {
-        tmp_mk_nb[i] = tmp_mk_nb[i - 1] * tmp_mk_ne[i - 1];
+/**
+ * @brief Add ALiBi (Attention with Linear Biases) positional biases to the attention mask.
+ *
+ * This function computes the ALiBi slopes for each attention head (if max_bias > 0),
+ * multiplies them with the attention mask to produce bias tensors, and adds these biases
+ * to the destination tensor (@p dst).
+ *
+ * The function performs necessary broadcasting of the mask and slope tensors to match
+ * the shape of the destination tensor, then applies element-wise multiplication and addition
+ * using CANN operators.
+ *
+ * @param ctx         CANN backend context for memory management and operator execution.
+ * @param mask        Input attention mask tensor, assumed to be contiguous.
+ * @param dst         Destination tensor to which ALiBi biases will be added.
+ * @param dst_ptr     Pointer to the memory of the destination tensor.
+ * @param max_bias    Maximum bias value controlling the slope scaling.
+ *
+ * @note
+ * - Write data into dst_ptr using only the shape information of the dst tensor.
+ * - `GGML_MAX_DIMS + 2` is used to extend tensor dimensions for broadcasting.
+ */
+static void aclnn_add_alibi(ggml_backend_cann_context& ctx, ggml_tensor* mask,
+    ggml_tensor* dst, void* dst_ptr, float max_bias) {
+    void* slope_buffer = nullptr;
+    void* bias_buffer = nullptr;
+
+    if (max_bias > 0.0f) {
+        int64_t n_heads = dst->ne[2];
+        ggml_cann_pool_alloc slope_allocator(ctx.pool(), n_heads * sizeof(float));
+        slope_buffer = slope_allocator.get();
+        ggml_cann_pool_alloc bias_allocator(
+                    ctx.pool(), ggml_nelements(dst) * ggml_element_size(dst));
+        bias_buffer = bias_allocator.get();
+        aclnn_get_slope(ctx, n_heads, slope_buffer, max_bias);
     }
-    aclTensor* tmp_mk_tensor = ggml_cann_create_tensor(
-        tmp_mk_base_buffer, ggml_cann_type_mapping(dst->type),
-        ggml_type_size(dst->type), tmp_mk_ne, tmp_mk_nb, GGML_MAX_DIMS,
-        ACL_FORMAT_ND);
 
-    // acl_position * mk
-    int64_t tmp_output_ne[] = {src_ne[0], src_ne[1], src_ne[2], src_ne[3]};
-    size_t tmp_output_nb[GGML_MAX_DIMS];
-    tmp_output_nb[0] = ggml_type_size(dst->type);
-    for (int i = 1; i < GGML_MAX_DIMS; i++) {
-        tmp_output_nb[i] = tmp_output_nb[i - 1] * tmp_output_ne[i - 1];
+    // broadcast for mask, slop and dst;
+    int64_t nr2 = dst->ne[2] / mask->ne[2];
+    int64_t nr3 = dst->ne[3] / mask->ne[3];
+
+    // broadcast the mask across rows
+    int64_t mask_ne[] = { mask->ne[0], dst->ne[1], mask->ne[2], 1, mask->ne[3], 1 };
+    size_t  mask_nb[] = {
+        mask_nb[0] = mask->nb[0], mask_nb[1] = mask->nb[1], mask_nb[2] = mask->nb[2],
+        mask_nb[3] = mask->nb[2], mask_nb[4] = mask->nb[3], mask_nb[5] = mask->nb[3]
+    };
+
+    int64_t dst_ne[] = { dst->ne[0], dst->ne[1], mask->ne[2], nr2, mask->ne[3], nr3 };
+    size_t  dst_nb[] = {
+        dst_nb[0] = dst->nb[0], dst_nb[1] = dst->nb[1], dst_nb[2] = dst->nb[2],
+        dst_nb[3] = dst->nb[2], dst_nb[4] = dst->nb[3], dst_nb[5] = dst->nb[3]
+    };
+
+    // slope is a 1 dim tensor, slope.ne2 == dst.ne2
+    int64_t slope_ne[] = { 1, 1, mask->ne[2], nr2, 1, 1 };
+    size_t  slope_nb[GGML_MAX_DIMS + 2];
+    slope_nb[0] = sizeof(float);
+    for (int i = 1; i < GGML_MAX_DIMS + 2; i++) {
+        slope_nb[i] = slope_nb[i - 1] * slope_ne[i - 1];
     }
-    ggml_cann_pool_alloc output_allocator(ctx.pool(), ggml_nbytes(dst));
-    void* tmp_output_buffer = output_allocator.get();
-    aclTensor* tmp_output_tensor = ggml_cann_create_tensor(
-        tmp_output_buffer, ggml_cann_type_mapping(dst->type),
-        ggml_type_size(dst->type), tmp_output_ne, tmp_output_nb, GGML_MAX_DIMS,
-        ACL_FORMAT_ND);
-    aclnn_mul(ctx, acl_position, tmp_mk_tensor, tmp_output_tensor);
 
-    // add
-    aclnn_add(ctx, tmp_output_tensor, acl_src, acl_dst);
-    ggml_cann_release_resources(ctx, tmp_arange1_tensor, tmp_arange2_tensor,
-        tmp_mk_base1_tensor, tmp_mk_base2_tensor, tmp_mk_base_tensor,
-        tmp_arange_tensor, tmp_mk_tensor, tmp_output_tensor);
+    aclTensor* acl_slope = ggml_cann_create_tensor(
+                            slope_buffer, ACL_FLOAT, sizeof(float),
+                            slope_ne, slope_nb, GGML_MAX_DIMS + 2);
+    aclTensor* acl_mask = ggml_cann_create_tensor(
+                            mask, mask_ne, mask_nb, GGML_MAX_DIMS + 2);
+
+    // write data into dst_ptr using only the shape information of the dst tensor.
+    aclTensor* acl_dst  = ggml_cann_create_tensor(
+                            dst_ptr, ggml_cann_type_mapping(dst->type),
+                            ggml_type_size(dst->type), dst_ne, dst_nb,
+                            GGML_MAX_DIMS + 2);
+
+    if (max_bias > 0.0f) {
+        int64_t bias_ne[] = { mask->ne[0], dst->ne[1], mask->ne[2], nr2, mask->ne[3], 1 };
+        size_t  bias_nb[GGML_MAX_DIMS + 2];
+        bias_nb[0] = sizeof(float);
+        for (int i = 1; i < GGML_MAX_DIMS + 2; i++) {
+            bias_nb[i] = bias_nb[i - 1] * bias_ne[i - 1];
+        }
+        aclTensor* bias_tensor = ggml_cann_create_tensor(
+                                    bias_buffer, ACL_FLOAT, sizeof(float),
+                                    bias_ne, bias_nb, GGML_MAX_DIMS + 2);
+
+        aclnn_mul(ctx, acl_slope, acl_mask, bias_tensor);
+        aclnn_add(ctx, acl_dst, bias_tensor);
+        ggml_cann_release_resources(ctx, bias_tensor);
+    } else {
+        aclnn_add(ctx, acl_dst, acl_mask);
+    }
+    ggml_cann_release_resources(ctx, acl_slope, acl_mask, acl_dst);
 }
 
-void ggml_cann_cpy(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+void ggml_cann_cpy(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     ggml_cann_dup(ctx, dst);
 }
 
@@ -1501,118 +1537,41 @@ void ggml_cann_cpy(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
  * @param acl_dst The destination tensor where the softmax results will be
  * stored.
  */
-static void aclnn_softmax(ggml_backend_cann_context& ctx, aclTensor* acl_src,
-                          int64_t dim, aclTensor* acl_dst) {
+static void aclnn_softmax(ggml_backend_cann_context & ctx,
+    aclTensor* acl_src, int64_t dim, aclTensor * acl_dst) {
     GGML_CANN_CALL_ACLNN_OP(ctx, Softmax, acl_src, dim, acl_dst);
 }
 
-void ggml_cann_softmax(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+void ggml_cann_softmax(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     ggml_tensor* src0 = dst->src[0];
     ggml_tensor* src1 = dst->src[1];  // mask
 
     aclTensor* acl_src0 = ggml_cann_create_tensor(src0);
-    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+    aclTensor* acl_dst  = ggml_cann_create_tensor(dst);
 
-    float scale = 1.0f;
+    float scale    = 1.0f;
     float max_bias = 0.0f;
 
-    memcpy(&scale, (float*)dst->op_params + 0, sizeof(float));
-    memcpy(&max_bias, (float*)dst->op_params + 1, sizeof(float));
+    memcpy(&scale, (float *) dst->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float));
 
     // input mul scale
     aclScalar* acl_scale = aclCreateScalar(&scale, aclDataType::ACL_FLOAT);
+    ggml_cann_pool_alloc src_tensor_allocator(ctx.pool(), ggml_nbytes(src0));
+    void* src_tensor_buffer = src_tensor_allocator.get();
+    aclTensor* softmax_tensor = ggml_cann_create_tensor(
+        src_tensor_buffer, ggml_cann_type_mapping(src0->type),
+        ggml_element_size(src0), src0->ne, src0->nb,GGML_MAX_DIMS);
 
-    size_t n_bytes = ggml_nbytes(src0);
-    ggml_cann_pool_alloc mul_scale_allocator(ctx.pool(), n_bytes);
-    void* input_mul_scale_buffer = mul_scale_allocator.get();
-    aclTensor* acl_input_mul_scale_tensor = ggml_cann_create_tensor(
-        input_mul_scale_buffer, ACL_FLOAT, ggml_type_size(src0->type), src0->ne,
-        src0->nb, GGML_MAX_DIMS);
-
-    bool inplace = false;
-    aclnn_muls(ctx, acl_src0, scale, acl_input_mul_scale_tensor, inplace);
+    aclnn_muls(ctx, acl_src0, scale, softmax_tensor, false);
 
     // mask
-    aclTensor* acl_src1_fp32_tensor = nullptr;
-    aclTensor* tmp_mask_tensor = nullptr;
-    ggml_cann_pool_alloc src1_fp32_allocator(ctx.pool());
     if (src1) {
-        const bool use_f16 = src1->type == GGML_TYPE_F16;
-        if (use_f16) {
-            // cast to fp32
-            size_t n_bytes = ggml_nelements(src1) * sizeof(float_t);
-            size_t src1_fp32_nb[GGML_MAX_DIMS];
-            src1_fp32_nb[0] = sizeof(float_t);
-            for (int i = 1; i < GGML_MAX_DIMS; i++) {
-                src1_fp32_nb[i] = src1_fp32_nb[i - 1] * src1->ne[i - 1];
-            }
-            src1_fp32_allocator.alloc(n_bytes);
-            void* src1_fp32_buffer = src1_fp32_allocator.get();
-            acl_src1_fp32_tensor = ggml_cann_create_tensor(
-                src1_fp32_buffer, ACL_FLOAT, sizeof(float), src1->ne,
-                src1_fp32_nb, GGML_MAX_DIMS);
-            aclTensor* acl_src1 = ggml_cann_create_tensor(src1);
-            aclnn_cast(ctx, acl_src1, acl_src1_fp32_tensor, ACL_FLOAT);
-            ggml_cann_release_resources(ctx, acl_src1);
-        } else {
-            acl_src1_fp32_tensor = ggml_cann_create_tensor(src1);
-        }
-
-        // broadcast the mask across rows, only use ne11 of ne01 in mask
-        if (src1->ne[1] != src0->ne[1]) {
-            // mask shape: [1,1,ne11,ne10]
-            int64_t tmp_mask_ne[] = {src0->ne[0], src0->ne[1], 1, 1};
-            size_t tmp_mask_nb[GGML_MAX_DIMS];
-            tmp_mask_nb[0] = sizeof(float_t);
-            for (int i = 1; i < GGML_MAX_DIMS; i++) {
-                tmp_mask_nb[i] = tmp_mask_nb[i - 1] * tmp_mask_ne[i - 1];
-            }
-            tmp_mask_tensor = ggml_cann_create_tensor(
-                src1->data, ACL_FLOAT, sizeof(float), tmp_mask_ne, tmp_mask_nb,
-                GGML_MAX_DIMS, ACL_FORMAT_ND);
-        }
-
-        // alibi
-        const int n_head = src0->ne[2];
-        const size_t src_nb0 = src0->nb[0];
-
-        n_bytes = ggml_nbytes(dst);
-        ggml_cann_pool_alloc output_allocator(ctx.pool(), n_bytes);
-        void* output_buffer = output_allocator.get();
-        aclTensor* alibi_output_tensor = ggml_cann_create_tensor(
-            output_buffer, ACL_FLOAT, ggml_type_size(dst->type), dst->ne,
-            dst->nb, GGML_MAX_DIMS);
-        if (max_bias <= 0.0f) {
-            // slope = 1.0
-            if (tmp_mask_tensor) {
-                aclnn_add(ctx, tmp_mask_tensor, acl_input_mul_scale_tensor,
-                          alibi_output_tensor);
-            } else {
-                aclnn_add(ctx, acl_src1_fp32_tensor, acl_input_mul_scale_tensor,
-                          alibi_output_tensor);
-            }
-        } else {
-            // slope != 1.0
-            if (tmp_mask_tensor) {
-                aclnn_alibi(ctx, acl_input_mul_scale_tensor, tmp_mask_tensor,
-                            alibi_output_tensor, n_head, src0->ne, src_nb0,
-                            max_bias, dst);
-            } else {
-                aclnn_alibi(ctx, acl_input_mul_scale_tensor,
-                            acl_src1_fp32_tensor, alibi_output_tensor, n_head,
-                            src0->ne, src_nb0, max_bias, dst);
-            }
-        }
-
-        // softmax
-        aclnn_softmax(ctx, alibi_output_tensor, 3, acl_dst);
-        ggml_cann_release_resources(ctx, alibi_output_tensor);
-    } else {
-        aclnn_softmax(ctx, acl_input_mul_scale_tensor, 3, acl_dst);
+        aclnn_add_alibi(ctx, src1, src0, src_tensor_buffer, max_bias);
     }
-
-    ggml_cann_release_resources(ctx, acl_src0, acl_src1_fp32_tensor, acl_dst,
-        acl_scale, acl_input_mul_scale_tensor, tmp_mask_tensor);
+    // softmax
+    aclnn_softmax(ctx, softmax_tensor, 3, acl_dst);
+    ggml_cann_release_resources(ctx, acl_src0, acl_dst, acl_scale, softmax_tensor);
 }
 
 /**
@@ -3208,104 +3167,24 @@ void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst){
             // Compute the slope if needed. Derived from ggml_cann_softmax().
             if(maxBias != 0.0f){
                 // alibi
-                const int64_t ne2_ne3 = src0->ne[2] * src0->ne[3];
-                const int64_t n_head = src0->ne[2];
-                const int n_heads_log2_floor = 1u << (uint32_t)floor(log2(n_head));
-                float m0 = powf(2.0f, -(maxBias) / n_heads_log2_floor);
-                float m1 = powf(2.0f, -(maxBias / 2.0f) / n_heads_log2_floor);
-                // init arange
-                ggml_cann_pool_alloc arange_allocator(ctx.pool(),
-                                                    ne2_ne3 * faElemSize);
-                void* tmp_arange_buffer = arange_allocator.get();
-
-                // arange1: [1, ..., n_heads_log2_floor+1)
-                float start = 1;
-                float stop = n_heads_log2_floor + 1;
-                float step = 1;
-                int64_t n_elements_arange = n_heads_log2_floor;
-
-                int64_t tmp_arange1_ne[] = {n_heads_log2_floor};
-                size_t tmp_arange1_nb[] = {faElemSize};
-                aclTensor* tmp_arange1_tensor = ggml_cann_create_tensor(
-                    tmp_arange_buffer, faDataType, faElemSize,
-                    tmp_arange1_ne, tmp_arange1_nb,
-                    GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
-
-                aclnn_arange(ctx, tmp_arange1_tensor, start, stop, step, n_elements_arange);
-
-                aclTensor* tmp_arange2_tensor = nullptr;
-                if (n_heads_log2_floor < ne2_ne3) {
-                    // arange2: [1, ..., 2 * (k - n_heads_log2_floor) + 1)
-                    start = 1;
-                    stop = 2 * (ne2_ne3 - n_heads_log2_floor) + 1;
-                    step = 2;
-                    n_elements_arange = ne2_ne3 - n_heads_log2_floor;
-                    int64_t tmp_arange2_ne[] = {ne2_ne3 - n_heads_log2_floor};
-                    size_t tmp_arange2_nb[] = {faElemSize};
-
-                    aclTensor* tmp_arange2_tensor = ggml_cann_create_tensor(
-                        (char*)tmp_arange_buffer +
-                            n_heads_log2_floor * faElemSize,
-                        faDataType, faElemSize,
-                        tmp_arange2_ne, tmp_arange2_nb, GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
-                    aclnn_arange(ctx, tmp_arange2_tensor, start, stop, step,
-                                n_elements_arange);
+                const int64_t n_heads = src0->ne[2];
+                ggml_cann_pool_alloc slope_allocator(ctx.pool(), n_heads * sizeof(float));
+                void* slope_buffer = slope_allocator.get();
+                aclnn_get_slope(ctx, n_heads, slope_buffer, maxBias);
+
+                int64_t slope_ne[] = {1, 1, n_heads, 1};
+                size_t slope_nb[GGML_MAX_DIMS];
+                slope_nb[0] = sizeof(float);
+                for(int i = 1;i<GGML_MAX_DIMS;i++) {
+                    slope_nb[i] = slope_nb[i-1] * slope_ne[0];
                 }
 
-                // init mk_base
-                ggml_cann_pool_alloc mk_base_allocator(ctx.pool(),
-                                                    ne2_ne3 * faElemSize);
-                void* tmp_mk_base_buffer = mk_base_allocator.get();
-                int64_t tmp_mk_base1_ne[] = {n_heads_log2_floor};
-                size_t tmp_mk_base1_nb[] = {faElemSize};
-                aclTensor* tmp_mk_base1_tensor = ggml_cann_create_tensor(
-                    tmp_mk_base_buffer, faDataType, faElemSize,
-                    tmp_mk_base1_ne, tmp_mk_base1_nb,
-                    GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
-
-                aclnn_fill_scalar(ctx, m0, tmp_mk_base1_tensor);
-
-                aclTensor* tmp_mk_base2_tensor = nullptr;
-                if (n_heads_log2_floor < ne2_ne3) {
-                    int64_t tmp_mk_base2_ne[] = {ne2_ne3 - n_heads_log2_floor};
-                    size_t tmp_mk_base2_nb[] = {faElemSize};
-                    aclTensor* tmp_mk_base2_tensor = ggml_cann_create_tensor(
-                        (char*)tmp_mk_base_buffer +
-                            n_heads_log2_floor * faElemSize,
-                        faDataType, faElemSize,
-                        tmp_mk_base2_ne, tmp_mk_base2_nb, GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
-                    aclnn_fill_scalar(ctx, m1, tmp_mk_base2_tensor);
-                }
+                aclTensor* slope_tensor = ggml_cann_create_tensor(
+                    slope_buffer, ACL_FLOAT, sizeof(float),
+                    slope_ne, slope_nb, GGML_MAX_DIMS);
+                GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMul, bcast_pse_tensor, slope_tensor);
 
-                // init mk
-                int64_t tmp_mk_base_ne[] = {ne2_ne3};
-                size_t tmp_mk_base_nb[] = {faElemSize};
-                aclTensor* tmp_mk_base_tensor = ggml_cann_create_tensor(
-                    tmp_mk_base_buffer, faDataType, faElemSize,
-                    tmp_mk_base_ne, tmp_mk_base_nb,
-                    GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
-                aclTensor* tmp_arange_tensor = ggml_cann_create_tensor(
-                    tmp_arange_buffer, faDataType, faElemSize,
-                    tmp_mk_base_ne, tmp_mk_base_nb,
-                    GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
-                aclnn_pow_tensor_tensor(ctx, tmp_mk_base_tensor, tmp_arange_tensor);
-
-                // reshape mk
-                int64_t tmp_mk_ne[] = {1, 1, src0->ne[2], src0->ne[3]};
-                size_t tmp_mk_nb[GGML_MAX_DIMS];
-                tmp_mk_nb[0] = faElemSize;
-                for (int i = 1; i < GGML_MAX_DIMS; i++) {
-                    tmp_mk_nb[i] = tmp_mk_nb[i - 1] * tmp_mk_ne[i - 1];
-                }
-                aclTensor* tmp_mk_tensor = ggml_cann_create_tensor(
-                    tmp_mk_base_buffer, faDataType, faElemSize,
-                    tmp_mk_ne, tmp_mk_nb, GGML_MAX_DIMS,
-                    ACL_FORMAT_ND);
-                GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMul, bcast_pse_tensor, tmp_mk_tensor);
-
-                ggml_cann_release_resources(ctx, tmp_arange1_tensor, tmp_arange2_tensor,
-                    tmp_mk_base1_tensor, tmp_mk_base2_tensor, tmp_mk_base_tensor,
-                    tmp_arange_tensor, tmp_mk_tensor);
+                ggml_cann_release_resources(ctx, slope_tensor);
             }
         }
 
diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
index cf575b36750..3d3520f1959 100755
--- a/ggml/src/ggml-cann/ggml-cann.cpp
+++ b/ggml/src/ggml-cann/ggml-cann.cpp
@@ -2391,7 +2391,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 // only support F32 and F16.
                 return false;
             }
-            return true;
+            return ggml_is_contiguous(op);
         } break;
         case GGML_OP_CONT: {
             // TODO: support GGML_TYPE_BF16
@@ -2456,8 +2456,9 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             // value of paddingW should be at most half of kernelW
             return (p0 <= (k0 / 2)) && (p1 <= (k1 / 2));
         }
-        case GGML_OP_SUM:
         case GGML_OP_DUP:
+            return ggml_is_contiguous(op);
+        case GGML_OP_SUM:
         case GGML_OP_IM2COL:
         case GGML_OP_CONCAT:
         case GGML_OP_REPEAT:
@@ -2503,9 +2504,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             if (op->src[2]) {
                 return false;
             }
-            // TODO: support broadcast
-            // ref: https://github.com/ggml-org/llama.cpp/pull/14435
-            return !op->src[1] || (op->src[1]->ne[2] == 1 && op->src[1]->ne[3] == 1);
+            return true;
         case GGML_OP_FLASH_ATTN_EXT:{
             // derived from [ggml-cuda.cu]
             if(op->src[1]->type != GGML_TYPE_F16 || op->src[2]->type != GGML_TYPE_F16){
@@ -2532,11 +2531,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 // DeepSeek MLA
                 return false;
             }
-            // TODO: support broadcast
-            // ref: https://github.com/ggml-org/llama.cpp/pull/14435
-            if (op->src[0]->ne[3] != 1) {
-                return false;
-            }
             float logitSoftcap = 0.0f;
             memcpy(&logitSoftcap,  (float*)op->op_params + 2, sizeof(float));
             if(logitSoftcap != 0.0f) {

From 74cabac71f7600f14b79bd6282d3172177709345 Mon Sep 17 00:00:00 2001
From: R0CKSTAR <yeahdongcn@gmail.com>
Date: Tue, 12 Aug 2025 10:02:51 +0800
Subject: [PATCH 52/86] musa: fix failures in test-backend-ops for mul_mat_id
 op (llama/15236)

* musa: fix failures in test-backend-ops for mul_mat_id op

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

* Address review comments

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

---------

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
---
 ggml/src/ggml-cuda/common.cuh | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
index 2e5d48797fa..a23da57e3a1 100644
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -312,11 +312,11 @@ static bool turing_mma_available(const int cc) {
 }
 
 static bool ampere_mma_available(const int cc) {
-    return cc < GGML_CUDA_CC_OFFSET_AMD && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_AMPERE;
+    return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_AMPERE;
 }
 
 static bool cp_async_available(const int cc) {
-    return cc < GGML_CUDA_CC_OFFSET_AMD && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_AMPERE;
+    return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_AMPERE;
 }
 
 static constexpr __device__ int ggml_cuda_get_physical_warp_size() {

From 5691b54f793621c7c602f3fb4907eb8d5a8b14ad Mon Sep 17 00:00:00 2001
From: Chenguang Li <757486878@qq.com>
Date: Tue, 12 Aug 2025 16:12:13 +0800
Subject: [PATCH 53/86] CANN: GGML_OP_CPY optimization (llama/15070)

Signed-off-by: noemotiovon <757486878@qq.com>
---
 ggml/src/ggml-cann/aclnn_ops.cpp | 76 +++++++++++++-------------------
 ggml/src/ggml-cann/ggml-cann.cpp |  3 +-
 2 files changed, 32 insertions(+), 47 deletions(-)

diff --git a/ggml/src/ggml-cann/aclnn_ops.cpp b/ggml/src/ggml-cann/aclnn_ops.cpp
index 0b409ce87d2..259a2928b1f 100755
--- a/ggml/src/ggml-cann/aclnn_ops.cpp
+++ b/ggml/src/ggml-cann/aclnn_ops.cpp
@@ -753,69 +753,55 @@ static void cann_copy(ggml_backend_cann_context& ctx, aclTensor* acl_src,
 void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_tensor* src0 = dst->src[0];
 
-    aclTensor* acl_src = ggml_cann_create_tensor(src0);
-    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
     if (ggml_are_same_shape(src0, dst)) {
+        aclTensor* acl_src = ggml_cann_create_tensor(src0);
+        aclTensor* acl_dst = ggml_cann_create_tensor(dst);
         if (dst->type == src0->type) {
             cann_copy(ctx, acl_src, acl_dst);
         } else {
             aclnn_cast(ctx, acl_src, acl_dst, ggml_cann_type_mapping(dst->type));
         }
+        ggml_cann_release_resources(ctx, acl_src, acl_dst);
     } else {
-        if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst)) {
-            if (dst->type == src0->type) {
-                size_t cpy_size = ggml_nbytes(dst);
-                ggml_cann_async_memcpy(ctx, dst->data, src0->data, cpy_size,
-                    ACL_MEMCPY_DEVICE_TO_DEVICE);
-                return;
-            } else {
-                ggml_cann_pool_alloc src_buffer_allocator(
-                    ctx.pool(),
-                    ggml_nelements(dst) * ggml_type_size(dst->type));
-                void* src_trans_buffer = src_buffer_allocator.get();
-                size_t src_trans_nb[GGML_MAX_DIMS];
-                src_trans_nb[0] = ggml_type_size(dst->type);
-                for (int i = 1; i < GGML_MAX_DIMS; i++) {
-                    src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1];
-                }
-                aclTensor* src_trans_tensor = ggml_cann_create_tensor(
-                    src_trans_buffer, ggml_cann_type_mapping(dst->type),
-                    ggml_type_size(dst->type), src0->ne, src_trans_nb,
-                    GGML_MAX_DIMS);
-
-                aclnn_cast(ctx, acl_src, src_trans_tensor, ggml_cann_type_mapping(dst->type));
-                size_t cpy_size = ggml_nbytes(dst);
-                ggml_cann_async_memcpy(ctx, dst->data, src_trans_buffer, cpy_size,
-                    ACL_MEMCPY_DEVICE_TO_DEVICE);
-                ggml_cann_release_resources(ctx, src_trans_tensor);
-                return;
-            }
-        } else if (ggml_is_contiguous(dst)) {
-            ggml_cann_pool_alloc src_buffer_allocator(
-                ctx.pool(), ggml_nelements(dst) * ggml_type_size(dst->type));
-            void* src_trans_buffer = src_buffer_allocator.get();
+        void* src_trans_buffer = src0->data;
+        ggml_cann_pool_alloc src_buffer_allocator;
+        if (!ggml_is_contiguous(src0)) {
+            aclTensor* acl_src = ggml_cann_create_tensor(src0);
+            src_buffer_allocator.alloc(ctx.pool(),
+                ggml_nelements(src0) * ggml_type_size(src0->type));
+            src_trans_buffer = src_buffer_allocator.get();
             size_t src_trans_nb[GGML_MAX_DIMS];
-            src_trans_nb[0] = ggml_type_size(dst->type);
+            src_trans_nb[0] = ggml_type_size(src0->type);
             for (int i = 1; i < GGML_MAX_DIMS; i++) {
                 src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1];
             }
             aclTensor* src_trans_tensor = ggml_cann_create_tensor(
-                src_trans_buffer, ggml_cann_type_mapping(dst->type),
-                ggml_type_size(dst->type), src0->ne, src_trans_nb,
+                src_trans_buffer, ggml_cann_type_mapping(src0->type),
+                ggml_type_size(src0->type), src0->ne, src_trans_nb,
                 GGML_MAX_DIMS);
+            cann_copy(ctx, acl_src, src_trans_tensor);
+            ggml_cann_release_resources(ctx, acl_src, src_trans_tensor);
+        }
 
-            aclnn_cast(ctx, acl_src, src_trans_tensor, ggml_cann_type_mapping(dst->type));
+        size_t src_reshape_nb[GGML_MAX_DIMS];
+        src_reshape_nb[0] = ggml_type_size(src0->type);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            src_reshape_nb[i] = src_reshape_nb[i - 1] * dst->ne[i - 1];
+        }
 
-            size_t cpy_size = ggml_nbytes(dst);
-            ggml_cann_async_memcpy(ctx, dst->data, src_trans_buffer, cpy_size,
-                ACL_MEMCPY_DEVICE_TO_DEVICE);
-            ggml_cann_release_resources(ctx, src_trans_tensor);
-            return;
+        aclTensor* trans_acl_src = ggml_cann_create_tensor(src_trans_buffer,
+            ggml_cann_type_mapping(src0->type),ggml_type_size(src0->type),
+            dst->ne, src_reshape_nb, GGML_MAX_DIMS, ACL_FORMAT_ND);
+        aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+
+        if (dst->type == src0->type) {
+            cann_copy(ctx, trans_acl_src, acl_dst);
         } else {
-            GGML_ABORT("Unsupport dst is not contiguous.");
+            aclnn_cast(ctx, trans_acl_src, acl_dst, ggml_cann_type_mapping(dst->type));
         }
+        ggml_cann_release_resources(ctx, trans_acl_src, acl_dst);
     }
-    ggml_cann_release_resources(ctx, acl_src, acl_dst);
+    return;
 }
 
 /**
diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
index 3d3520f1959..cb8af42ebf9 100755
--- a/ggml/src/ggml-cann/ggml-cann.cpp
+++ b/ggml/src/ggml-cann/ggml-cann.cpp
@@ -2391,7 +2391,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 // only support F32 and F16.
                 return false;
             }
-            return ggml_is_contiguous(op);
+            return true;
         } break;
         case GGML_OP_CONT: {
             // TODO: support GGML_TYPE_BF16
@@ -2457,7 +2457,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             return (p0 <= (k0 / 2)) && (p1 <= (k1 / 2));
         }
         case GGML_OP_DUP:
-            return ggml_is_contiguous(op);
         case GGML_OP_SUM:
         case GGML_OP_IM2COL:
         case GGML_OP_CONCAT:

From ce7f91a1daf2d1684844f6c13990023537e05458 Mon Sep 17 00:00:00 2001
From: Aman Gupta <amangupta052@gmail.com>
Date: Tue, 12 Aug 2025 17:21:45 +0800
Subject: [PATCH 54/86] CUDA cmake: add `-lineinfo` for easier debug
 (llama/15260)

---
 ggml/src/ggml-cuda/CMakeLists.txt | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/ggml/src/ggml-cuda/CMakeLists.txt b/ggml/src/ggml-cuda/CMakeLists.txt
index 98ed29bc9c1..bce07ac3628 100644
--- a/ggml/src/ggml-cuda/CMakeLists.txt
+++ b/ggml/src/ggml-cuda/CMakeLists.txt
@@ -120,6 +120,10 @@ if (CUDAToolkit_FOUND)
 
     set(CUDA_FLAGS -use_fast_math -extended-lambda)
 
+    if (GGML_CUDA_DEBUG)
+        list(APPEND CUDA_FLAGS -lineinfo)
+    endif()
+
     if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.8")
         # Options are:
         # - none (not recommended)

From e0815176e641fa2f74faad3ec32f01cd2b414aec Mon Sep 17 00:00:00 2001
From: rmatif <kingrealriadh@gmail.com>
Date: Tue, 12 Aug 2025 11:42:41 +0200
Subject: [PATCH 55/86] opencl: allow mixed f16/f32 `add` (llama/15140)

---
 ggml/src/ggml-opencl/ggml-opencl.cpp | 191 +++++++++++++++++----------
 ggml/src/ggml-opencl/kernels/add.cl  |  50 +++++--
 2 files changed, 162 insertions(+), 79 deletions(-)

diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
index 8ba1e00df7b..fc838684ac6 100644
--- a/ggml/src/ggml-opencl/ggml-opencl.cpp
+++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
@@ -2481,6 +2481,13 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
         case GGML_OP_SCALE:
             return op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]);
         case GGML_OP_ADD:
+            if (op->type == GGML_TYPE_F16) {
+                const bool src0_ok = op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_F32;
+                const bool src1_ok = op->src[1]->type == GGML_TYPE_F16 || op->src[1]->type == GGML_TYPE_F32;
+                if (src0_ok && src1_ok) {
+                    return true;
+                }
+            }
         case GGML_OP_MUL:
         case GGML_OP_DIV:
         case GGML_OP_SUB:
@@ -3717,34 +3724,30 @@ static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const
     GGML_ASSERT(dst);
     GGML_ASSERT(dst->extra);
 
-    GGML_ASSERT(src0->type == src1->type);
-    GGML_ASSERT(src0->type == dst->type);
-    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
-
-    const int  ne00 = src0->ne[0];
-    const int  ne01 = src0->ne[1];
-    const int  ne02 = src0->ne[2];
-    const int  ne03 = src0->ne[3];
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
 
     const cl_ulong nb00 = src0->nb[0];
     const cl_ulong nb01 = src0->nb[1];
     const cl_ulong nb02 = src0->nb[2];
     const cl_ulong nb03 = src0->nb[3];
 
-    const int  ne10 = src1->ne[0];
-    const int  ne11 = src1->ne[1];
-    const int  ne12 = src1->ne[2];
-    const int  ne13 = src1->ne[3]; UNUSED(ne13);
+    const int ne10 = src1->ne[0];
+    const int ne11 = src1->ne[1];
+    const int ne12 = src1->ne[2];
+    const int ne13 = src1->ne[3];
 
     const cl_ulong nb10 = src1->nb[0];
     const cl_ulong nb11 = src1->nb[1];
     const cl_ulong nb12 = src1->nb[2];
-    const cl_ulong nb13 = src1->nb[3]; UNUSED(nb13);
+    const cl_ulong nb13 = src1->nb[3];
 
-    const int  ne0  = dst->ne[0];
-    const int  ne1  = dst->ne[1];
-    const int  ne2  = dst->ne[2];
-    const int  ne3  = dst->ne[3];
+    const int ne0  = dst->ne[0];
+    const int ne1  = dst->ne[1];
+    const int ne2  = dst->ne[2];
+    const int ne3  = dst->ne[3];
 
     const cl_ulong nb0  = dst->nb[0];
     const cl_ulong nb1  = dst->nb[1];
@@ -3761,68 +3764,114 @@ static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const
     cl_ulong offset1 = extra1->offset + src1->view_offs;
     cl_ulong offsetd = extrad->offset + dst->view_offs;
 
-    bool bcast_row = false;
     cl_kernel kernel;
 
-    if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) {
-        GGML_ASSERT(ggml_is_contiguous(src0));
+    const bool bcast_row = ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0;
 
-        // src1 is a row
+    if (bcast_row) {
+        GGML_ASSERT(ggml_is_contiguous(src0));
         GGML_ASSERT(ne11 == 1);
+    }
 
-        bcast_row = true;
-        int ne = ne00 / 4;
-
-        if (src0->type == GGML_TYPE_F32) {
+    if (dst->type == GGML_TYPE_F32) {
+        GGML_ASSERT(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32);
+        if (bcast_row) {
             kernel = backend_ctx->kernel_add_row;
+            const int ne = ne00 / 4;
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne));
         } else {
-            kernel = backend_ctx->kernel_add_row_f16;
-        }
-
-        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
-        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
-        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra1->data_device));
-        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
-        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extrad->data_device));
-        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
-        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne));
-    } else {
-        if (src0->type == GGML_TYPE_F32) {
             kernel = backend_ctx->kernel_add;
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne03));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb00));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne13));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb10));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int),      &ne2));
+            CL_CHECK(clSetKernelArg(kernel, 25, sizeof(int),      &ne3));
+            CL_CHECK(clSetKernelArg(kernel, 26, sizeof(cl_ulong), &nb0));
+            CL_CHECK(clSetKernelArg(kernel, 27, sizeof(cl_ulong), &nb1));
+            CL_CHECK(clSetKernelArg(kernel, 28, sizeof(cl_ulong), &nb2));
+            CL_CHECK(clSetKernelArg(kernel, 29, sizeof(cl_ulong), &nb3));
+        }
+    } else if (dst->type == GGML_TYPE_F16) {
+        GGML_ASSERT(src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_F32);
+        GGML_ASSERT(src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32);
+        const int type_src0 = (src0->type == GGML_TYPE_F32);
+        const int type_src1 = (src1->type == GGML_TYPE_F32);
+        if (bcast_row) {
+            kernel = backend_ctx->kernel_add_row_f16;
+            const int ne = ne00 / 4;
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne));
+            CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),      &type_src0));
+            CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),      &type_src1));
         } else {
             kernel = backend_ctx->kernel_add_f16;
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne03));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb00));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne13));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb10));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int),      &ne2));
+            CL_CHECK(clSetKernelArg(kernel, 25, sizeof(int),      &ne3));
+            CL_CHECK(clSetKernelArg(kernel, 26, sizeof(cl_ulong), &nb0));
+            CL_CHECK(clSetKernelArg(kernel, 27, sizeof(cl_ulong), &nb1));
+            CL_CHECK(clSetKernelArg(kernel, 28, sizeof(cl_ulong), &nb2));
+            CL_CHECK(clSetKernelArg(kernel, 29, sizeof(cl_ulong), &nb3));
+            CL_CHECK(clSetKernelArg(kernel, 30, sizeof(int),      &type_src0));
+            CL_CHECK(clSetKernelArg(kernel, 31, sizeof(int),      &type_src1));
         }
-
-        CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
-        CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
-        CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
-        CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
-        CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
-        CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
-        CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
-        CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
-        CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
-        CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne03));
-        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb00));
-        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb01));
-        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb02));
-        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb03));
-        CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne10));
-        CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne11));
-        CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne12));
-        CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne13));
-        CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb10));
-        CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb11));
-        CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb12));
-        CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &nb13));
-        CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &ne0));
-        CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &ne1));
-        CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int),      &ne2));
-        CL_CHECK(clSetKernelArg(kernel, 25, sizeof(int),      &ne3));
-        CL_CHECK(clSetKernelArg(kernel, 26, sizeof(cl_ulong), &nb0));
-        CL_CHECK(clSetKernelArg(kernel, 27, sizeof(cl_ulong), &nb1));
-        CL_CHECK(clSetKernelArg(kernel, 28, sizeof(cl_ulong), &nb2));
-        CL_CHECK(clSetKernelArg(kernel, 29, sizeof(cl_ulong), &nb3));
+    } else {
+        GGML_ASSERT(false && "unsupported data types for add");
     }
 
     if (bcast_row) {
@@ -3832,13 +3881,13 @@ static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const
 
         size_t * local_work_size_ptr = local_work_size;
         if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
-            local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
+            local_work_size_ptr = nullptr;
         }
 
-        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+        backend_ctx->enqueue_ndrange_kernel(kernel, 1, global_work_size, local_work_size_ptr, dst);
     } else {
         unsigned int nth = MIN(64, ne0);
-        size_t global_work_size[] = {ne01*nth, (size_t)ne02, (size_t)ne03};
+        size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
         size_t local_work_size[] = {nth, 1, 1};
 
         backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
diff --git a/ggml/src/ggml-opencl/kernels/add.cl b/ggml/src/ggml-opencl/kernels/add.cl
index 8bc926c8893..509bf17344e 100644
--- a/ggml/src/ggml-opencl/kernels/add.cl
+++ b/ggml/src/ggml-opencl/kernels/add.cl
@@ -112,7 +112,9 @@ kernel void kernel_add_f16(
         ulong nb0,
         ulong nb1,
         ulong nb2,
-        ulong nb3
+        ulong nb3,
+        int type_src0,
+        int type_src1
 ) {
     src0 = src0 + offset0;
     src1 = src1 + offset1;
@@ -132,25 +134,57 @@ kernel void kernel_add_f16(
 
     for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
         const int i10 = i0 % ne10;
-        *((global half *)(dst_ptr + i0*nb0)) = *((global half *)(src0_ptr + i0*nb00)) + *((global half *)(src1_ptr + i10*nb10));
+
+        half v0, v1;
+        if (type_src0 == 1) {
+            v0 = convert_half(*((global float *)(src0_ptr + i0*nb00)));
+        } else {
+            v0 = *((global half *)(src0_ptr + i0*nb00));
+        }
+
+        if (type_src1 == 1) {
+            v1 = convert_half(*((global float *)(src1_ptr + i10*nb10)));
+        } else {
+            v1 = *((global half *)(src1_ptr + i10*nb10));
+        }
+
+        *((global half *)(dst_ptr + i0*nb0)) = v0 + v1;
     }
 }
 
 kernel void kernel_add_row_f16(
-        global half4 * src0,
+        global char * src0,
         ulong  offset0,
-        global half4 * src1,
+        global char * src1,
         ulong  offset1,
         global half4 * dst,
         ulong  offsetd,
-        int ne
+        int ne,
+        int type_src0,
+        int type_src1
 ) {
-    src0 = (global half4*)((global char*)src0 + offset0);
-    src1 = (global half4*)((global char*)src1 + offset1);
     dst = (global half4*)((global char*)dst + offsetd);
 
     // This performs better than using %.
     uint gid = get_global_id(0);
     uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
-    dst[gid] = src0[gid] + src1[idx1];
+
+    half4 v0, v1;
+    if (type_src0 == 1) {
+        global float4* src0_f32 = (global float4*)((global char*)src0 + offset0);
+        v0 = convert_half4(src0_f32[gid]);
+    } else {
+        global half4* src0_f16 = (global half4*)((global char*)src0 + offset0);
+        v0 = src0_f16[gid];
+    }
+
+    if (type_src1 == 1) {
+        global float4* src1_f32 = (global float4*)((global char*)src1 + offset1);
+        v1 = convert_half4(src1_f32[idx1]);
+    } else {
+        global half4* src1_f16 = (global half4*)((global char*)src1 + offset1);
+        v1 = src1_f16[idx1];
+    }
+
+    dst[gid] = v0 + v1;
 }

From 2114917d38ee2e90d8ad0e15e0a5345b3ae7bccc Mon Sep 17 00:00:00 2001
From: Romain Biessy <romain.biessy@codeplay.com>
Date: Tue, 12 Aug 2025 13:58:22 +0200
Subject: [PATCH 56/86] sycl: Fix and disable more configurations of mul_mat
 (llama/15151)

* sycl: Fix and disable more configurations of mul_mat

* Disable more configurations
---
 ggml/src/ggml-sycl/ggml-sycl.cpp | 35 +++++++++++++++++++++++++-------
 1 file changed, 28 insertions(+), 7 deletions(-)

diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
index 3992dad01d5..a0a650e92e4 100644
--- a/ggml/src/ggml-sycl/ggml-sycl.cpp
+++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
@@ -2705,9 +2705,9 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
                                                 " : converting src1 to fp16");
 
         // iterate tensor dims and find the slowest moving dim and stride
-        int64_t last_dim=0;
-        int64_t last_str=0;
-        int64_t largest_str=0;
+        int last_dim=0;
+        int last_str=0;
+        size_t largest_str=0;
         for(int i = 0; i< 4; i++){
             // last stride is always the largest
             if(src1->nb[i] == largest_str){
@@ -2783,7 +2783,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
             auto launch_gemm_for_batches = [&ctx, queue](const sycl::half *src0,
                                                 const sycl::half *src1, float *dst,
                                                 int64_t a0, int64_t a1, int64_t batcha,
-                                                int64_t b0, int64_t b1, int64_t batchb,
+                                                int64_t /*b0*/, int64_t b1, int64_t batchb,
                                                 int64_t sa0, int64_t sa1, int64_t sa2,
                                                 int64_t sb0, int64_t sb1, int64_t sb2,
                                                 int64_t sd2) {
@@ -2832,14 +2832,26 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
                 }
             };
 
-            bool cont_batches_a = nb02 * ne02 == nb03;
-            bool cont_batches_b = nb12 * ne12 == nb13;
-            if (cont_batches_a && cont_batches_b) {
+            const bool cont_batches_dim2_a = nb02 * ne02 == nb03;
+            const bool cont_batches_dim2_b = nb12 * ne12 == nb13;
+            const bool cont_batches_dim3_a = ne02 == 1 && nb02 * ne01 == nb03;
+            const bool cont_batches_dim3_b = ne12 == 1 && nb12 * ne11 == nb13;
+            if (cont_batches_dim2_a && cont_batches_dim2_b) {
+                // A batch is considered contiguous if the dimension 2 is not strided
                 int64_t batches0 = ne02 * ne03;
                 int64_t batches1 = ne12 * ne13;
                 launch_gemm_for_batches(src0_f16, src1_f16, dst_ddf, ne00, ne01, batches0,
                         ne10, ne11, batches1, str_a0, str_a1, str_a2, str_b0, str_b1,
                         str_b2, nb2 / sizeof(float));
+            } else if (cont_batches_dim3_a && cont_batches_dim3_b) {
+                // This case is similar to the one above with the difference that only the batch in dimension 3 is used and the dimension 2 is of size 1.
+                int64_t batches0 = ne02 * ne03;
+                int64_t batches1 = ne12 * ne13;
+                int64_t str_a3 = nb03 / type_size_src0;
+                int64_t str_b3 = nb13 / type_size_src1;
+                launch_gemm_for_batches(src0_f16, src1_f16, dst_ddf, ne00, ne01, batches0,
+                        ne10, ne11, batches1, str_a0, str_a1, str_a3, str_b0, str_b1,
+                        str_b3, nb2 / sizeof(float));
             } else {
                 for (int64_t b_a = 0; b_a < ne03; b_a++) {
                     const sycl::half *src0_f16_shifted
@@ -4215,6 +4227,15 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                     // FIXME: keep a list of supported types to avoid breaking the backend when a new type is added
                     return false;
                 }
+                // TODO: The configuration below needs more work to be supported with oneDNN
+                if (ggml_is_permuted(a) && !ggml_is_contiguous(a) && a->ne[2] > 1 && a->ne[3] > 1) {
+                    return false;
+                }
+                // TODO: This specific configuration can fail with oneDNN and needs more debugging
+                if (!ggml_is_permuted(a) && ggml_is_permuted(b) && b->ne[2] > 1 && b->ne[3] > 1 &&
+                    a->ne[0] > 128 && a->ne[2] == 1 && src0_type == GGML_TYPE_F16) {
+                    return false;
+                }
                 return true;
             }
         case GGML_OP_OUT_PROD:

From c9a94eaaaacb608d839ba02dfcd5348f90d9d1dd Mon Sep 17 00:00:00 2001
From: uvos <carl@uvos.xyz>
Date: Tue, 12 Aug 2025 22:15:12 +0200
Subject: [PATCH 57/86] HIP: disable sync warp shuffel operators from clr
 amd_warp_sync_functions.h (llama/15273)

---
 ggml/src/ggml-cuda/fattn-wmma-f16.cu | 1 -
 ggml/src/ggml-cuda/vendors/hip.h     | 2 +-
 2 files changed, 1 insertion(+), 2 deletions(-)

diff --git a/ggml/src/ggml-cuda/fattn-wmma-f16.cu b/ggml/src/ggml-cuda/fattn-wmma-f16.cu
index fdc4d17da2d..6bc7943ccd5 100644
--- a/ggml/src/ggml-cuda/fattn-wmma-f16.cu
+++ b/ggml/src/ggml-cuda/fattn-wmma-f16.cu
@@ -15,7 +15,6 @@ namespace wmma = mtmusa::wmma;
 namespace wmma = nvcuda::wmma;
 #endif // GGML_USE_MUSA
 #elif defined(GGML_HIP_ROCWMMA_FATTN) && defined(FP16_MMA_AVAILABLE)
-#undef HIP_ENABLE_WARP_SYNC_BUILTINS // conflicts with rocWMMA headers
 #include <rocwmma/rocwmma.hpp>
 namespace wmma = rocwmma;
 #endif // !defined(GGML_USE_HIP)
diff --git a/ggml/src/ggml-cuda/vendors/hip.h b/ggml/src/ggml-cuda/vendors/hip.h
index c31f3192322..96f8bc75e96 100644
--- a/ggml/src/ggml-cuda/vendors/hip.h
+++ b/ggml/src/ggml-cuda/vendors/hip.h
@@ -1,6 +1,6 @@
 #pragma once
 
-#define HIP_ENABLE_WARP_SYNC_BUILTINS 1
+#define HIP_DISABLE_WARP_SYNC_BUILTINS 1
 #include <hip/hip_runtime.h>
 #include <hipblas/hipblas.h>
 #include <hip/hip_fp16.h>

From 77938f9be4a7c337592489e599d7ac37346ea0f6 Mon Sep 17 00:00:00 2001
From: Tak-RS <snosk.t@gmail.com>
Date: Wed, 13 Aug 2025 14:54:30 +0900
Subject: [PATCH 58/86] ggml-rpc: chunk send()/recv() to avoid EINVAL for very
 large tensors over RPC (macOS & others) (llama/15188)

* ggml-rpc: chunk send()/recv() to avoid EINVAL for very large tensors over RPC (macOS & others). Fixes #15055

* ggml-rpc: rename RPC_IO_CHUNK->MAX_CHUNK_SIZE, use std::min() for cap, switch to GGML_LOG_ERROR, handle 0-length send/recv

* rpc: drop n==0 special case in send_data(); retry in loop per review

* rpc: remove trailing whitespace in send_data()

---------

Co-authored-by: Shinnosuke Takagi <nosuke@nosukenoMacBook-Pro.local>
---
 ggml/src/ggml-rpc/ggml-rpc.cpp | 23 ++++++++++++++++++-----
 1 file changed, 18 insertions(+), 5 deletions(-)

diff --git a/ggml/src/ggml-rpc/ggml-rpc.cpp b/ggml/src/ggml-rpc/ggml-rpc.cpp
index df6ba54076d..e84ff93efc3 100644
--- a/ggml/src/ggml-rpc/ggml-rpc.cpp
+++ b/ggml/src/ggml-rpc/ggml-rpc.cpp
@@ -29,9 +29,12 @@
 #include <cstring>
 #include <fstream>
 #include <filesystem>
+#include <algorithm>
 
 namespace fs = std::filesystem;
 
+static constexpr size_t MAX_CHUNK_SIZE = 1024ull * 1024ull * 1024ull; // 1 GiB
+
 #ifdef _WIN32
 typedef SOCKET sockfd_t;
 using ssize_t = __int64;
@@ -323,11 +326,14 @@ static std::shared_ptr<socket_t> create_server_socket(const char * host, int por
 static bool send_data(sockfd_t sockfd, const void * data, size_t size) {
     size_t bytes_sent = 0;
     while (bytes_sent < size) {
-        ssize_t n = send(sockfd, (const char *)data + bytes_sent, size - bytes_sent, 0);
+        size_t size_to_send = std::min(size - bytes_sent, MAX_CHUNK_SIZE);
+        ssize_t n = send(sockfd, (const char *)data + bytes_sent, size_to_send, 0);
         if (n < 0) {
+            GGML_LOG_ERROR("send failed (bytes_sent=%zu, size_to_send=%zu)\n",
+                           bytes_sent, size_to_send);
             return false;
         }
-        bytes_sent += n;
+        bytes_sent += (size_t)n;
     }
     return true;
 }
@@ -335,11 +341,18 @@ static bool send_data(sockfd_t sockfd, const void * data, size_t size) {
 static bool recv_data(sockfd_t sockfd, void * data, size_t size) {
     size_t bytes_recv = 0;
     while (bytes_recv < size) {
-        ssize_t n = recv(sockfd, (char *)data + bytes_recv, size - bytes_recv, 0);
-        if (n <= 0) {
+        size_t size_to_recv = std::min(size - bytes_recv, MAX_CHUNK_SIZE);
+        ssize_t n = recv(sockfd, (char *)data + bytes_recv, size_to_recv, 0);
+        if (n < 0) {
+            GGML_LOG_ERROR("recv failed (bytes_recv=%zu, size_to_recv=%zu)\n",
+                           bytes_recv, size_to_recv);
+            return false;
+        }
+        if (n == 0) {
+            GGML_LOG_ERROR("recv returned 0 (peer closed?)\n");
             return false;
         }
-        bytes_recv += n;
+        bytes_recv += (size_t)n;
     }
     return true;
 }

From 0b0748f98186f2517ceef6874b45b7214059d550 Mon Sep 17 00:00:00 2001
From: Oliver Simons <osimons@nvidia.com>
Date: Wed, 13 Aug 2025 10:04:46 +0200
Subject: [PATCH 59/86] CUDA: Optimize `reduce_rows_f32` kernel, leading up to
 25x perf improvement on kernel-level and 10% perf increase for Gemma3n
 (llama/15132)

* Factor out `reduce_rows_f32` from common.cuh

This increases iteration cycle speed by not having to recompile
every kernel all the time

* Hide memory-latency by loop unrolling in reduce_rows_f32

* Further optimizations to `reduce_rows_f32`

1. Increase threadblock size to better hide latency of memory requests.
   As a consequence of bigger threadblocks, do 2-step summation, using
   shared memory to communicate results between invocations
2. Use sum_temp array to reduce waits on sum
3. Adjust num_unroll to reflext bigger threadblock
4. Improve default block_dims, increase support for more block_dims

* Add perf tests for `reduce_rows_f32` kernel

* Add heuristic to toggle 128/512 threads based on sm count

Break even point was the minimum of the following multiples.

| GPU Model                     | Nrow SM Count Multiple |
| -----------                   | -----------            |
| RTX 4000 SFF ADA              | 2.0x                   |
| RTX 6000 ADA                  | 2.5x                   |
| RTX PRO 6000 Blackwell Max-Q  | 3.04x                  |
| RTX PRO 4500 Blackwell	| 3.15x                  |

* Ensure perf gains also for small ncols and large nrows

Alternative to this, one could have also made the number of unrollings
template-able, but that would require compiling the kernel multiple
times, increasing binary size unnecessarily

* Modify perf and unit-tests

* Apply auto-formatting by clang

* Fix CI build failure

See https://github.com/ggml-org/llama.cpp/actions/runs/16798370266/job/47573716079?pr=15132#step:7:486
Building with VS generator worked though.

* Remove sm_count property from `ggml_backend_cuda_context`

Requested by @JohannesGaessler, and should fix remaining CI issues as a
side-effect

* Add CUB-based implementation for GGML_OP_MEAN

Currently this branch is only executed for nrows==1

* Add heuristics to execute CUB branch only when it brings perf

Heuristics were determined on the following HW:

* RTX 4000 SFF ADA
* RTX 6000 ADA
* RTX PRO 6000 Blackwell Max-Q
* RTX PRO 4500 Blackwell

* Add unit-test for CUB-based mean

Tests should run with CUDA Graphs enabled per default on NVGPUs

* Rename `USE_CUB` to `GGML_CUDA_USE_CUB`

Suggested by @JohannesGaessler

* Unindent Preprocessor directives

See
https://github.com/ggml-org/llama.cpp/pull/15132#discussion_r2269213506
---
 ggml/src/ggml-cuda/common.cuh      | 24 +++-----------
 ggml/src/ggml-cuda/mean.cu         | 52 +++++++++++++++++++++++++++--
 ggml/src/ggml-cuda/reduce_rows.cuh | 53 ++++++++++++++++++++++++++++++
 ggml/src/ggml-cuda/sum.cu          | 16 ++++-----
 ggml/src/ggml-cuda/sumrows.cu      | 25 +++++++++++---
 5 files changed, 134 insertions(+), 36 deletions(-)
 create mode 100644 ggml/src/ggml-cuda/reduce_rows.cuh

diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
index a23da57e3a1..5a2a3478d26 100644
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -87,6 +87,10 @@
 #define GGML_CUDA_CC_IS_QY2(cc)      (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_NG)
 #define GGML_CUDA_CC_IS_NG(cc)       (cc >= GGML_CUDA_CC_NG)
 
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
+#    define GGML_CUDA_USE_CUB
+#endif  // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
+
 #ifdef __CUDA_ARCH_LIST__
 constexpr bool ggml_cuda_has_arch_impl(int) {
     return false;
@@ -420,26 +424,6 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
 #endif // FP16_AVAILABLE
 }
 
-// Row reduction kernel template - compute sum (norm=false) or mean (norm=true)
-template<bool norm>
-static __global__ void reduce_rows_f32(const float * x, float * dst, const int ncols) {
-    const int row = blockIdx.x;
-    const int col = threadIdx.x;
-
-    float sum = 0.0f;
-    for (int i = col; i < ncols; i += blockDim.x) {
-        sum += x[row * ncols + i];
-    }
-
-    sum = warp_reduce_sum(sum);
-
-    if (col != 0) {
-        return;
-    }
-
-    dst[row] = norm ? sum / ncols : sum;
-}
-
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ int warp_reduce_all(int x) {
 #ifdef GGML_USE_HIP
diff --git a/ggml/src/ggml-cuda/mean.cu b/ggml/src/ggml-cuda/mean.cu
index 4b238a3998b..2ad493239b1 100644
--- a/ggml/src/ggml-cuda/mean.cu
+++ b/ggml/src/ggml-cuda/mean.cu
@@ -1,4 +1,14 @@
 #include "mean.cuh"
+#include "reduce_rows.cuh"
+
+#ifdef GGML_CUDA_USE_CUB
+#include <cub/cub.cuh>
+using namespace cub;
+#endif  // GGML_CUDA_USE_CUB
+
+template <typename T> __global__ void divide_by_count(T * result, size_t count) {
+    *result /= static_cast<T>(count);
+}
 
 void ggml_cuda_op_mean(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0   = dst->src[0];
@@ -13,7 +23,45 @@ void ggml_cuda_op_mean(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const int64_t ncols = src0->ne[0];
     const int64_t nrows = ggml_nrows(src0);
 
-    const dim3 block_dims(WARP_SIZE, 1, 1);
+// Special case for reducing vectors
+#ifdef GGML_CUDA_USE_CUB
+    cudaStreamCaptureStatus iscapturing;
+    CUDA_CHECK(cudaStreamIsCapturing(stream, &iscapturing));
+    if ((nrows == 1) &&
+            // CUDA_GRAPHS_DISABLED
+            ((ncols > 65536) &&
+             ((ctx.cuda_graph->instance == nullptr) && (iscapturing == cudaStreamCaptureStatusNone) ||
+              ctx.cuda_graph->disable_due_to_gpu_arch || ctx.cuda_graph->disable_due_to_too_many_updates ||
+              ctx.cuda_graph->disable_due_to_failed_graph_capture)) ||
+        // CUDA_GRAPHS ENABLED
+        ((ncols > 32768) &&
+         !((ctx.cuda_graph->instance == nullptr) && (iscapturing == cudaStreamCaptureStatusNone) ||
+           ctx.cuda_graph->disable_due_to_gpu_arch || ctx.cuda_graph->disable_due_to_too_many_updates ||
+           ctx.cuda_graph->disable_due_to_failed_graph_capture))) {
+        // Single row - use device-wide reduction
+        size_t           tmp_size = 0;
+        ggml_cuda_pool & pool     = ctx.pool();
+
+        DeviceReduce::Sum(nullptr, tmp_size, src0_d, dst_d, ncols, stream);
+
+        ggml_cuda_pool_alloc<uint8_t> tmp_alloc(pool, tmp_size);
+        DeviceReduce::Sum(tmp_alloc.ptr, tmp_size, src0_d, dst_d, ncols, stream);
+
+        // Divide by ncols
+        divide_by_count<float><<<1, 1, 0, stream>>>(dst_d, ncols);
+        return;
+    }
+#endif
+
     const dim3 block_nums(nrows, 1, 1);
-    reduce_rows_f32</*norm*/ true><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+
+    const int id  = ggml_cuda_get_device();
+    const int nsm = ggml_cuda_info().devices[id].nsm;
+    if ((nrows / nsm) < 2) {
+        const dim3 block_dims(512, 1, 1);
+        reduce_rows_f32</*norm=*/true><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+    } else {
+        const dim3 block_dims(ncols < 1024 ? 32 : 128, 1, 1);
+        reduce_rows_f32</*norm=*/true><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+    }
 }
diff --git a/ggml/src/ggml-cuda/reduce_rows.cuh b/ggml/src/ggml-cuda/reduce_rows.cuh
new file mode 100644
index 00000000000..6bee204136b
--- /dev/null
+++ b/ggml/src/ggml-cuda/reduce_rows.cuh
@@ -0,0 +1,53 @@
+#include "common.cuh"
+
+// Row reduction kernel template - compute sum (norm=false) or mean (norm=true)
+template <bool norm>
+static __global__ void reduce_rows_f32(const float * __restrict__ x, float * __restrict__ dst, const int ncols) {
+    const int row = blockIdx.x;
+    const int col = threadIdx.x;
+
+    float     sum        = 0.0f;
+    const int num_unroll = 8;
+    float     temp[num_unroll];
+    float     sum_temp[num_unroll] = { 0.0f };
+    for (int i = col; i < ncols;) {
+        for (int j = 0; j < num_unroll; ++j) {
+            if (i < ncols) {
+                temp[j] = x[row * ncols + i];
+            } else {
+                temp[j] = 0;
+            }
+            i += blockDim.x;
+        }
+        for (int j = 0; j < num_unroll; ++j) {
+            sum_temp[j] += temp[j];
+        }
+    }
+    for (int j = 0; j < num_unroll; ++j) {
+        sum += sum_temp[j];
+    }
+
+    // sum up partial sums
+    sum = warp_reduce_sum(sum);
+    if (blockDim.x > WARP_SIZE) {
+        assert((blockDim.x <= 1024) && (blockDim.x % WARP_SIZE) == 0);
+        __shared__ float s_sum[32];
+        const int        warp_id = threadIdx.x / WARP_SIZE;
+        const int        lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = sum;
+        }
+        __syncthreads();
+        sum = 0.0f;
+        if (lane_id < (blockDim.x / WARP_SIZE)) {
+            sum = s_sum[lane_id];
+        }
+        sum = warp_reduce_sum(sum);
+    }
+
+    if (col != 0) {
+        return;
+    }
+
+    dst[row] = norm ? sum / ncols : sum;
+}
diff --git a/ggml/src/ggml-cuda/sum.cu b/ggml/src/ggml-cuda/sum.cu
index eb3d7cdba98..c56257b4406 100644
--- a/ggml/src/ggml-cuda/sum.cu
+++ b/ggml/src/ggml-cuda/sum.cu
@@ -1,19 +1,15 @@
-#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
-#define USE_CUB
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
+#include "sum.cuh"
+#include "sumrows.cuh"
 
-#ifdef USE_CUB
+#ifdef GGML_CUDA_USE_CUB
 #include <cub/cub.cuh>
 using namespace cub;
-#endif // USE_CUB
-
-#include "sumrows.cuh"
-#include "sum.cuh"
+#endif  // GGML_CUDA_USE_CUB
 
 #include <cstdint>
 
 void sum_f32_cuda(ggml_cuda_pool & pool, const float * x, float * dst, const int64_t ne, cudaStream_t stream) {
-#ifdef USE_CUB
+#ifdef GGML_CUDA_USE_CUB
     size_t tmp_size = 0;
     DeviceReduce::Sum(nullptr,       tmp_size, x, dst, ne, stream);
     ggml_cuda_pool_alloc<uint8_t> tmp_alloc(pool, tmp_size);
@@ -23,7 +19,7 @@ void sum_f32_cuda(ggml_cuda_pool & pool, const float * x, float * dst, const int
     // For AMD there is rocPRIM which could be used as a drop-in replacement via hipcub but this would require C++11 -> C++14.
     sum_rows_f32_cuda(x, dst, ne, 1, stream);
     GGML_UNUSED(pool);
-#endif // USE_CUB
+#endif // GGML_CUDA_USE_CUB
 }
 
 void ggml_cuda_op_sum(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
diff --git a/ggml/src/ggml-cuda/sumrows.cu b/ggml/src/ggml-cuda/sumrows.cu
index 2eee08fa073..4025771aadb 100644
--- a/ggml/src/ggml-cuda/sumrows.cu
+++ b/ggml/src/ggml-cuda/sumrows.cu
@@ -1,9 +1,17 @@
+#include "reduce_rows.cuh"
 #include "sumrows.cuh"
 
 void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    const dim3 block_dims(WARP_SIZE, 1, 1);
+    const int  id  = ggml_cuda_get_device();
+    const int  nsm = ggml_cuda_info().devices[id].nsm;
     const dim3 block_nums(nrows, 1, 1);
-    reduce_rows_f32</*norm*/false><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+    if ((nrows / nsm) < 2) {
+        const dim3 block_dims(512, 1, 1);
+        reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+    } else {
+        const dim3 block_dims(ncols < 1024 ? 32 : 128, 1, 1);
+        reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+    }
 }
 
 void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -19,8 +27,17 @@ void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const int64_t ncols = src0->ne[0];
     const int64_t nrows = ggml_nrows(src0);
 
-    const dim3 block_dims(WARP_SIZE, 1, 1);
     const dim3 block_nums(nrows, 1, 1);
 
-    reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+    const int id  = ggml_cuda_get_device();
+    const int nsm = ggml_cuda_info().devices[id].nsm;
+    if ((nrows / nsm) < 2) {
+        // Increase num threads to 512 for small nrows to better hide the latency
+        const dim3 block_dims(512, 1, 1);
+        reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+    } else {
+        // Enough active SMs to hide latency, use smaller blocks to allow better scheduling
+        const dim3 block_dims(ncols < 1024 ? 32 : 128, 1, 1);
+        reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+    }
 }

From 16088b510f68220288f3fc081954aa456900394a Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Wed, 13 Aug 2025 11:09:39 +0300
Subject: [PATCH 60/86] ggml : repack block_iq4_nlx8 (llama/14904)

ggml-ci
---
 ggml/src/ggml-cpu/arch-fallback.h     |   14 +
 ggml/src/ggml-cpu/arch/x86/repack.cpp | 2213 +++++++++++++------------
 ggml/src/ggml-cpu/repack.cpp          |  251 ++-
 ggml/src/ggml-cpu/repack.h            |   11 +
 4 files changed, 1361 insertions(+), 1128 deletions(-)

diff --git a/ggml/src/ggml-cpu/arch-fallback.h b/ggml/src/ggml-cpu/arch-fallback.h
index b62e3158ee7..f476127995b 100644
--- a/ggml/src/ggml-cpu/arch-fallback.h
+++ b/ggml/src/ggml-cpu/arch-fallback.h
@@ -40,18 +40,22 @@
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64)
 // repack.cpp
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
 // repack.cpp
@@ -80,12 +84,14 @@
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #elif defined(__loongarch64)
 // quants.c
 #define quantize_row_q8_K_generic quantize_row_q8_K
@@ -103,12 +109,14 @@
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #elif defined(__riscv)
 // quants.c
 #define quantize_row_q8_K_generic quantize_row_q8_K
@@ -133,11 +141,13 @@
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #elif defined(__s390x__)
 // quants.c
 #define quantize_row_q8_K_generic quantize_row_q8_K
@@ -164,12 +174,14 @@
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #elif defined(__wasm__)
 // quants.c
 #define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
@@ -195,10 +207,12 @@
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #endif
diff --git a/ggml/src/ggml-cpu/arch/x86/repack.cpp b/ggml/src/ggml-cpu/arch/x86/repack.cpp
index 37933a4b23f..d95bb6d8aaf 100644
--- a/ggml/src/ggml-cpu/arch/x86/repack.cpp
+++ b/ggml/src/ggml-cpu/arch/x86/repack.cpp
@@ -511,38 +511,34 @@ void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
 #endif
 }
 
-void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+//
+// GEMV/GEMM templates
+//
+
+#if defined(__AVX2__) || defined(__AVX512F__)
+
+// GEMV for 8x blocks of 32 4-bit quants with a single scale factor per block
+template<typename block_tx8>
+static void gemv_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc, __m256i signextendlut) {
+    static_assert(
+            std::is_same_v<block_tx8, block_q4_0x8> ||
+            std::is_same_v<block_tx8, block_iq4_nlx8>,
+            "Unsupported block type");
+
     const int qk = QK8_0;
     const int nb = n / qk;
-    const int ncols_interleaved = 8;
-    const int blocklen = 8;
-
-    assert (n % qk == 0);
-    assert (nc % ncols_interleaved == 0);
 
-    UNUSED(s);
     UNUSED(bs);
-    UNUSED(vx);
-    UNUSED(vy);
-    UNUSED(nr);
-    UNUSED(nc);
-    UNUSED(nb);
-    UNUSED(ncols_interleaved);
-    UNUSED(blocklen);
 
-#if defined(__AVX2__)
-    // Lookup table to convert signed nibbles to signed bytes
-    __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
-    signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
     __m128i changemask = _mm_set_epi8(15, 14, 7, 6, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0);
     __m256i finalpermutemask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
 
     // Permute mask used for easier vector processing at later stages
     const __m256i m4b = _mm256_set1_epi8(0x0F);
 
-    int64_t b_nb = n / QK4_0;
+    int64_t b_nb = n / 32;
 
-    const block_q4_0x8 * b_ptr_start = (const block_q4_0x8 *)vx;
+    const block_tx8  * b_ptr_start = (const block_tx8  *)vx;
     const block_q8_0 * a_ptr_start = (const block_q8_0 *)vy;
 
     // Process Q8_0 blocks one by one
@@ -551,17 +547,17 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
         // Pointers to LHS blocks of block_q8_0 format
         const block_q8_0 * a_ptr = a_ptr_start + (y * nb);
 
-        // Take group of eight block_q4_0x8 structures at each pass of the loop and perform dot product operation
+        // Take group of eight blocks at each pass of the loop and perform dot product operation
         for (int64_t x = 0; x < nc / 8; x++) {
 
             // Pointers to RHS blocks
-            const block_q4_0x8 * b_ptr = b_ptr_start + (x * b_nb);
+            const block_tx8 * b_ptr = b_ptr_start + (x * b_nb);
 
             // Master FP accumulator
             __m256 acc_row = _mm256_setzero_ps();
 
             for (int64_t b = 0; b < nb; b++) {
-                // Load 8 blocks of Q4_0 interleaved as 8 bytes (B0 - B7)
+                // Load 8 blocks of 32 interleaved as 8 bytes (B0 - B7)
                 const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
                 const __m256i rhs_raw_vec_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs) + 1);
                 const __m256i rhs_raw_vec_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs) + 2);
@@ -578,8 +574,13 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                 const __m256i rhs_vec_0123_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m4b)); // B0(24-31) B1(24-31) B2(24-31) B3(24-31)
                 const __m256i rhs_vec_4567_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m4b)); // B4(24-31) B5(24-31) B6(24-31) B7(24-31)
 
-                // Load the scale values for the 8 blocks interleaved in block_q4_0x8
-                const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, changemask);
+                // Load the scale values for the 8 blocks interleaved in block_tx8
+                __m256 col_scale_f32;
+                if constexpr (
+                        std::is_same_v<block_tx8, block_q4_0x8> ||
+                        std::is_same_v<block_tx8, block_iq4_nlx8>) {
+                    col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, changemask);
+                }
 
                 // Load and convert to FP32 scale from block_q8_0
                 const __m256 row_scale_f32 = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(a_ptr[b].d));
@@ -620,240 +621,782 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
             _mm256_storeu_ps(s + (y * nr + x * 8), acc_row);
         }
     }
-    return;
-
-#endif
-    ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
-void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
-    const int qk = QK_K;
-    const int nb = n / qk;
-    const int ncols_interleaved = 8;
-    const int blocklen = 8;
-    static const uint32_t kmask1 = 0x3f3f3f3f;
-    static const uint32_t kmask2 = 0x0f0f0f0f;
-    static const uint32_t kmask3 = 0x03030303;
-
-    assert (n % qk == 0);
-    assert (nc % ncols_interleaved == 0);
+// GEMM for 8x blocks of 32 4-bit quants with a single scale factor per block
+template<typename block_tx8>
+static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc, __m256i signextendlut) {
+    static_assert(
+            std::is_same_v<block_tx8, block_q4_0x8> ||
+            std::is_same_v<block_tx8, block_iq4_nlx8>,
+            "Unsupported block type");
 
-    UNUSED(s);
-    UNUSED(bs);
-    UNUSED(vx);
-    UNUSED(vy);
-    UNUSED(nr);
-    UNUSED(nc);
-    UNUSED(nb);
-    UNUSED(ncols_interleaved);
-    UNUSED(blocklen);
+    const int qk = QK8_0;
+    const int nb = n / qk;
 
-#if defined(__AVX2__)
-    // Lookup table to convert signed nibbles to signed bytes
-    __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
-    signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
-    // Shuffle masks to rearrange delta and scale values to multiply with appropriate scales
-    __m128i deltamask = _mm_set_epi8(15, 14, 7, 6, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0);
-    __m128i scalemask = _mm_set_epi8(7, 7, 3, 3, 6, 6, 2, 2, 5, 5, 1, 1, 4, 4, 0, 0);
-    // Permute mask used for easier vector processing at later stages
-    __m256i finalpermutemask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
+    const block_tx8    * b_ptr_start = (const block_tx8    *)vx;
+    const block_q8_0x4 * a_ptr_start = (const block_q8_0x4 *)vy;
 
-    // Mask to extract nibbles from bytes
+    int64_t b_nb = n / 32;
+    int64_t y = 0;
+    // Mask to mask out nibbles from packed bytes
     const __m256i m4b = _mm256_set1_epi8(0x0F);
+    const __m128i loadMask = _mm_blend_epi32(_mm_setzero_si128(), _mm_set1_epi32(0xFFFFFFFF), 3);
+    // Permute mask used for easier vector processing at later stages
+    __m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4);
+    int64_t xstart = 0;
+    int anr = nr - nr%16; // Used to align nr with boundary of 16
+#ifdef __AVX512F__
+    int anc = nc - nc%16; // Used to align nc with boundary of 16
+                          // Mask to mask out nibbles from packed bytes expanded to 512 bit length
+    const __m512i m4bexpanded = _mm512_set1_epi8(0x0F);
+    // Lookup table to convert signed nibbles to signed bytes expanded to 512 bit length
+    __m512i signextendlutexpanded = _mm512_inserti32x8(_mm512_castsi256_si512(signextendlut), signextendlut, 1);
 
-    int64_t b_nb = n / QK_K;
-
-    const block_q4_Kx8 * b_ptr_start = (const block_q4_Kx8 *)vx;
-    const block_q8_K * a_ptr_start = (const block_q8_K *)vy;
+    // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation
+    for (; y < anr / 4; y += 4) {
 
-    // Process Q8_K blocks one by one
-    for (int64_t y = 0; y < nr; y++) {
+        const block_q8_0x4 * a_ptrs[4];
 
-        // Pointers to LHS blocks of block_q8_K format
-        const block_q8_K * a_ptr = a_ptr_start + (y * nb);
+        a_ptrs[0] = a_ptr_start + (y * nb);
+        for (int i = 0; i < 3; ++i) {
+            a_ptrs[i + 1] = a_ptrs[i] + nb;
+        }
 
-        // Take group of eight interleaved block_q4_K structures at each pass of the loop and perform dot product operation
-        for (int64_t x = 0; x < nc / 8; x++) {
+        // Take group of two block_tx8 structures at each pass of the loop and perform dot product operation
+        for (int64_t x = 0; x < anc / 8; x += 2) {
 
-            // Pointers to RHS blocks
-            const block_q4_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
+            const block_tx8 * b_ptr_0 = b_ptr_start + ((x)     * b_nb);
+            const block_tx8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
 
             // Master FP accumulators
-            __m256 acc_row = _mm256_setzero_ps();
-            __m256 acc_min_rows = _mm256_setzero_ps();
+            __m512 acc_rows[16];
+            for (int i = 0; i < 16; i++) {
+                acc_rows[i] = _mm512_setzero_ps();
+            }
 
             for (int64_t b = 0; b < nb; b++) {
+                // Load the sixteen blocks of quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF
+                const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs));
+                const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32));
+                const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64));
+                const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96));
+
+                const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs));
+                const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32));
+                const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64));
+                const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96));
+
+                // Save the values in the following vectors in the formats B0B1B4B5B8B9BCBD, B2B3B6B7BABBBEBF for further processing and storing of values
+                const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+                const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+                const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+                const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+
+                const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240);
+                const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240);
+                const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240);
+                const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240);
+
+                const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
+                const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
+                const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
+                const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
 
-                // Load and convert to FP32 scale from block_q8_K
-                const __m256 row_scale_f32 = _mm256_set1_ps((a_ptr[b].d));
-
-                // Load the scale values for the 8 blocks interleaved in block_q4_Kx8
-                // col_scale_f32 rearranged so as to multiply with appropriate quants
-                const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, deltamask);
-                const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin);
+                // 4-bit -> 8-bit - Sign is maintained
+                const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7)
+                const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7)
 
-                __m256i iacc_b = _mm256_setzero_si256();
-                __m256i iacc_min_b = _mm256_setzero_si256();
+                const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15)
+                const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15)
 
-                const __m256i q8sums = _mm256_loadu_si256((const __m256i * )(a_ptr[b].bsums));
-                __m256i q8s = _mm256_castsi128_si256(_mm_hadd_epi16(_mm256_castsi256_si128(q8sums), _mm256_extracti128_si256(q8sums, 1)));
-                q8s = _mm256_permute2f128_si256(q8s, q8s, 0);
+                const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23)
+                const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23)
 
-                // Processes two sub blocks from each Q4_K in each iteration
-                for (int sb = 0; sb < QK_K / 64; sb++) {
+                const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31)
+                const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31)
 
-                    // Load the eight block_q4_K for two sub blocks quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
-                    const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + sb * 256));
-                    const __m256i rhs_raw_vec_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 32 + sb * 256));
-                    const __m256i rhs_raw_vec_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 64 + sb * 256));
-                    const __m256i rhs_raw_vec_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 96 + sb * 256));
-                    const __m256i rhs_raw_vec_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 128 + sb * 256));
-                    const __m256i rhs_raw_vec_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 160 + sb * 256));
-                    const __m256i rhs_raw_vec_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 192 + sb * 256));
-                    const __m256i rhs_raw_vec_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 224 + sb * 256));
+                // Shuffle pattern one - right side input
+                const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
+                const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
 
-                    // 4-bit -> 8-bit
-                    // Values of the first sub block of eight block_q4_K structures for the sb loop
-                    const __m256i rhs_vec_0123_00 = _mm256_and_si256(rhs_raw_vec_0123_0, m4b);
-                    const __m256i rhs_vec_4567_00 = _mm256_and_si256(rhs_raw_vec_4567_0, m4b);
-                    const __m256i rhs_vec_0123_01 = _mm256_and_si256(rhs_raw_vec_0123_1, m4b);
-                    const __m256i rhs_vec_4567_01 = _mm256_and_si256(rhs_raw_vec_4567_1, m4b);
-                    const __m256i rhs_vec_0123_02 = _mm256_and_si256(rhs_raw_vec_0123_2, m4b);
-                    const __m256i rhs_vec_4567_02 = _mm256_and_si256(rhs_raw_vec_4567_2, m4b);
-                    const __m256i rhs_vec_0123_03 = _mm256_and_si256(rhs_raw_vec_0123_3, m4b);
-                    const __m256i rhs_vec_4567_03 = _mm256_and_si256(rhs_raw_vec_4567_3, m4b);
+                const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
+                const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
 
-                    // Values of the second sub block of eight block_q4_K structures when sb = 1
-                    const __m256i rhs_vec_0123_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 4), m4b);
-                    const __m256i rhs_vec_4567_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 4), m4b);
-                    const __m256i rhs_vec_0123_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m4b);
-                    const __m256i rhs_vec_4567_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m4b);
-                    const __m256i rhs_vec_0123_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 4), m4b);
-                    const __m256i rhs_vec_4567_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 4), m4b);
-                    const __m256i rhs_vec_0123_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 4), m4b);
-                    const __m256i rhs_vec_4567_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 4), m4b);
+                const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
+                const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
 
-                    uint32_t utmp_0[4], utmp_1[4];
+                const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
+                const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
 
-                    // Scales and Mins of corresponding sub blocks from different Q8_K structures are stored together
-                    // The below block is for eg to extract first sub block's scales and mins from different Q4_K structures for the sb loop
-                    memcpy(utmp_0, b_ptr[b].scales + 24 * sb, 12);
-                    utmp_0[3] = ((utmp_0[2] >> 4) & kmask2) | (((utmp_0[1] >> 6) & kmask3) << 4);
-                    const uint32_t uaux_0 = utmp_0[1] & kmask1;
-                    utmp_0[1] = (utmp_0[2] & kmask2) | (((utmp_0[0] >> 6) & kmask3) << 4);
-                    utmp_0[2] = uaux_0;
-                    utmp_0[0] &= kmask1;
+                // Shuffle pattern two - right side input
 
-                    // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop
-                    memcpy(utmp_1, b_ptr[b].scales + 12 + sb * 24, 12);
-                    utmp_1[3] = ((utmp_1[2] >> 4) & kmask2) | (((utmp_1[1] >> 6) & kmask3) << 4);
-                    const uint32_t uaux_1 = utmp_1[1] & kmask1;
-                    utmp_1[1] = (utmp_1[2] & kmask2) | (((utmp_1[0] >> 6) & kmask3) << 4);
-                    utmp_1[2] = uaux_1;
-                    utmp_1[0] &= kmask1;
+                const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
+                const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
 
-                    // Scales of first sub block in the sb loop
-                    const __m128i mins_and_scales_0 = _mm_set_epi32(utmp_0[3], utmp_0[2], utmp_0[1], utmp_0[0]);
-                    __m128i scales_rearrange_0 = _mm_shuffle_epi8(mins_and_scales_0, scalemask);
-                    __m256i scales_0 = _mm256_cvtepu8_epi16(scales_rearrange_0);
+                const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
+                const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
 
-                    // Scales of second sub block in the sb loop
-                    __m128i mins_and_scales_1 = _mm_set_epi32(utmp_1[3], utmp_1[2], utmp_1[1], utmp_1[0]);
-                    __m128i scales_rearrange_1 = _mm_shuffle_epi8(mins_and_scales_1, scalemask);
-                    __m256i scales_1 = _mm256_cvtepu8_epi16(scales_rearrange_1);
+                const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
+                const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
 
-                    // Mins of first and second sub block of Q4_K block are arranged side by side
-                    __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_unpacklo_epi8(_mm_shuffle_epi32(mins_and_scales_0, 78), _mm_shuffle_epi32(mins_and_scales_1, 78)));
+                const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
+                const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
 
-                    // Load the two sub block values corresponding to sb in block_q8_K in batches of 16 bytes and replicate the same across 256 bit vector
-                    __m256i lhs_vec_00 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + sb * 64)));
-                    __m256i lhs_vec_01 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 16 + sb * 64)));
-                    __m256i lhs_vec_10 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 32 + sb * 64)));
-                    __m256i lhs_vec_11 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 48 + sb * 64)));
+                // Scale values - Load the weight scale values of two block_tx8
+                __m512 col_scale_f32;
+                if constexpr (
+                        std::is_same_v<block_tx8, block_q4_0x8> ||
+                        std::is_same_v<block_tx8, block_iq4_nlx8>) {
+                    col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
+                }
 
-                    lhs_vec_00 = _mm256_permute2f128_si256(lhs_vec_00, lhs_vec_00, 0);
-                    lhs_vec_01 = _mm256_permute2f128_si256(lhs_vec_01, lhs_vec_01, 0);
-                    lhs_vec_10 = _mm256_permute2f128_si256(lhs_vec_10, lhs_vec_10, 0);
-                    lhs_vec_11 = _mm256_permute2f128_si256(lhs_vec_11, lhs_vec_11, 0);
+                // Process LHS in pairs of rows
+                for (int rp = 0; rp < 4; rp++) {
 
-                    // Dot product done within 32 bit lanes and accumulated in the same vector
-                    // First done for first sub block and thenn for second sub block in each sb
-                    // B0(0-3) B4(0-3) B1(0-3) B5(0-3) B2(0-3) B6(0-3) B3(0-3) B7(0-3) with A0(0-3)
-                    // B0(4-7) B4(4-7) B1(4-7) B5(4-7) B2(4-7) B6(4-7) B3(4-7) B7(4-7) with A0(4-7)
-                    // ...........................................................................
-                    // B0(28-31) B4(28-31) B1(28-31) B5(28-31) B2(28-31) B6(28-31) B3(28-31) B7(28-31) with A0(28-31)
+                    // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
+                    // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector
+                    __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs)));
+                    __m256i lhs_mat_ymm_01_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 0);
+                    __m256i lhs_mat_ymm_23_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 17);
+                    __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32)));
+                    __m256i lhs_mat_ymm_01_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 0);
+                    __m256i lhs_mat_ymm_23_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 17);
+                    __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64)));
+                    __m256i lhs_mat_ymm_01_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 0);
+                    __m256i lhs_mat_ymm_23_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 17);
+                    __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96)));
+                    __m256i lhs_mat_ymm_01_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 0);
+                    __m256i lhs_mat_ymm_23_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 17);
 
+                    __m512i lhs_mat_01_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_0), lhs_mat_ymm_01_0, 1);
+                    __m512i lhs_mat_23_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_0), lhs_mat_ymm_23_0, 1);
+                    __m512i lhs_mat_01_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_1), lhs_mat_ymm_01_1, 1);
+                    __m512i lhs_mat_23_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_1), lhs_mat_ymm_23_1, 1);
+                    __m512i lhs_mat_01_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_2), lhs_mat_ymm_01_2, 1);
+                    __m512i lhs_mat_23_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_2), lhs_mat_ymm_23_2, 1);
+                    __m512i lhs_mat_01_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_3), lhs_mat_ymm_01_3, 1);
+                    __m512i lhs_mat_23_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_3), lhs_mat_ymm_23_3, 1);
 
-                    __m256i iacc_0 = _mm256_setzero_si256();
-                    __m256i iacc_1 = _mm256_setzero_si256();
+                    // Shuffle pattern one - left side input
 
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_00 ,_mm256_shuffle_epi32(rhs_vec_4567_00, 177), 170), _mm256_shuffle_epi32(lhs_vec_00, 0)));
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_00, 177) ,rhs_vec_4567_00, 170), _mm256_shuffle_epi32(lhs_vec_00, 85)));
+                    const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
+                    const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
 
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_01 ,_mm256_shuffle_epi32(rhs_vec_4567_01, 177), 170), _mm256_shuffle_epi32(lhs_vec_00, 170)));
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_01, 177) ,rhs_vec_4567_01, 170), _mm256_shuffle_epi32(lhs_vec_00, 255)));
+                    const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
+                    const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
 
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_02 ,_mm256_shuffle_epi32(rhs_vec_4567_02, 177), 170), _mm256_shuffle_epi32(lhs_vec_01, 0)));
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_02, 177) ,rhs_vec_4567_02, 170), _mm256_shuffle_epi32(lhs_vec_01, 85)));
+                    const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
+                    const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
 
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_03 ,_mm256_shuffle_epi32(rhs_vec_4567_03, 177), 170), _mm256_shuffle_epi32(lhs_vec_01, 170)));
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_03, 177) ,rhs_vec_4567_03, 170), _mm256_shuffle_epi32(lhs_vec_01, 255)));
+                    const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
+                    const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
 
-                    iacc_0 = _mm256_madd_epi16(iacc_0, scales_0);
+                    // Shuffle pattern two - left side input
 
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_10 ,_mm256_shuffle_epi32(rhs_vec_4567_10, 177), 170), _mm256_shuffle_epi32(lhs_vec_10, 0)));
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_10, 177) ,rhs_vec_4567_10, 170), _mm256_shuffle_epi32(lhs_vec_10, 85)));
+                    const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
+                    const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
 
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_11 ,_mm256_shuffle_epi32(rhs_vec_4567_11, 177), 170), _mm256_shuffle_epi32(lhs_vec_10, 170)));
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_11, 177) ,rhs_vec_4567_11, 170), _mm256_shuffle_epi32(lhs_vec_10, 255)));
+                    const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
+                    const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
 
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_12 ,_mm256_shuffle_epi32(rhs_vec_4567_12, 177), 170), _mm256_shuffle_epi32(lhs_vec_11, 0)));
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_12, 177) ,rhs_vec_4567_12, 170), _mm256_shuffle_epi32(lhs_vec_11, 85)));
+                    const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
+                    const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
 
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_13 ,_mm256_shuffle_epi32(rhs_vec_4567_13, 177), 170), _mm256_shuffle_epi32(lhs_vec_11, 170)));
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_13, 177) ,rhs_vec_4567_13, 170), _mm256_shuffle_epi32(lhs_vec_11, 255)));
+                    const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
+                    const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
 
-                    iacc_1 = _mm256_madd_epi16(iacc_1, scales_1);
+                    // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+                    // Resembles MMLAs into 2x2 matrices in ARM Version
+                    const __m512i zero = _mm512_setzero_epi32();
+                    __m512i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_01_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_01_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_01_0_sp1, rhs_mat_014589CD_0_sp1);
+                    __m512i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367ABEF_0_sp1);
+                    __m512i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_23_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_23_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_23_0_sp1, rhs_mat_014589CD_0_sp1);
+                    __m512i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367ABEF_0_sp1);
+                    __m512i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_01_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_01_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_01_0_sp2, rhs_mat_014589CD_0_sp2);
+                    __m512i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367ABEF_0_sp2);
+                    __m512i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_23_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_23_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_23_0_sp2, rhs_mat_014589CD_0_sp2);
+                    __m512i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367ABEF_0_sp2);
 
-                    // Accumulate the iacc value for one sb
-                    __m256i iacc_sb = _mm256_add_epi32(iacc_0, iacc_1);
+                    // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+                    __m512i iacc_mat_00 = _mm512_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
+                    __m512i iacc_mat_01 = _mm512_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
+                    __m512i iacc_mat_10 = _mm512_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
+                    __m512i iacc_mat_11 = _mm512_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
 
-                    // Broadcast the bsums of the two sub blocks  of the iteration of Q8_K across the vector
-                    // Multiply-Add with corresponding mins of Q4_Kx8 with bsums
-                    __m256i q8s_sb = _mm256_shuffle_epi32(q8s, 0);
-                    __m256i iacc_min_sb = _mm256_madd_epi16(q8s_sb, mins_01);
-                    q8s = _mm256_bsrli_epi128(q8s, 4);
 
-                    // Accumulate for the complete block
-                    iacc_b = _mm256_add_epi32(iacc_b, iacc_sb);
-                    iacc_min_b = _mm256_add_epi32(iacc_min_b, iacc_min_sb);
+                    // Straighten out to make 4 row vectors
+                    __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
+                    __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
+                    __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
+                    __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
+
+                    // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
+                    const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptrs[rp][b].d), loadMask), 68);
+                    const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16);
+
+                    // Multiply with appropiate scales and accumulate
+                    acc_rows[rp * 4]     = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)),   acc_rows[rp * 4]);
+                    acc_rows[rp * 4 + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)),  acc_rows[rp * 4 + 1]);
+                    acc_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
+                    acc_rows[rp * 4 + 3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[rp * 4 + 3]);
                 }
+            }
 
-                // Multiply-Add with scale values for the complete super block
-                acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_b), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row);
-                acc_min_rows = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_min_b), _mm256_mul_ps(col_dmin_f32, row_scale_f32), acc_min_rows);
+            // Store the accumulated values
+            for (int i = 0; i < 16; i++) {
+                _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
+            }
+        }
+    }
+
+    // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation
+    for (; y < nr / 4; y ++) {
+        const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb);
+
+        // Take group of two block_tx8 structures at each pass of the loop and perform dot product operation
+        for (int64_t x = 0; x < anc / 8; x += 2) {
+
+            const block_tx8 * b_ptr_0 = b_ptr_start + ((x)     * b_nb);
+            const block_tx8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
 
+            // Master FP accumulators
+            __m512 acc_rows[4];
+            for (int i = 0; i < 4; i++) {
+                acc_rows[i] = _mm512_setzero_ps();
             }
 
-            // Accumulated output values permuted so as to be stored in appropriate order post accumulation
-            acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask);
-            _mm256_storeu_ps(s + (y * nr + x * 8), _mm256_sub_ps(acc_row, acc_min_rows));
+            for (int64_t b = 0; b < nb; b++) {
+                // Load the sixteen blocks of quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF
+                const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs));
+                const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32));
+                const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64));
+                const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96));
+
+                const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs));
+                const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32));
+                const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64));
+                const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96));
+
+                // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess
+                const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+                const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+                const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+                const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+
+                const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240);
+                const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240);
+                const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240);
+                const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240);
+
+                const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
+                const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
+                const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
+                const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
+
+                // 4-bit -> 8-bit - Sign is maintained
+                const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7)
+                const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7)
+
+                const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15)
+                const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15)
+
+                const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23)
+                const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23)
+
+                const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31)
+                const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31)
+
+                // Shuffle pattern one - right side input
+                const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
+                const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
+
+                const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
+                const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
+
+                const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
+                const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
+
+                const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
+                const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
+
+                // Shuffle pattern two - right side input
+
+                const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
+                const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
+
+                const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
+                const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
+
+                const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
+                const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
+
+                const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
+                const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
+
+
+                // Scale values - Load the weight scale values of two block_tx8
+                __m512 col_scale_f32;
+                if constexpr (
+                        std::is_same_v<block_tx8, block_q4_0x8> ||
+                        std::is_same_v<block_tx8, block_iq4_nlx8>) {
+                    col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
+                }
+
+                // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
+                // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector
+                __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs)));
+                __m256i lhs_mat_ymm_01_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 0);
+                __m256i lhs_mat_ymm_23_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 17);
+                __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32)));
+                __m256i lhs_mat_ymm_01_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 0);
+                __m256i lhs_mat_ymm_23_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 17);
+                __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64)));
+                __m256i lhs_mat_ymm_01_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 0);
+                __m256i lhs_mat_ymm_23_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 17);
+                __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96)));
+                __m256i lhs_mat_ymm_01_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 0);
+                __m256i lhs_mat_ymm_23_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 17);
+
+                __m512i lhs_mat_01_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_0), lhs_mat_ymm_01_0, 1);
+                __m512i lhs_mat_23_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_0), lhs_mat_ymm_23_0, 1);
+                __m512i lhs_mat_01_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_1), lhs_mat_ymm_01_1, 1);
+                __m512i lhs_mat_23_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_1), lhs_mat_ymm_23_1, 1);
+                __m512i lhs_mat_01_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_2), lhs_mat_ymm_01_2, 1);
+                __m512i lhs_mat_23_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_2), lhs_mat_ymm_23_2, 1);
+                __m512i lhs_mat_01_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_3), lhs_mat_ymm_01_3, 1);
+                __m512i lhs_mat_23_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_3), lhs_mat_ymm_23_3, 1);
+
+                // Shuffle pattern one - left side input
+
+                const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
+                const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+
+                const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
+                const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+
+                const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
+                const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+
+                const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
+                const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+
+                // Shuffle pattern two - left side input
+
+                const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
+                const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+
+                const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
+                const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+
+                const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
+                const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+
+                const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
+                const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+
+                // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+                // Resembles MMLAs into 2x2 matrices in ARM Version
+                const __m512i zero = _mm512_setzero_epi32();
+                __m512i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_01_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_01_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_01_0_sp1, rhs_mat_014589CD_0_sp1);
+                __m512i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367ABEF_0_sp1);
+                __m512i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_23_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_23_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_23_0_sp1, rhs_mat_014589CD_0_sp1);
+                __m512i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367ABEF_0_sp1);
+                __m512i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_01_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_01_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_01_0_sp2, rhs_mat_014589CD_0_sp2);
+                __m512i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367ABEF_0_sp2);
+                __m512i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_23_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_23_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_23_0_sp2, rhs_mat_014589CD_0_sp2);
+                __m512i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367ABEF_0_sp2);
+
+                // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+                __m512i iacc_mat_00 = _mm512_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
+                __m512i iacc_mat_01 = _mm512_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
+                __m512i iacc_mat_10 = _mm512_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
+                __m512i iacc_mat_11 = _mm512_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+
+
+                // Straighten out to make 4 row vectors
+                __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
+                __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
+                __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
+                __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
+
+                // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
+                const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptr[b].d), loadMask), 68);
+                const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16);
+
+                // Multiply with appropiate scales and accumulate
+                acc_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)),   acc_rows[0]);
+                acc_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)),  acc_rows[1]);
+                acc_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
+                acc_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]);
+            }
+
+            // Store the accumulated values
+            for (int i = 0; i < 4; i++) {
+                _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
+            }
         }
     }
+    if (anc != nc) {
+        xstart = anc/8;
+        y = 0;
+    }
+#endif // __AVX512F__
 
-#else
-    UNUSED(kmask1);
-    UNUSED(kmask2);
-    UNUSED(kmask3);
-    ggml_gemv_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+    // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation
+
+    for (; y < anr / 4; y += 4) {
+        const block_q8_0x4 * a_ptrs[4];
+
+        a_ptrs[0] = a_ptr_start + (y * nb);
+        for (int i = 0; i < 3; ++i) {
+            a_ptrs[i + 1] = a_ptrs[i] + nb;
+        }
+
+        // Take group of eight block_tx8 structures at each pass of the loop and perform dot product operation
+        for (int64_t x = xstart; x < nc / 8; x++) {
+
+            const block_tx8 * b_ptr = b_ptr_start + (x * b_nb);
+
+            // Master FP accumulators
+            __m256 acc_rows[16];
+            for (int i = 0; i < 16; i++) {
+                acc_rows[i] = _mm256_setzero_ps();
+            }
+
+            for (int64_t b = 0; b < nb; b++) {
+                // Load the eight blocks of quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
+                const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
+                const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32));
+                const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64));
+                const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96));
+
+                // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
+                const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+                const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+                const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+                const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+
+                // 4-bit -> 8-bit - Sign is maintained
+                const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7)
+                const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7)
+
+                const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15)
+                const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15)
+
+                const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23)
+                const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23)
+
+                const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31)
+                const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31)
+
+                // Shuffle pattern one - right side input
+                const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136);  //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3)
+                const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136);  //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3)
+
+                const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136);  //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11)
+                const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136);  //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11)
+
+                const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136);  //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19)
+                const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136);  //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19)
+
+                const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136);  //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27)
+                const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136);  //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27)
+
+                // Shuffle pattern two - right side input
+
+                const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221);  //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7)
+                const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221);  //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7)
+
+                const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221);  //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15)
+                const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221);  //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15)
+
+                const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221);  //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23)
+                const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221);  //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23)
+
+                const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221);  //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31)
+                const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221);  //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31)
+
+                // Scale values - Load the wight scale values of block_tx8
+                __m256 col_scale_f32;
+                if constexpr (
+                        std::is_same_v<block_tx8, block_q4_0x8> ||
+                        std::is_same_v<block_tx8, block_iq4_nlx8>) {
+                    col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
+                }
+
+                // Process LHS in groups of four
+                for (int rp = 0; rp < 4; rp++) {
+                    // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
+                    // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
+                    __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs)));
+                    __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0);
+                    __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17);
+                    __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32)));
+                    __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0);
+                    __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17);
+                    __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64)));
+                    __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0);
+                    __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17);
+                    __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96)));
+                    __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0);
+                    __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17);
+
+                    // Shuffle pattern one - left side input
+                    const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
+                    const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+
+                    const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
+                    const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+
+                    const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
+                    const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+
+                    const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
+                    const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+
+                    // Shuffle pattern two - left side input
+                    const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
+                    const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+
+                    const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
+                    const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+
+                    const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
+                    const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+
+                    const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
+                    const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+
+                    // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+                    // Resembles MMLAs into 2x2 matrices in ARM Version
+                    const __m256i zero = _mm256_setzero_si256();
+                    __m256i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1);
+                    __m256i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1);
+                    __m256i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1);
+                    __m256i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1);
+                    __m256i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2);
+                    __m256i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2);
+                    __m256i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2);
+                    __m256i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2);
+
+                    // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+                    __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
+                    __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
+                    __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
+                    __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+
+                    // Straighten out to make 4 row vectors
+                    __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
+                    __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
+                    __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
+                    __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
+
+                    // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
+                    const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask);
+
+                    // Multiply with appropiate scales and accumulate
+                    acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
+                    acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
+                    acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
+                    acc_rows[rp * 4 + 3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32,  255)), acc_rows[rp * 4 + 3]);
+                }
+            }
+
+            // Store the accumulated values
+            for (int i = 0; i < 16; i++) {
+                _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
+            }
+        }
+    }
+
+    // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation
+    for (; y < nr / 4; y ++) {
+        const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb);
+
+        // Load the eight blocks of quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
+        for (int64_t x = xstart; x < nc / 8; x++) {
+            const block_tx8 * b_ptr = b_ptr_start + (x * b_nb);
+
+            // Master FP accumulators
+            __m256 acc_rows[4];
+            for (int i = 0; i < 4; i++) {
+                acc_rows[i] = _mm256_setzero_ps();
+            }
+
+            for (int64_t b = 0; b < nb; b++) {
+                // Load the eight block_q8_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
+                const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
+                const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32));
+                const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64));
+                const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96));
+
+                // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess
+                const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+                const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+                const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+                const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+
+                // 4-bit -> 8-bit - Sign is maintained
+                const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b));  //B0(0-7) B1(0-7) B4(0-7) B5(0-7)
+                const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b));  //B2(0-7) B3(0-7) B6(0-7) B7(0-7)
+
+                const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b));  //B0(8-15) B1(8-15) B4(8-15) B5(8-15)
+                const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b));  //B2(8-15) B3(8-15) B6(8-15) B7(8-15)
+
+                const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b));  //B0(16-23) B1(16-23) B4(16-23) B5(16-23)
+                const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b));  //B2(16-23) B3(16-23) B6(16-23) B7(16-23)
+
+                const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b));  //B0(24-31) B1(24-31) B4(24-31) B5(24-31)
+                const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b));  //B2(24-31) B3(24-31) B6(24-31) B7(24-31)
+
+                // Shuffle pattern one - right side input
+                const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136);  //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3)
+                const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136);  //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3)
+
+                const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136);  //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11)
+                const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136);  //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11)
+
+                const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136);  //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19)
+                const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136);  //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19)
+
+                const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136);  //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27)
+                const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136);  //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27)
+
+                // Shuffle pattern two - right side input
+
+                const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221);  //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7)
+                const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221);  //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7)
+
+                const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221);  //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15)
+                const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221);  //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15)
+
+                const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221);  //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23)
+                const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221);  //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23)
+
+                const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221);  //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31)
+                const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221);  //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31)
+
+                // Scale values - Load the wight scale values of block_tx8
+                __m256 col_scale_f32;
+                if constexpr (
+                        std::is_same_v<block_tx8, block_q4_0x8> ||
+                        std::is_same_v<block_tx8, block_iq4_nlx8>) {
+                    col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
+                }
+
+                // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
+                // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
+                __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs)));
+                __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0);
+                __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17);
+                __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32)));
+                __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0);
+                __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17);
+                __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64)));
+                __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0);
+                __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17);
+                __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96)));
+                __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0);
+                __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17);
+
+                // Shuffle pattern one - left side input
+
+                const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
+                const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+
+                const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
+                const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+
+                const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
+                const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+
+                const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
+                const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+
+                // Shuffle pattern two - left side input
+
+                const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
+                const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+
+                const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
+                const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+
+                const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
+                const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+
+                const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
+                const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+
+                // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+                // Resembles MMLAs into 2x2 matrices in ARM Version
+                const __m256i zero = _mm256_setzero_si256();
+                __m256i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1);
+                __m256i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1);
+                __m256i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1);
+                __m256i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1);
+                __m256i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2);
+                __m256i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2);
+                __m256i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2);
+                __m256i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2);
+
+                // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+                __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
+                __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
+                __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
+                __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+
+
+                // Straighten out to make 4 row vectors
+                __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
+                __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
+                __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
+                __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
+
+                // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
+                const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptr[b].d, loadMask);
+
+                // Multiply with appropiate scales and accumulate
+                acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
+                acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
+                acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
+                acc_rows[3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]);
+            }
+
+            // Store the accumulated values
+            for (int i = 0; i < 4; i++) {
+                _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
+            }
+        }
+    }
+}
+
+#endif // defined(__AVX2__) || defined(__AVX512F__)
+
+void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+#if defined(__AVX2__) || defined(__AVX512F__)
+    {
+        // Lookup table to convert signed nibbles to signed bytes
+        __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
+        signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
+
+        gemv_q4_b32_8x8_q8_0_lut_avx<block_q4_0x8>(n, s, bs, vx, vy, nr, nc, signextendlut);
+
+        return;
+    }
 #endif
+
+    ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
-void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK_K;
     const int nb = n / qk;
     const int ncols_interleaved = 8;
     const int blocklen = 8;
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
 
     assert (n % qk == 0);
     assert (nc % ncols_interleaved == 0);
@@ -872,21 +1415,18 @@ void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     // Lookup table to convert signed nibbles to signed bytes
     __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
     signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
-    // Shuffle masks to rearrange delta values to multiply with appropriate scales
+    // Shuffle masks to rearrange delta and scale values to multiply with appropriate scales
     __m128i deltamask = _mm_set_epi8(15, 14, 7, 6, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0);
+    __m128i scalemask = _mm_set_epi8(7, 7, 3, 3, 6, 6, 2, 2, 5, 5, 1, 1, 4, 4, 0, 0);
     // Permute mask used for easier vector processing at later stages
     __m256i finalpermutemask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
 
-    const __m256i m3b = _mm256_set1_epi8(3);
-    const __m128i m4b_sse = _mm_set1_epi8(0xF);
-
-    //Mask to get appropriate scales
-    __m128i scalemask1 = _mm_set_epi8(14,14,6,6,12,12,4,4,10,10,2,2,8,8,0,0);
-    __m128i scalemask2 = _mm_set_epi8(15,15,7,7,13,13,5,5,11,11,3,3,9,9,1,1);
+    // Mask to extract nibbles from bytes
+    const __m256i m4b = _mm256_set1_epi8(0x0F);
 
     int64_t b_nb = n / QK_K;
 
-    const block_q2_Kx8 * b_ptr_start = (const block_q2_Kx8 *)vx;
+    const block_q4_Kx8 * b_ptr_start = (const block_q4_Kx8 *)vx;
     const block_q8_K * a_ptr_start = (const block_q8_K *)vy;
 
     // Process Q8_K blocks one by one
@@ -895,11 +1435,11 @@ void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
         // Pointers to LHS blocks of block_q8_K format
         const block_q8_K * a_ptr = a_ptr_start + (y * nb);
 
-        // Take group of eight interleaved block_q2_K structures at each pass of the loop and perform dot product operation
-        for(int64_t x = 0; x < nc / 8; x++) {
+        // Take group of eight interleaved block_q4_K structures at each pass of the loop and perform dot product operation
+        for (int64_t x = 0; x < nc / 8; x++) {
 
             // Pointers to RHS blocks
-            const block_q2_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
+            const block_q4_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
 
             // Master FP accumulators
             __m256 acc_row = _mm256_setzero_ps();
@@ -907,10 +1447,10 @@ void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
             for (int64_t b = 0; b < nb; b++) {
 
-                // Load and convert to FP32 delta from block_q8_K
+                // Load and convert to FP32 scale from block_q8_K
                 const __m256 row_scale_f32 = _mm256_set1_ps((a_ptr[b].d));
 
-                // Load the delta values for the 8 blocks interleaved in block_q2_Kx8
+                // Load the scale values for the 8 blocks interleaved in block_q4_Kx8
                 // col_scale_f32 rearranged so as to multiply with appropriate quants
                 const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, deltamask);
                 const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin);
@@ -918,11 +1458,15 @@ void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                 __m256i iacc_b = _mm256_setzero_si256();
                 __m256i iacc_min_b = _mm256_setzero_si256();
 
-                // Processes eight sub blocks from each Q2_K in each iteration
-                for(int sb = 0; sb < QK_K / 128; sb++) {
+                const __m256i q8sums = _mm256_loadu_si256((const __m256i * )(a_ptr[b].bsums));
+                __m256i q8s = _mm256_castsi128_si256(_mm_hadd_epi16(_mm256_castsi256_si128(q8sums), _mm256_extracti128_si256(q8sums, 1)));
+                q8s = _mm256_permute2f128_si256(q8s, q8s, 0);
 
-                    // Load the eight block_q2_K for eight sub blocks quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
-                    const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + sb * 256));
+                // Processes two sub blocks from each Q4_K in each iteration
+                for (int sb = 0; sb < QK_K / 64; sb++) {
+
+                    // Load the eight block_q4_K for two sub blocks quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
+                    const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + sb * 256));
                     const __m256i rhs_raw_vec_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 32 + sb * 256));
                     const __m256i rhs_raw_vec_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 64 + sb * 256));
                     const __m256i rhs_raw_vec_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 96 + sb * 256));
@@ -931,982 +1475,482 @@ void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                     const __m256i rhs_raw_vec_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 192 + sb * 256));
                     const __m256i rhs_raw_vec_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 224 + sb * 256));
 
-                    // 2-bit -> 8-bit
-                    // Values of the 0th,2nd,4th,6th sub blocks of eight block_q2_K structures for the sb loop
-                    const __m256i rhs_vec_0123_00 = _mm256_and_si256(rhs_raw_vec_0123_0, m3b); //B00(0-7) B01(0-7) B02(0-7) B03(0-7)
-                    const __m256i rhs_vec_0123_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 2), m3b); //B20(0-7) B21(0-7) B22(0-7) B23(0-7)
-                    const __m256i rhs_vec_0123_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 4), m3b); //B40(0-7) B41(0-7) B42(0-7) B43(0-7)
-                    const __m256i rhs_vec_0123_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 6), m3b); //B60(0-7) B61(0-7) B62(0-7) B63(0-7)
-
-                    const __m256i rhs_vec_4567_00 = _mm256_and_si256(rhs_raw_vec_4567_0, m3b); //B04(0-7) B05(0-7) B06(0-7) B07(0-7)
-                    const __m256i rhs_vec_4567_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 2), m3b); //B24(0-7) B25(0-7) B26(0-7) B27(0-7)
-                    const __m256i rhs_vec_4567_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 4), m3b); //B44(0-7) B45(0-7) B46(0-7) B47(0-7)
-                    const __m256i rhs_vec_4567_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 6), m3b); //B64(0-7) B65(0-7) B66(0-7) B67(0-7)
-
-                    const __m256i rhs_vec_0123_01 = _mm256_and_si256(rhs_raw_vec_0123_1, m3b); //B00(8-15) B01(8-15) B02(8-15) B03(8-15)
-                    const __m256i rhs_vec_0123_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 2), m3b); //B20(8-15) B21(8-15) B22(8-15) B23(8-15)
-                    const __m256i rhs_vec_0123_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m3b); //B40(8-15) B41(8-15) B42(8-15) B43(8-15)
-                    const __m256i rhs_vec_0123_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 6), m3b); //B60(8-15) B61(8-15) B62(8-15) B63(8-15)
-
-                    const __m256i rhs_vec_4567_01 = _mm256_and_si256(rhs_raw_vec_4567_1, m3b); //B04(8-15) B05(8-15) B06(8-15) B07(8-15)
-                    const __m256i rhs_vec_4567_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 2), m3b); //B24(8-15) B25(8-15) B26(8-15) B27(8-15)
-                    const __m256i rhs_vec_4567_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m3b); //B44(8-15) B45(8-15) B46(8-15) B47(8-15)
-                    const __m256i rhs_vec_4567_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 6), m3b); //B64(8-15) B65(8-15) B66(8-15) B67(8-15)
-
-                    // Values of the 1st,3rd,5th,7th sub blocks of eight block_q2_K structures for the sb loop
-                    const __m256i rhs_vec_0123_10 = _mm256_and_si256(rhs_raw_vec_0123_2, m3b); //B10(0-7) B11(0-7) B12(0-7) B13(0-7)
-                    const __m256i rhs_vec_0123_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 2), m3b); //B30(0-7) B31(0-7) B32(0-7) B33(0-7)
-                    const __m256i rhs_vec_0123_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 4), m3b); //B50(0-7) B51(0-7) B52(0-7) B53(0-7)
-                    const __m256i rhs_vec_0123_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 6), m3b); //B70(0-7) B71(0-7) B72(0-7) B73(0-7)
-
-                    const __m256i rhs_vec_4567_10 = _mm256_and_si256(rhs_raw_vec_4567_2, m3b); //B14(0-7) B15(0-7) B16(0-7) B17(0-7)
-                    const __m256i rhs_vec_4567_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 2), m3b); //B34(0-7) B35(0-7) B36(0-7) B37(0-7)
-                    const __m256i rhs_vec_4567_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 4), m3b); //B54(0-7) B55(0-7) B56(0-7) B57(0-7)
-                    const __m256i rhs_vec_4567_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 6), m3b); //B74(0-7) B75(0-7) B76(0-7) B77(0-7)
-
-                    const __m256i rhs_vec_0123_11 = _mm256_and_si256(rhs_raw_vec_0123_3, m3b); //B10(8-15) B11(8-15) B12(8-15) B13(8-15)
-                    const __m256i rhs_vec_0123_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 2), m3b); //B30(8-15) B31(8-15) B32(8-15) B33(8-15)
-                    const __m256i rhs_vec_0123_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 4), m3b); //B50(8-15) B51(8-15) B52(8-15) B53(8-15)
-                    const __m256i rhs_vec_0123_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 6), m3b); //B70(8-15) B71(8-15) B72(8-15) B73(8-15)
-
-                    const __m256i rhs_vec_4567_11 = _mm256_and_si256(rhs_raw_vec_4567_3, m3b); //B14(8-15) B15(8-15) B16(8-15) B17(8-15)
-                    const __m256i rhs_vec_4567_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 2), m3b); //B34(8-15) B35(8-15) B36(8-15) B37(8-15)
-                    const __m256i rhs_vec_4567_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 4), m3b); //B54(8-15) B55(8-15) B56(8-15) B57(8-15)
-                    const __m256i rhs_vec_4567_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 6), m3b); //B74(8-15) B75(8-15) B76(8-15) B77(8-15)
-
-                    //Scales and Mins of corresponding sub blocks from different Q2_K structures are stored together
-                    //s00 m00  s01 m01   s10 m10  s11 m11  s20 m20  s21 m21   s30 m30  s31 m31  s40 m40  s41 m41   s50 m50  s51 m51  s60 m60  s61 m61   s70 m70  s71 m71
-
-                    const __m128i mins_and_scales_01 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + sb * 64));
-                    const __m128i mins_and_scales_23 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 16 + sb * 64));
-                    const __m128i mins_and_scales_45 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 32 + sb * 64));
-                    const __m128i mins_and_scales_67 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 48 + sb * 64));
-
-                    // Extract scales which is lower half from mins_and_scales
-                    const __m128i scales_01 = _mm_and_si128(mins_and_scales_01, m4b_sse);
-                    const __m128i scales_23 = _mm_and_si128(mins_and_scales_23, m4b_sse);
-                    const __m128i scales_45 = _mm_and_si128(mins_and_scales_45, m4b_sse);
-                    const __m128i scales_67 = _mm_and_si128(mins_and_scales_67, m4b_sse);
-
-                    // Extract mins which is upper half from mins_and_scales
-                    const __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_01, 4), m4b_sse));
-                    const __m256i mins_23 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_23, 4), m4b_sse));
-                    const __m256i mins_45 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_45, 4), m4b_sse));
-                    const __m256i mins_67 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_67, 4), m4b_sse));
-
-                    // Scales of sub blocks in the sb loop
-                    // Scales of the 0th sub block from each super block
-                    __m128i scales_rearrange_0 = _mm_shuffle_epi8(scales_01, scalemask1);
-                    __m256i scales_0 = _mm256_cvtepu8_epi16(scales_rearrange_0);
-
-                    // Scales of the 1st sub block from each super block
-                    __m128i scales_rearrange_1 = _mm_shuffle_epi8(scales_01, scalemask2);
-                    __m256i scales_1 = _mm256_cvtepu8_epi16(scales_rearrange_1);
-
-                    // Scales of the 2nd sub block from each super block
-                    __m128i scales_rearrange_2 = _mm_shuffle_epi8(scales_23, scalemask1);
-                    __m256i scales_2 = _mm256_cvtepu8_epi16(scales_rearrange_2);
-
-                    // Scales of the 3rd sub block from each super block
-                    __m128i scales_rearrange_3 = _mm_shuffle_epi8(scales_23, scalemask2);
-                    __m256i scales_3 = _mm256_cvtepu8_epi16(scales_rearrange_3);
-
-                    // Scales of the 4th sub block from each super block
-                    __m128i scales_rearrange_4 = _mm_shuffle_epi8(scales_45, scalemask1);
-                    __m256i scales_4 = _mm256_cvtepu8_epi16(scales_rearrange_4);
-
-                    // Scales of the 5th sub block from each super block
-                    __m128i scales_rearrange_5 = _mm_shuffle_epi8(scales_45, scalemask2);
-                    __m256i scales_5 = _mm256_cvtepu8_epi16(scales_rearrange_5);
-
-                    // Scales of the 6th sub block from each super block
-                    __m128i scales_rearrange_6 = _mm_shuffle_epi8(scales_67, scalemask1);
-                    __m256i scales_6 = _mm256_cvtepu8_epi16(scales_rearrange_6);
-
-                    // Scales of the 7th sub block from each super block
-                    __m128i scales_rearrange_7 = _mm_shuffle_epi8(scales_67, scalemask2);
-                    __m256i scales_7 = _mm256_cvtepu8_epi16(scales_rearrange_7);
-
-                    // Load the sub block values corresponding to sb in block_q8_K in batches of 16 bytes and replicate the same across 256 bit vector
-                    __m256i lhs_vec_0 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + sb * 128)));
-                    __m256i lhs_vec_1 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 16 + sb * 128)));
-                    __m256i lhs_vec_2 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 32 + sb * 128)));
-                    __m256i lhs_vec_3 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 48 + sb * 128)));
-                    __m256i lhs_vec_4 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 64 + sb * 128)));
-                    __m256i lhs_vec_5 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 80 + sb * 128)));
-                    __m256i lhs_vec_6 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 96 + sb * 128)));
-                    __m256i lhs_vec_7 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 112 + sb * 128)));
-
-                    lhs_vec_0 = _mm256_permute2f128_si256(lhs_vec_0, lhs_vec_0, 0);
-                    lhs_vec_1 = _mm256_permute2f128_si256(lhs_vec_1, lhs_vec_1, 0);
-                    lhs_vec_2 = _mm256_permute2f128_si256(lhs_vec_2, lhs_vec_2, 0);
-                    lhs_vec_3 = _mm256_permute2f128_si256(lhs_vec_3, lhs_vec_3, 0);
-                    lhs_vec_4 = _mm256_permute2f128_si256(lhs_vec_4, lhs_vec_4, 0);
-                    lhs_vec_5 = _mm256_permute2f128_si256(lhs_vec_5, lhs_vec_5, 0);
-                    lhs_vec_6 = _mm256_permute2f128_si256(lhs_vec_6, lhs_vec_6, 0);
-                    lhs_vec_7 = _mm256_permute2f128_si256(lhs_vec_7, lhs_vec_7, 0);
-
-                    __m256i iacc_0 = _mm256_setzero_si256();
-                    __m256i iacc_1 = _mm256_setzero_si256();
-                    __m256i iacc_2 = _mm256_setzero_si256();
-                    __m256i iacc_3 = _mm256_setzero_si256();
-                    __m256i iacc_4 = _mm256_setzero_si256();
-                    __m256i iacc_5 = _mm256_setzero_si256();
-                    __m256i iacc_6 = _mm256_setzero_si256();
-                    __m256i iacc_7 = _mm256_setzero_si256();
-
-                    // Dot product done within 32 bit lanes and accumulated in the same vector
-                    // First done for 0th sub block and then for seven (1st - 7th) other sub blocks processed for each sb (sb < QK_K/128 loop)                    // B0(0-3) B4(0-3) B1(0-3) B5(0-3) B2(0-3) B6(0-3) B3(0-3) B7(0-3) with A0(0-3)
-                    // B0(4-7) B4(4-7) B1(4-7) B5(4-7) B2(4-7) B6(4-7) B3(4-7) B7(4-7) with A0(4-7)
-                    // B0(8-11) B4(8-11) B1(8-11) B5(8-11) B2(8-11) B6(8-11) B3(8-11) B7(8-11) with A0(8-11)
-                    // B0(12-15) B4(12-15) B1(12-15) B5(12-15) B2(12-15) B6(12-15) B3(12-15) B7(12-15) with A0(12-15)
-
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_00 ,_mm256_shuffle_epi32(rhs_vec_4567_00, 177), 170), _mm256_shuffle_epi32(lhs_vec_0, 0)));
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_00, 177) ,rhs_vec_4567_00, 170), _mm256_shuffle_epi32(lhs_vec_0, 85)));
-
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_01 ,_mm256_shuffle_epi32(rhs_vec_4567_01, 177), 170), _mm256_shuffle_epi32(lhs_vec_0, 170)));
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_01, 177) ,rhs_vec_4567_01, 170), _mm256_shuffle_epi32(lhs_vec_0, 255)));
-
-                    iacc_0 = _mm256_madd_epi16(iacc_0, scales_0);
-
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_10 ,_mm256_shuffle_epi32(rhs_vec_4567_10, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 0)));
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_10, 177) ,rhs_vec_4567_10, 170), _mm256_shuffle_epi32(lhs_vec_1, 85)));
-
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_11 ,_mm256_shuffle_epi32(rhs_vec_4567_11, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 170)));
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_11, 177) ,rhs_vec_4567_11, 170), _mm256_shuffle_epi32(lhs_vec_1, 255)));
-
-                    iacc_1 = _mm256_madd_epi16(iacc_1, scales_1);
-
-                    iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_20 ,_mm256_shuffle_epi32(rhs_vec_4567_20, 177), 170), _mm256_shuffle_epi32(lhs_vec_2, 0)));
-                    iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_20, 177) ,rhs_vec_4567_20, 170), _mm256_shuffle_epi32(lhs_vec_2, 85)));
-
-                    iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_21 ,_mm256_shuffle_epi32(rhs_vec_4567_21, 177), 170), _mm256_shuffle_epi32(lhs_vec_2, 170)));
-                    iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_21, 177) ,rhs_vec_4567_21, 170), _mm256_shuffle_epi32(lhs_vec_2, 255)));
-
-                    iacc_2 = _mm256_madd_epi16(iacc_2, scales_2);
-
-                    iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_30 ,_mm256_shuffle_epi32(rhs_vec_4567_30, 177), 170), _mm256_shuffle_epi32(lhs_vec_3, 0)));
-                    iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_30, 177) ,rhs_vec_4567_30, 170), _mm256_shuffle_epi32(lhs_vec_3, 85)));
-
-                    iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_31 ,_mm256_shuffle_epi32(rhs_vec_4567_31, 177), 170), _mm256_shuffle_epi32(lhs_vec_3, 170)));
-                    iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_31, 177) ,rhs_vec_4567_31, 170), _mm256_shuffle_epi32(lhs_vec_3, 255)));
-
-                    iacc_3 = _mm256_madd_epi16(iacc_3, scales_3);
-
-                    iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_40 ,_mm256_shuffle_epi32(rhs_vec_4567_40, 177), 170), _mm256_shuffle_epi32(lhs_vec_4, 0)));
-                    iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_40, 177) ,rhs_vec_4567_40, 170), _mm256_shuffle_epi32(lhs_vec_4, 85)));
-
-                    iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_41 ,_mm256_shuffle_epi32(rhs_vec_4567_41, 177), 170), _mm256_shuffle_epi32(lhs_vec_4, 170)));
-                    iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_41, 177) ,rhs_vec_4567_41, 170), _mm256_shuffle_epi32(lhs_vec_4, 255)));
-
-                    iacc_4 = _mm256_madd_epi16(iacc_4, scales_4);
-
-                    iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_50 ,_mm256_shuffle_epi32(rhs_vec_4567_50, 177), 170), _mm256_shuffle_epi32(lhs_vec_5, 0)));
-                    iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_50, 177) ,rhs_vec_4567_50, 170), _mm256_shuffle_epi32(lhs_vec_5, 85)));
-
-                    iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_51 ,_mm256_shuffle_epi32(rhs_vec_4567_51, 177), 170), _mm256_shuffle_epi32(lhs_vec_5, 170)));
-                    iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_51, 177) ,rhs_vec_4567_51, 170), _mm256_shuffle_epi32(lhs_vec_5, 255)));
-
-                    iacc_5 = _mm256_madd_epi16(iacc_5, scales_5);
-
-                    iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_60 ,_mm256_shuffle_epi32(rhs_vec_4567_60, 177), 170), _mm256_shuffle_epi32(lhs_vec_6, 0)));
-                    iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_60, 177) ,rhs_vec_4567_60, 170), _mm256_shuffle_epi32(lhs_vec_6, 85)));
-
-                    iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_61 ,_mm256_shuffle_epi32(rhs_vec_4567_61, 177), 170), _mm256_shuffle_epi32(lhs_vec_6, 170)));
-                    iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_61, 177) ,rhs_vec_4567_61, 170), _mm256_shuffle_epi32(lhs_vec_6, 255)));
-
-                    iacc_6 = _mm256_madd_epi16(iacc_6, scales_6);
-
-                    iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_70 ,_mm256_shuffle_epi32(rhs_vec_4567_70, 177), 170), _mm256_shuffle_epi32(lhs_vec_7, 0)));
-                    iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_70, 177) ,rhs_vec_4567_70, 170), _mm256_shuffle_epi32(lhs_vec_7, 85)));
-
-                    iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_71 ,_mm256_shuffle_epi32(rhs_vec_4567_71, 177), 170), _mm256_shuffle_epi32(lhs_vec_7, 170)));
-                    iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_71, 177) ,rhs_vec_4567_71, 170), _mm256_shuffle_epi32(lhs_vec_7, 255)));
-
-                    iacc_7 = _mm256_madd_epi16(iacc_7, scales_7);
-
-                    // Accumulate the iacc value for one sb
-                    __m256i iacc_sb = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_0, iacc_1), _mm256_add_epi32(iacc_2, iacc_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_4, iacc_5), _mm256_add_epi32(iacc_6, iacc_7)));
-
-                    __m128i q8sums = _mm_loadu_si128((const __m128i *)(a_ptr[b].bsums + sb * 8));
-                    __m256i q8s = _mm256_castsi128_si256(q8sums);
-                    q8s= _mm256_permute2f128_si256(q8s, q8s, 0);
-
-                    // Broadcast the bsums of the two corresponding subblocks of q8_k
-                    // Multiply-Add with corresponding mins of Q2_Kx8 with bsums
-                    __m256i iacc_min_sb_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 0), mins_01);
-                    __m256i iacc_min_sb_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 85), mins_23);
-                    __m256i iacc_min_sb_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 170), mins_45);
-                    __m256i iacc_min_sb_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 255), mins_67);
-
-                    __m256i iacc_min_sb = _mm256_add_epi32(_mm256_add_epi32(iacc_min_sb_01, iacc_min_sb_23), _mm256_add_epi32(iacc_min_sb_45,iacc_min_sb_67));
-
-                    // Accumulate for the complete block
-                    iacc_b = _mm256_add_epi32(iacc_b, iacc_sb);
-                    iacc_min_b = _mm256_add_epi32(iacc_min_b, iacc_min_sb);
-                }
-
-                //Multiply-Add with scale values for complete super block
-                acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_b), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row);
-                acc_min_rows = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_min_b), _mm256_mul_ps(col_dmin_f32, row_scale_f32), acc_min_rows);
-            }
-            // Accumulated output values permuted so as to be stored in appropriate order post accumulation
-            acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask);
-            _mm256_storeu_ps(s + (y * nr + x * 8), _mm256_sub_ps(acc_row, acc_min_rows));
-        }
-    }
-#else
-
-    ggml_gemv_q2_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
-
-#endif
-}
-
-void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
-    const int qk = QK8_0;
-    const int nb = n / qk;
-    const int ncols_interleaved = 8;
-    const int blocklen = 8;
-
-    assert (n % qk == 0);
-    assert (nr % 4 == 0);
-    assert (nc % ncols_interleaved == 0);
-
-    UNUSED(s);
-    UNUSED(bs);
-    UNUSED(vx);
-    UNUSED(vy);
-    UNUSED(nr);
-    UNUSED(nc);
-    UNUSED(nb);
-    UNUSED(ncols_interleaved);
-    UNUSED(blocklen);
-
-#if defined(__AVX2__) || defined(__AVX512F__)
-    {
-        const block_q4_0x8 * b_ptr_start = (const block_q4_0x8 *)vx;
-        const block_q8_0x4 * a_ptr_start = (const block_q8_0x4 *)vy;
-        int64_t b_nb = n / QK4_0;
-        int64_t y = 0;
-        // Mask to mask out nibbles from packed bytes
-        const __m256i m4b = _mm256_set1_epi8(0x0F);
-        const __m128i loadMask = _mm_blend_epi32(_mm_setzero_si128(), _mm_set1_epi32(0xFFFFFFFF), 3);
-        // Lookup table to convert signed nibbles to signed bytes
-        __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
-        signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
-        // Permute mask used for easier vector processing at later stages
-        __m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4);
-        int64_t xstart = 0;
-        int anr = nr - nr%16; // Used to align nr with boundary of 16
-    #ifdef __AVX512F__
-        int anc = nc - nc%16; // Used to align nc with boundary of 16
-        // Mask to mask out nibbles from packed bytes expanded to 512 bit length
-        const __m512i m4bexpanded = _mm512_set1_epi8(0x0F);
-        // Lookup table to convert signed nibbles to signed bytes expanded to 512 bit length
-        __m512i signextendlutexpanded = _mm512_inserti32x8(_mm512_castsi256_si512(signextendlut), signextendlut, 1);
-
-        // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation
-        for (; y < anr / 4; y += 4) {
-
-            const block_q8_0x4 * a_ptrs[4];
-
-            a_ptrs[0] = a_ptr_start + (y * nb);
-            for (int i = 0; i < 3; ++i) {
-                a_ptrs[i + 1] = a_ptrs[i] + nb;
-            }
-
-            // Take group of two block_q4_0x8 structures at each pass of the loop and perform dot product operation
-            for (int64_t x = 0; x < anc / 8; x += 2) {
-
-                const block_q4_0x8 * b_ptr_0 = b_ptr_start + ((x)     * b_nb);
-                const block_q4_0x8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
-
-                // Master FP accumulators
-                __m512 acc_rows[16];
-                for (int i = 0; i < 16; i++) {
-                    acc_rows[i] = _mm512_setzero_ps();
-                }
-
-                for (int64_t b = 0; b < nb; b++) {
-                    // Load the sixteen block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF
-                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs));
-                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32));
-                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64));
-                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96));
-
-                    const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs));
-                    const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32));
-                    const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64));
-                    const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96));
-
-                    // Save the values in the following vectors in the formats B0B1B4B5B8B9BCBD, B2B3B6B7BABBBEBF for further processing and storing of values
-                    const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
-                    const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
-
-                    const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240);
-                    const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240);
-
-                    const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
-                    const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
-                    const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
-                    const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
-
-                    // 4-bit -> 8-bit - Sign is maintained
-                    const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7)
-                    const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7)
-
-                    const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15)
-                    const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15)
-
-                    const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23)
-                    const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23)
-
-                    const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31)
-                    const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31)
-
-                    // Shuffle pattern one - right side input
-                    const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
-                    const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
-
-                    const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
-                    const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
-
-                    const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
-                    const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
-
-                    const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
-                    const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
-
-                    // Shuffle pattern two - right side input
-
-                    const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
-                    const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
-
-                    const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
-                    const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
-
-                    const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
-                    const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
-
-                    const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
-                    const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
-
-                    // Scale values - Load the weight scale values of two block_q4_0x8
-                    const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
-
-                    // Process LHS in pairs of rows
-                    for (int rp = 0; rp < 4; rp++) {
-
-                        // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
-                        // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector
-                        __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs)));
-                        __m256i lhs_mat_ymm_01_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 0);
-                        __m256i lhs_mat_ymm_23_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 17);
-                        __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32)));
-                        __m256i lhs_mat_ymm_01_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 0);
-                        __m256i lhs_mat_ymm_23_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 17);
-                        __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64)));
-                        __m256i lhs_mat_ymm_01_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 0);
-                        __m256i lhs_mat_ymm_23_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 17);
-                        __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96)));
-                        __m256i lhs_mat_ymm_01_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 0);
-                        __m256i lhs_mat_ymm_23_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 17);
-
-                        __m512i lhs_mat_01_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_0), lhs_mat_ymm_01_0, 1);
-                        __m512i lhs_mat_23_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_0), lhs_mat_ymm_23_0, 1);
-                        __m512i lhs_mat_01_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_1), lhs_mat_ymm_01_1, 1);
-                        __m512i lhs_mat_23_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_1), lhs_mat_ymm_23_1, 1);
-                        __m512i lhs_mat_01_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_2), lhs_mat_ymm_01_2, 1);
-                        __m512i lhs_mat_23_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_2), lhs_mat_ymm_23_2, 1);
-                        __m512i lhs_mat_01_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_3), lhs_mat_ymm_01_3, 1);
-                        __m512i lhs_mat_23_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_3), lhs_mat_ymm_23_3, 1);
-
-                        // Shuffle pattern one - left side input
-
-                        const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
-                        const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
-
-                        const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
-                        const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
-
-                        const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
-                        const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
-
-                        const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
-                        const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
-
-                        // Shuffle pattern two - left side input
-
-                        const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
-                        const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
-
-                        const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
-                        const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
-
-                        const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
-                        const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
-
-                        const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
-                        const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
-
-                        // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
-                        // Resembles MMLAs into 2x2 matrices in ARM Version
-                        const __m512i zero = _mm512_setzero_epi32();
-                        __m512i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_01_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_01_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_01_0_sp1, rhs_mat_014589CD_0_sp1);
-                        __m512i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367ABEF_0_sp1);
-                        __m512i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_23_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_23_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_23_0_sp1, rhs_mat_014589CD_0_sp1);
-                        __m512i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367ABEF_0_sp1);
-                        __m512i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_01_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_01_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_01_0_sp2, rhs_mat_014589CD_0_sp2);
-                        __m512i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367ABEF_0_sp2);
-                        __m512i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_23_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_23_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_23_0_sp2, rhs_mat_014589CD_0_sp2);
-                        __m512i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367ABEF_0_sp2);
-
-                        // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
-                        __m512i iacc_mat_00 = _mm512_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
-                        __m512i iacc_mat_01 = _mm512_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
-                        __m512i iacc_mat_10 = _mm512_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
-                        __m512i iacc_mat_11 = _mm512_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
-
-
-                        // Straighten out to make 4 row vectors
-                        __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
-                        __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
-                        __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
-                        __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
-
-                        // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
-                        const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptrs[rp][b].d), loadMask), 68);
-                        const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16);
-
-                        // Multiply with appropiate scales and accumulate
-                        acc_rows[rp * 4]     = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)),   acc_rows[rp * 4]);
-                        acc_rows[rp * 4 + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)),  acc_rows[rp * 4 + 1]);
-                        acc_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
-                        acc_rows[rp * 4 + 3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[rp * 4 + 3]);
-                    }
-                }
-
-                // Store the accumulated values
-                for (int i = 0; i < 16; i++) {
-                    _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
-                }
-            }
-        }
-        // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation
-        for (; y < nr / 4; y ++) {
-
-            const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb);
-
-            // Take group of two block_q4_0x8 structures at each pass of the loop and perform dot product operation
-            for (int64_t x = 0; x < anc / 8; x += 2) {
-
-                const block_q4_0x8 * b_ptr_0 = b_ptr_start + ((x)     * b_nb);
-                const block_q4_0x8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
-
-                // Master FP accumulators
-                __m512 acc_rows[4];
-                for (int i = 0; i < 4; i++) {
-                    acc_rows[i] = _mm512_setzero_ps();
-                }
-
-                for (int64_t b = 0; b < nb; b++) {
-                    // Load the sixteen block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF
-                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs));
-                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32));
-                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64));
-                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96));
-
-                    const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs));
-                    const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32));
-                    const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64));
-                    const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96));
-
-                    // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess
-                    const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
-                    const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
-
-                    const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240);
-                    const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240);
+                    // 4-bit -> 8-bit
+                    // Values of the first sub block of eight block_q4_K structures for the sb loop
+                    const __m256i rhs_vec_0123_00 = _mm256_and_si256(rhs_raw_vec_0123_0, m4b);
+                    const __m256i rhs_vec_4567_00 = _mm256_and_si256(rhs_raw_vec_4567_0, m4b);
+                    const __m256i rhs_vec_0123_01 = _mm256_and_si256(rhs_raw_vec_0123_1, m4b);
+                    const __m256i rhs_vec_4567_01 = _mm256_and_si256(rhs_raw_vec_4567_1, m4b);
+                    const __m256i rhs_vec_0123_02 = _mm256_and_si256(rhs_raw_vec_0123_2, m4b);
+                    const __m256i rhs_vec_4567_02 = _mm256_and_si256(rhs_raw_vec_4567_2, m4b);
+                    const __m256i rhs_vec_0123_03 = _mm256_and_si256(rhs_raw_vec_0123_3, m4b);
+                    const __m256i rhs_vec_4567_03 = _mm256_and_si256(rhs_raw_vec_4567_3, m4b);
 
-                    const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
-                    const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
-                    const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
-                    const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
+                    // Values of the second sub block of eight block_q4_K structures when sb = 1
+                    const __m256i rhs_vec_0123_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 4), m4b);
+                    const __m256i rhs_vec_4567_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 4), m4b);
+                    const __m256i rhs_vec_0123_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m4b);
+                    const __m256i rhs_vec_4567_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m4b);
+                    const __m256i rhs_vec_0123_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 4), m4b);
+                    const __m256i rhs_vec_4567_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 4), m4b);
+                    const __m256i rhs_vec_0123_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 4), m4b);
+                    const __m256i rhs_vec_4567_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 4), m4b);
 
-                    // 4-bit -> 8-bit - Sign is maintained
-                    const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7)
-                    const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7)
+                    uint32_t utmp_0[4], utmp_1[4];
 
-                    const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15)
-                    const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15)
+                    // Scales and Mins of corresponding sub blocks from different Q8_K structures are stored together
+                    // The below block is for eg to extract first sub block's scales and mins from different Q4_K structures for the sb loop
+                    memcpy(utmp_0, b_ptr[b].scales + 24 * sb, 12);
+                    utmp_0[3] = ((utmp_0[2] >> 4) & kmask2) | (((utmp_0[1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_0 = utmp_0[1] & kmask1;
+                    utmp_0[1] = (utmp_0[2] & kmask2) | (((utmp_0[0] >> 6) & kmask3) << 4);
+                    utmp_0[2] = uaux_0;
+                    utmp_0[0] &= kmask1;
 
-                    const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23)
-                    const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23)
+                    // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop
+                    memcpy(utmp_1, b_ptr[b].scales + 12 + sb * 24, 12);
+                    utmp_1[3] = ((utmp_1[2] >> 4) & kmask2) | (((utmp_1[1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_1 = utmp_1[1] & kmask1;
+                    utmp_1[1] = (utmp_1[2] & kmask2) | (((utmp_1[0] >> 6) & kmask3) << 4);
+                    utmp_1[2] = uaux_1;
+                    utmp_1[0] &= kmask1;
 
-                    const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31)
-                    const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31)
+                    // Scales of first sub block in the sb loop
+                    const __m128i mins_and_scales_0 = _mm_set_epi32(utmp_0[3], utmp_0[2], utmp_0[1], utmp_0[0]);
+                    __m128i scales_rearrange_0 = _mm_shuffle_epi8(mins_and_scales_0, scalemask);
+                    __m256i scales_0 = _mm256_cvtepu8_epi16(scales_rearrange_0);
 
-                    // Shuffle pattern one - right side input
-                    const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
-                    const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
+                    // Scales of second sub block in the sb loop
+                    __m128i mins_and_scales_1 = _mm_set_epi32(utmp_1[3], utmp_1[2], utmp_1[1], utmp_1[0]);
+                    __m128i scales_rearrange_1 = _mm_shuffle_epi8(mins_and_scales_1, scalemask);
+                    __m256i scales_1 = _mm256_cvtepu8_epi16(scales_rearrange_1);
 
-                    const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
-                    const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
+                    // Mins of first and second sub block of Q4_K block are arranged side by side
+                    __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_unpacklo_epi8(_mm_shuffle_epi32(mins_and_scales_0, 78), _mm_shuffle_epi32(mins_and_scales_1, 78)));
 
-                    const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
-                    const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
+                    // Load the two sub block values corresponding to sb in block_q8_K in batches of 16 bytes and replicate the same across 256 bit vector
+                    __m256i lhs_vec_00 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + sb * 64)));
+                    __m256i lhs_vec_01 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 16 + sb * 64)));
+                    __m256i lhs_vec_10 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 32 + sb * 64)));
+                    __m256i lhs_vec_11 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 48 + sb * 64)));
 
-                    const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
-                    const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
+                    lhs_vec_00 = _mm256_permute2f128_si256(lhs_vec_00, lhs_vec_00, 0);
+                    lhs_vec_01 = _mm256_permute2f128_si256(lhs_vec_01, lhs_vec_01, 0);
+                    lhs_vec_10 = _mm256_permute2f128_si256(lhs_vec_10, lhs_vec_10, 0);
+                    lhs_vec_11 = _mm256_permute2f128_si256(lhs_vec_11, lhs_vec_11, 0);
 
-                    // Shuffle pattern two - right side input
+                    // Dot product done within 32 bit lanes and accumulated in the same vector
+                    // First done for first sub block and thenn for second sub block in each sb
+                    // B0(0-3) B4(0-3) B1(0-3) B5(0-3) B2(0-3) B6(0-3) B3(0-3) B7(0-3) with A0(0-3)
+                    // B0(4-7) B4(4-7) B1(4-7) B5(4-7) B2(4-7) B6(4-7) B3(4-7) B7(4-7) with A0(4-7)
+                    // ...........................................................................
+                    // B0(28-31) B4(28-31) B1(28-31) B5(28-31) B2(28-31) B6(28-31) B3(28-31) B7(28-31) with A0(28-31)
 
-                    const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
-                    const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
 
-                    const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
-                    const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
+                    __m256i iacc_0 = _mm256_setzero_si256();
+                    __m256i iacc_1 = _mm256_setzero_si256();
 
-                    const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
-                    const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_00 ,_mm256_shuffle_epi32(rhs_vec_4567_00, 177), 170), _mm256_shuffle_epi32(lhs_vec_00, 0)));
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_00, 177) ,rhs_vec_4567_00, 170), _mm256_shuffle_epi32(lhs_vec_00, 85)));
 
-                    const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
-                    const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_01 ,_mm256_shuffle_epi32(rhs_vec_4567_01, 177), 170), _mm256_shuffle_epi32(lhs_vec_00, 170)));
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_01, 177) ,rhs_vec_4567_01, 170), _mm256_shuffle_epi32(lhs_vec_00, 255)));
 
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_02 ,_mm256_shuffle_epi32(rhs_vec_4567_02, 177), 170), _mm256_shuffle_epi32(lhs_vec_01, 0)));
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_02, 177) ,rhs_vec_4567_02, 170), _mm256_shuffle_epi32(lhs_vec_01, 85)));
 
-                    // Scale values - Load the weight scale values of two block_q4_0x8
-                    const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_03 ,_mm256_shuffle_epi32(rhs_vec_4567_03, 177), 170), _mm256_shuffle_epi32(lhs_vec_01, 170)));
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_03, 177) ,rhs_vec_4567_03, 170), _mm256_shuffle_epi32(lhs_vec_01, 255)));
 
-                    // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
-                    // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector
-                    __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs)));
-                    __m256i lhs_mat_ymm_01_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 0);
-                    __m256i lhs_mat_ymm_23_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 17);
-                    __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32)));
-                    __m256i lhs_mat_ymm_01_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 0);
-                    __m256i lhs_mat_ymm_23_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 17);
-                    __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64)));
-                    __m256i lhs_mat_ymm_01_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 0);
-                    __m256i lhs_mat_ymm_23_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 17);
-                    __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96)));
-                    __m256i lhs_mat_ymm_01_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 0);
-                    __m256i lhs_mat_ymm_23_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 17);
+                    iacc_0 = _mm256_madd_epi16(iacc_0, scales_0);
 
-                    __m512i lhs_mat_01_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_0), lhs_mat_ymm_01_0, 1);
-                    __m512i lhs_mat_23_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_0), lhs_mat_ymm_23_0, 1);
-                    __m512i lhs_mat_01_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_1), lhs_mat_ymm_01_1, 1);
-                    __m512i lhs_mat_23_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_1), lhs_mat_ymm_23_1, 1);
-                    __m512i lhs_mat_01_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_2), lhs_mat_ymm_01_2, 1);
-                    __m512i lhs_mat_23_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_2), lhs_mat_ymm_23_2, 1);
-                    __m512i lhs_mat_01_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_3), lhs_mat_ymm_01_3, 1);
-                    __m512i lhs_mat_23_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_3), lhs_mat_ymm_23_3, 1);
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_10 ,_mm256_shuffle_epi32(rhs_vec_4567_10, 177), 170), _mm256_shuffle_epi32(lhs_vec_10, 0)));
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_10, 177) ,rhs_vec_4567_10, 170), _mm256_shuffle_epi32(lhs_vec_10, 85)));
 
-                    // Shuffle pattern one - left side input
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_11 ,_mm256_shuffle_epi32(rhs_vec_4567_11, 177), 170), _mm256_shuffle_epi32(lhs_vec_10, 170)));
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_11, 177) ,rhs_vec_4567_11, 170), _mm256_shuffle_epi32(lhs_vec_10, 255)));
 
-                    const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
-                    const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_12 ,_mm256_shuffle_epi32(rhs_vec_4567_12, 177), 170), _mm256_shuffle_epi32(lhs_vec_11, 0)));
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_12, 177) ,rhs_vec_4567_12, 170), _mm256_shuffle_epi32(lhs_vec_11, 85)));
 
-                    const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
-                    const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_13 ,_mm256_shuffle_epi32(rhs_vec_4567_13, 177), 170), _mm256_shuffle_epi32(lhs_vec_11, 170)));
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_13, 177) ,rhs_vec_4567_13, 170), _mm256_shuffle_epi32(lhs_vec_11, 255)));
 
-                    const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
-                    const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+                    iacc_1 = _mm256_madd_epi16(iacc_1, scales_1);
 
-                    const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
-                    const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+                    // Accumulate the iacc value for one sb
+                    __m256i iacc_sb = _mm256_add_epi32(iacc_0, iacc_1);
 
-                    // Shuffle pattern two - left side input
+                    // Broadcast the bsums of the two sub blocks  of the iteration of Q8_K across the vector
+                    // Multiply-Add with corresponding mins of Q4_Kx8 with bsums
+                    __m256i q8s_sb = _mm256_shuffle_epi32(q8s, 0);
+                    __m256i iacc_min_sb = _mm256_madd_epi16(q8s_sb, mins_01);
+                    q8s = _mm256_bsrli_epi128(q8s, 4);
 
-                    const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
-                    const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+                    // Accumulate for the complete block
+                    iacc_b = _mm256_add_epi32(iacc_b, iacc_sb);
+                    iacc_min_b = _mm256_add_epi32(iacc_min_b, iacc_min_sb);
+                }
 
-                    const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
-                    const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+                // Multiply-Add with scale values for the complete super block
+                acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_b), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row);
+                acc_min_rows = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_min_b), _mm256_mul_ps(col_dmin_f32, row_scale_f32), acc_min_rows);
 
-                    const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
-                    const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+            }
 
-                    const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
-                    const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+            // Accumulated output values permuted so as to be stored in appropriate order post accumulation
+            acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask);
+            _mm256_storeu_ps(s + (y * nr + x * 8), _mm256_sub_ps(acc_row, acc_min_rows));
+        }
+    }
 
-                    // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
-                    // Resembles MMLAs into 2x2 matrices in ARM Version
-                    const __m512i zero = _mm512_setzero_epi32();
-                    __m512i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_01_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_01_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_01_0_sp1, rhs_mat_014589CD_0_sp1);
-                    __m512i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367ABEF_0_sp1);
-                    __m512i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_23_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_23_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_23_0_sp1, rhs_mat_014589CD_0_sp1);
-                    __m512i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367ABEF_0_sp1);
-                    __m512i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_01_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_01_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_01_0_sp2, rhs_mat_014589CD_0_sp2);
-                    __m512i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367ABEF_0_sp2);
-                    __m512i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_23_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_23_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_23_0_sp2, rhs_mat_014589CD_0_sp2);
-                    __m512i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367ABEF_0_sp2);
+#else
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    ggml_gemv_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+#endif
+}
 
-                    // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
-                    __m512i iacc_mat_00 = _mm512_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
-                    __m512i iacc_mat_01 = _mm512_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
-                    __m512i iacc_mat_10 = _mm512_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
-                    __m512i iacc_mat_11 = _mm512_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+#if defined(__AVX2__)
+    __m256i signextendlut = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i*)kvalues_iq4nl));
+    signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
 
+    gemv_q4_b32_8x8_q8_0_lut_avx<block_iq4_nlx8>(n, s, bs, vx, vy, nr, nc, signextendlut);
 
-                    // Straighten out to make 4 row vectors
-                    __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
-                    __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
-                    __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
-                    __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
+    return;
+#endif
 
-                    // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
-                    const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptr[b].d), loadMask), 68);
-                    const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16);
+    ggml_gemv_iq4_nl_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
+}
 
-                    // Multiply with appropiate scales and accumulate
-                    acc_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)),   acc_rows[0]);
-                    acc_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)),  acc_rows[1]);
-                    acc_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
-                    acc_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]);
-                }
+void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK_K;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
 
-                // Store the accumulated values
-                for (int i = 0; i < 4; i++) {
-                    _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
-                }
-            }
-        }
-        if (anc != nc) {
-            xstart = anc/8;
-            y = 0;
-        }
-    #endif // __AVX512F__
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
 
-        // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
 
-        for (; y < anr / 4; y += 4) {
-            const block_q8_0x4 * a_ptrs[4];
+#if defined(__AVX2__)
+    // Lookup table to convert signed nibbles to signed bytes
+    __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
+    signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
+    // Shuffle masks to rearrange delta values to multiply with appropriate scales
+    __m128i deltamask = _mm_set_epi8(15, 14, 7, 6, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0);
+    // Permute mask used for easier vector processing at later stages
+    __m256i finalpermutemask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
 
-            a_ptrs[0] = a_ptr_start + (y * nb);
-            for (int i = 0; i < 3; ++i) {
-                a_ptrs[i + 1] = a_ptrs[i] + nb;
-            }
+    const __m256i m3b = _mm256_set1_epi8(3);
+    const __m128i m4b_sse = _mm_set1_epi8(0xF);
 
-            // Take group of eight block_q4_0x8 structures at each pass of the loop and perform dot product operation
-            for (int64_t x = xstart; x < nc / 8; x++) {
+    //Mask to get appropriate scales
+    __m128i scalemask1 = _mm_set_epi8(14,14,6,6,12,12,4,4,10,10,2,2,8,8,0,0);
+    __m128i scalemask2 = _mm_set_epi8(15,15,7,7,13,13,5,5,11,11,3,3,9,9,1,1);
 
-                const block_q4_0x8 * b_ptr = b_ptr_start + (x * b_nb);
+    int64_t b_nb = n / QK_K;
 
-                // Master FP accumulators
-                __m256 acc_rows[16];
-                for (int i = 0; i < 16; i++) {
-                    acc_rows[i] = _mm256_setzero_ps();
-                }
+    const block_q2_Kx8 * b_ptr_start = (const block_q2_Kx8 *)vx;
+    const block_q8_K * a_ptr_start = (const block_q8_K *)vy;
 
-                for (int64_t b = 0; b < nb; b++) {
-                    // Load the eight block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
-                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
-                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32));
-                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64));
-                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96));
+    // Process Q8_K blocks one by one
+    for (int64_t y = 0; y < nr; y++) {
 
-                    // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
-                    const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
-                    const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+        // Pointers to LHS blocks of block_q8_K format
+        const block_q8_K * a_ptr = a_ptr_start + (y * nb);
 
-                    // 4-bit -> 8-bit - Sign is maintained
-                    const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7)
-                    const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7)
+        // Take group of eight interleaved block_q2_K structures at each pass of the loop and perform dot product operation
+        for(int64_t x = 0; x < nc / 8; x++) {
 
-                    const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15)
-                    const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15)
+            // Pointers to RHS blocks
+            const block_q2_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
 
-                    const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23)
-                    const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23)
+            // Master FP accumulators
+            __m256 acc_row = _mm256_setzero_ps();
+            __m256 acc_min_rows = _mm256_setzero_ps();
 
-                    const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31)
-                    const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31)
+            for (int64_t b = 0; b < nb; b++) {
 
-                    // Shuffle pattern one - right side input
-                    const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136);  //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3)
-                    const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136);  //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3)
+                // Load and convert to FP32 delta from block_q8_K
+                const __m256 row_scale_f32 = _mm256_set1_ps((a_ptr[b].d));
 
-                    const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136);  //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11)
-                    const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136);  //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11)
+                // Load the delta values for the 8 blocks interleaved in block_q2_Kx8
+                // col_scale_f32 rearranged so as to multiply with appropriate quants
+                const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, deltamask);
+                const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin);
 
-                    const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136);  //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19)
-                    const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136);  //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19)
+                __m256i iacc_b = _mm256_setzero_si256();
+                __m256i iacc_min_b = _mm256_setzero_si256();
 
-                    const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136);  //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27)
-                    const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136);  //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27)
+                // Processes eight sub blocks from each Q2_K in each iteration
+                for(int sb = 0; sb < QK_K / 128; sb++) {
 
-                    // Shuffle pattern two - right side input
+                    // Load the eight block_q2_K for eight sub blocks quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
+                    const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + sb * 256));
+                    const __m256i rhs_raw_vec_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 32 + sb * 256));
+                    const __m256i rhs_raw_vec_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 64 + sb * 256));
+                    const __m256i rhs_raw_vec_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 96 + sb * 256));
+                    const __m256i rhs_raw_vec_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 128 + sb * 256));
+                    const __m256i rhs_raw_vec_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 160 + sb * 256));
+                    const __m256i rhs_raw_vec_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 192 + sb * 256));
+                    const __m256i rhs_raw_vec_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 224 + sb * 256));
 
-                    const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221);  //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7)
-                    const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221);  //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7)
+                    // 2-bit -> 8-bit
+                    // Values of the 0th,2nd,4th,6th sub blocks of eight block_q2_K structures for the sb loop
+                    const __m256i rhs_vec_0123_00 = _mm256_and_si256(rhs_raw_vec_0123_0, m3b); //B00(0-7) B01(0-7) B02(0-7) B03(0-7)
+                    const __m256i rhs_vec_0123_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 2), m3b); //B20(0-7) B21(0-7) B22(0-7) B23(0-7)
+                    const __m256i rhs_vec_0123_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 4), m3b); //B40(0-7) B41(0-7) B42(0-7) B43(0-7)
+                    const __m256i rhs_vec_0123_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 6), m3b); //B60(0-7) B61(0-7) B62(0-7) B63(0-7)
 
-                    const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221);  //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15)
-                    const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221);  //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15)
+                    const __m256i rhs_vec_4567_00 = _mm256_and_si256(rhs_raw_vec_4567_0, m3b); //B04(0-7) B05(0-7) B06(0-7) B07(0-7)
+                    const __m256i rhs_vec_4567_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 2), m3b); //B24(0-7) B25(0-7) B26(0-7) B27(0-7)
+                    const __m256i rhs_vec_4567_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 4), m3b); //B44(0-7) B45(0-7) B46(0-7) B47(0-7)
+                    const __m256i rhs_vec_4567_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 6), m3b); //B64(0-7) B65(0-7) B66(0-7) B67(0-7)
 
-                    const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221);  //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23)
-                    const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221);  //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23)
+                    const __m256i rhs_vec_0123_01 = _mm256_and_si256(rhs_raw_vec_0123_1, m3b); //B00(8-15) B01(8-15) B02(8-15) B03(8-15)
+                    const __m256i rhs_vec_0123_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 2), m3b); //B20(8-15) B21(8-15) B22(8-15) B23(8-15)
+                    const __m256i rhs_vec_0123_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m3b); //B40(8-15) B41(8-15) B42(8-15) B43(8-15)
+                    const __m256i rhs_vec_0123_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 6), m3b); //B60(8-15) B61(8-15) B62(8-15) B63(8-15)
 
-                    const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221);  //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31)
-                    const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221);  //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31)
+                    const __m256i rhs_vec_4567_01 = _mm256_and_si256(rhs_raw_vec_4567_1, m3b); //B04(8-15) B05(8-15) B06(8-15) B07(8-15)
+                    const __m256i rhs_vec_4567_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 2), m3b); //B24(8-15) B25(8-15) B26(8-15) B27(8-15)
+                    const __m256i rhs_vec_4567_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m3b); //B44(8-15) B45(8-15) B46(8-15) B47(8-15)
+                    const __m256i rhs_vec_4567_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 6), m3b); //B64(8-15) B65(8-15) B66(8-15) B67(8-15)
 
-                    // Scale values - Load the wight scale values of block_q4_0x8
-                    const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
+                    // Values of the 1st,3rd,5th,7th sub blocks of eight block_q2_K structures for the sb loop
+                    const __m256i rhs_vec_0123_10 = _mm256_and_si256(rhs_raw_vec_0123_2, m3b); //B10(0-7) B11(0-7) B12(0-7) B13(0-7)
+                    const __m256i rhs_vec_0123_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 2), m3b); //B30(0-7) B31(0-7) B32(0-7) B33(0-7)
+                    const __m256i rhs_vec_0123_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 4), m3b); //B50(0-7) B51(0-7) B52(0-7) B53(0-7)
+                    const __m256i rhs_vec_0123_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 6), m3b); //B70(0-7) B71(0-7) B72(0-7) B73(0-7)
 
-                    // Process LHS in groups of four
-                    for (int rp = 0; rp < 4; rp++) {
-                        // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
-                        // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
-                        __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs)));
-                        __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0);
-                        __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17);
-                        __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32)));
-                        __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0);
-                        __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17);
-                        __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64)));
-                        __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0);
-                        __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17);
-                        __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96)));
-                        __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0);
-                        __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17);
+                    const __m256i rhs_vec_4567_10 = _mm256_and_si256(rhs_raw_vec_4567_2, m3b); //B14(0-7) B15(0-7) B16(0-7) B17(0-7)
+                    const __m256i rhs_vec_4567_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 2), m3b); //B34(0-7) B35(0-7) B36(0-7) B37(0-7)
+                    const __m256i rhs_vec_4567_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 4), m3b); //B54(0-7) B55(0-7) B56(0-7) B57(0-7)
+                    const __m256i rhs_vec_4567_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 6), m3b); //B74(0-7) B75(0-7) B76(0-7) B77(0-7)
 
-                        // Shuffle pattern one - left side input
-                        const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
-                        const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+                    const __m256i rhs_vec_0123_11 = _mm256_and_si256(rhs_raw_vec_0123_3, m3b); //B10(8-15) B11(8-15) B12(8-15) B13(8-15)
+                    const __m256i rhs_vec_0123_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 2), m3b); //B30(8-15) B31(8-15) B32(8-15) B33(8-15)
+                    const __m256i rhs_vec_0123_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 4), m3b); //B50(8-15) B51(8-15) B52(8-15) B53(8-15)
+                    const __m256i rhs_vec_0123_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 6), m3b); //B70(8-15) B71(8-15) B72(8-15) B73(8-15)
 
-                        const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
-                        const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+                    const __m256i rhs_vec_4567_11 = _mm256_and_si256(rhs_raw_vec_4567_3, m3b); //B14(8-15) B15(8-15) B16(8-15) B17(8-15)
+                    const __m256i rhs_vec_4567_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 2), m3b); //B34(8-15) B35(8-15) B36(8-15) B37(8-15)
+                    const __m256i rhs_vec_4567_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 4), m3b); //B54(8-15) B55(8-15) B56(8-15) B57(8-15)
+                    const __m256i rhs_vec_4567_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 6), m3b); //B74(8-15) B75(8-15) B76(8-15) B77(8-15)
 
-                        const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
-                        const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+                    //Scales and Mins of corresponding sub blocks from different Q2_K structures are stored together
+                    //s00 m00  s01 m01   s10 m10  s11 m11  s20 m20  s21 m21   s30 m30  s31 m31  s40 m40  s41 m41   s50 m50  s51 m51  s60 m60  s61 m61   s70 m70  s71 m71
 
-                        const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
-                        const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+                    const __m128i mins_and_scales_01 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + sb * 64));
+                    const __m128i mins_and_scales_23 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 16 + sb * 64));
+                    const __m128i mins_and_scales_45 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 32 + sb * 64));
+                    const __m128i mins_and_scales_67 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 48 + sb * 64));
 
-                        // Shuffle pattern two - left side input
-                        const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
-                        const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+                    // Extract scales which is lower half from mins_and_scales
+                    const __m128i scales_01 = _mm_and_si128(mins_and_scales_01, m4b_sse);
+                    const __m128i scales_23 = _mm_and_si128(mins_and_scales_23, m4b_sse);
+                    const __m128i scales_45 = _mm_and_si128(mins_and_scales_45, m4b_sse);
+                    const __m128i scales_67 = _mm_and_si128(mins_and_scales_67, m4b_sse);
 
-                        const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
-                        const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+                    // Extract mins which is upper half from mins_and_scales
+                    const __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_01, 4), m4b_sse));
+                    const __m256i mins_23 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_23, 4), m4b_sse));
+                    const __m256i mins_45 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_45, 4), m4b_sse));
+                    const __m256i mins_67 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_67, 4), m4b_sse));
 
-                        const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
-                        const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+                    // Scales of sub blocks in the sb loop
+                    // Scales of the 0th sub block from each super block
+                    __m128i scales_rearrange_0 = _mm_shuffle_epi8(scales_01, scalemask1);
+                    __m256i scales_0 = _mm256_cvtepu8_epi16(scales_rearrange_0);
 
-                        const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
-                        const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+                    // Scales of the 1st sub block from each super block
+                    __m128i scales_rearrange_1 = _mm_shuffle_epi8(scales_01, scalemask2);
+                    __m256i scales_1 = _mm256_cvtepu8_epi16(scales_rearrange_1);
 
-                        // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
-                        // Resembles MMLAs into 2x2 matrices in ARM Version
-                        const __m256i zero = _mm256_setzero_si256();
-                        __m256i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1);
-                        __m256i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1);
-                        __m256i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1);
-                        __m256i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1);
-                        __m256i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2);
-                        __m256i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2);
-                        __m256i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2);
-                        __m256i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2);
+                    // Scales of the 2nd sub block from each super block
+                    __m128i scales_rearrange_2 = _mm_shuffle_epi8(scales_23, scalemask1);
+                    __m256i scales_2 = _mm256_cvtepu8_epi16(scales_rearrange_2);
 
-                        // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
-                        __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
-                        __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
-                        __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
-                        __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+                    // Scales of the 3rd sub block from each super block
+                    __m128i scales_rearrange_3 = _mm_shuffle_epi8(scales_23, scalemask2);
+                    __m256i scales_3 = _mm256_cvtepu8_epi16(scales_rearrange_3);
 
-                        // Straighten out to make 4 row vectors
-                        __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
-                        __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
-                        __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
-                        __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
+                    // Scales of the 4th sub block from each super block
+                    __m128i scales_rearrange_4 = _mm_shuffle_epi8(scales_45, scalemask1);
+                    __m256i scales_4 = _mm256_cvtepu8_epi16(scales_rearrange_4);
 
-                        // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
-                        const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask);
+                    // Scales of the 5th sub block from each super block
+                    __m128i scales_rearrange_5 = _mm_shuffle_epi8(scales_45, scalemask2);
+                    __m256i scales_5 = _mm256_cvtepu8_epi16(scales_rearrange_5);
 
-                        // Multiply with appropiate scales and accumulate
-                        acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
-                        acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
-                        acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
-                        acc_rows[rp * 4 + 3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32,  255)), acc_rows[rp * 4 + 3]);
-                    }
-                }
+                    // Scales of the 6th sub block from each super block
+                    __m128i scales_rearrange_6 = _mm_shuffle_epi8(scales_67, scalemask1);
+                    __m256i scales_6 = _mm256_cvtepu8_epi16(scales_rearrange_6);
 
-                // Store the accumulated values
-                for (int i = 0; i < 16; i++) {
-                    _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
-                }
-            }
-        }
+                    // Scales of the 7th sub block from each super block
+                    __m128i scales_rearrange_7 = _mm_shuffle_epi8(scales_67, scalemask2);
+                    __m256i scales_7 = _mm256_cvtepu8_epi16(scales_rearrange_7);
 
-        // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation
-        for (; y < nr / 4; y ++) {
+                    // Load the sub block values corresponding to sb in block_q8_K in batches of 16 bytes and replicate the same across 256 bit vector
+                    __m256i lhs_vec_0 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + sb * 128)));
+                    __m256i lhs_vec_1 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 16 + sb * 128)));
+                    __m256i lhs_vec_2 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 32 + sb * 128)));
+                    __m256i lhs_vec_3 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 48 + sb * 128)));
+                    __m256i lhs_vec_4 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 64 + sb * 128)));
+                    __m256i lhs_vec_5 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 80 + sb * 128)));
+                    __m256i lhs_vec_6 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 96 + sb * 128)));
+                    __m256i lhs_vec_7 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 112 + sb * 128)));
 
-            const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb);
+                    lhs_vec_0 = _mm256_permute2f128_si256(lhs_vec_0, lhs_vec_0, 0);
+                    lhs_vec_1 = _mm256_permute2f128_si256(lhs_vec_1, lhs_vec_1, 0);
+                    lhs_vec_2 = _mm256_permute2f128_si256(lhs_vec_2, lhs_vec_2, 0);
+                    lhs_vec_3 = _mm256_permute2f128_si256(lhs_vec_3, lhs_vec_3, 0);
+                    lhs_vec_4 = _mm256_permute2f128_si256(lhs_vec_4, lhs_vec_4, 0);
+                    lhs_vec_5 = _mm256_permute2f128_si256(lhs_vec_5, lhs_vec_5, 0);
+                    lhs_vec_6 = _mm256_permute2f128_si256(lhs_vec_6, lhs_vec_6, 0);
+                    lhs_vec_7 = _mm256_permute2f128_si256(lhs_vec_7, lhs_vec_7, 0);
 
-            // Load the eight block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
-            for (int64_t x = xstart; x < nc / 8; x++) {
+                    __m256i iacc_0 = _mm256_setzero_si256();
+                    __m256i iacc_1 = _mm256_setzero_si256();
+                    __m256i iacc_2 = _mm256_setzero_si256();
+                    __m256i iacc_3 = _mm256_setzero_si256();
+                    __m256i iacc_4 = _mm256_setzero_si256();
+                    __m256i iacc_5 = _mm256_setzero_si256();
+                    __m256i iacc_6 = _mm256_setzero_si256();
+                    __m256i iacc_7 = _mm256_setzero_si256();
 
-                const block_q4_0x8 * b_ptr = b_ptr_start + (x * b_nb);
+                    // Dot product done within 32 bit lanes and accumulated in the same vector
+                    // First done for 0th sub block and then for seven (1st - 7th) other sub blocks processed for each sb (sb < QK_K/128 loop)                    // B0(0-3) B4(0-3) B1(0-3) B5(0-3) B2(0-3) B6(0-3) B3(0-3) B7(0-3) with A0(0-3)
+                    // B0(4-7) B4(4-7) B1(4-7) B5(4-7) B2(4-7) B6(4-7) B3(4-7) B7(4-7) with A0(4-7)
+                    // B0(8-11) B4(8-11) B1(8-11) B5(8-11) B2(8-11) B6(8-11) B3(8-11) B7(8-11) with A0(8-11)
+                    // B0(12-15) B4(12-15) B1(12-15) B5(12-15) B2(12-15) B6(12-15) B3(12-15) B7(12-15) with A0(12-15)
 
-                // Master FP accumulators
-                __m256 acc_rows[4];
-                for (int i = 0; i < 4; i++) {
-                    acc_rows[i] = _mm256_setzero_ps();
-                }
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_00 ,_mm256_shuffle_epi32(rhs_vec_4567_00, 177), 170), _mm256_shuffle_epi32(lhs_vec_0, 0)));
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_00, 177) ,rhs_vec_4567_00, 170), _mm256_shuffle_epi32(lhs_vec_0, 85)));
 
-                for (int64_t b = 0; b < nb; b++) {
-                    // Load the eight block_q8_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
-                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
-                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32));
-                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64));
-                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96));
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_01 ,_mm256_shuffle_epi32(rhs_vec_4567_01, 177), 170), _mm256_shuffle_epi32(lhs_vec_0, 170)));
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_01, 177) ,rhs_vec_4567_01, 170), _mm256_shuffle_epi32(lhs_vec_0, 255)));
 
-                    // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess
-                    const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
-                    const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+                    iacc_0 = _mm256_madd_epi16(iacc_0, scales_0);
 
-                    // 4-bit -> 8-bit - Sign is maintained
-                    const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b));  //B0(0-7) B1(0-7) B4(0-7) B5(0-7)
-                    const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b));  //B2(0-7) B3(0-7) B6(0-7) B7(0-7)
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_10 ,_mm256_shuffle_epi32(rhs_vec_4567_10, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 0)));
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_10, 177) ,rhs_vec_4567_10, 170), _mm256_shuffle_epi32(lhs_vec_1, 85)));
 
-                    const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b));  //B0(8-15) B1(8-15) B4(8-15) B5(8-15)
-                    const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b));  //B2(8-15) B3(8-15) B6(8-15) B7(8-15)
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_11 ,_mm256_shuffle_epi32(rhs_vec_4567_11, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 170)));
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_11, 177) ,rhs_vec_4567_11, 170), _mm256_shuffle_epi32(lhs_vec_1, 255)));
 
-                    const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b));  //B0(16-23) B1(16-23) B4(16-23) B5(16-23)
-                    const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b));  //B2(16-23) B3(16-23) B6(16-23) B7(16-23)
+                    iacc_1 = _mm256_madd_epi16(iacc_1, scales_1);
 
-                    const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b));  //B0(24-31) B1(24-31) B4(24-31) B5(24-31)
-                    const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b));  //B2(24-31) B3(24-31) B6(24-31) B7(24-31)
+                    iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_20 ,_mm256_shuffle_epi32(rhs_vec_4567_20, 177), 170), _mm256_shuffle_epi32(lhs_vec_2, 0)));
+                    iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_20, 177) ,rhs_vec_4567_20, 170), _mm256_shuffle_epi32(lhs_vec_2, 85)));
 
-                    // Shuffle pattern one - right side input
-                    const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136);  //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3)
-                    const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136);  //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3)
+                    iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_21 ,_mm256_shuffle_epi32(rhs_vec_4567_21, 177), 170), _mm256_shuffle_epi32(lhs_vec_2, 170)));
+                    iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_21, 177) ,rhs_vec_4567_21, 170), _mm256_shuffle_epi32(lhs_vec_2, 255)));
 
-                    const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136);  //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11)
-                    const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136);  //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11)
+                    iacc_2 = _mm256_madd_epi16(iacc_2, scales_2);
 
-                    const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136);  //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19)
-                    const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136);  //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19)
+                    iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_30 ,_mm256_shuffle_epi32(rhs_vec_4567_30, 177), 170), _mm256_shuffle_epi32(lhs_vec_3, 0)));
+                    iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_30, 177) ,rhs_vec_4567_30, 170), _mm256_shuffle_epi32(lhs_vec_3, 85)));
 
-                    const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136);  //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27)
-                    const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136);  //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27)
+                    iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_31 ,_mm256_shuffle_epi32(rhs_vec_4567_31, 177), 170), _mm256_shuffle_epi32(lhs_vec_3, 170)));
+                    iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_31, 177) ,rhs_vec_4567_31, 170), _mm256_shuffle_epi32(lhs_vec_3, 255)));
 
-                    // Shuffle pattern two - right side input
+                    iacc_3 = _mm256_madd_epi16(iacc_3, scales_3);
 
-                    const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221);  //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7)
-                    const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221);  //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7)
+                    iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_40 ,_mm256_shuffle_epi32(rhs_vec_4567_40, 177), 170), _mm256_shuffle_epi32(lhs_vec_4, 0)));
+                    iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_40, 177) ,rhs_vec_4567_40, 170), _mm256_shuffle_epi32(lhs_vec_4, 85)));
 
-                    const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221);  //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15)
-                    const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221);  //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15)
+                    iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_41 ,_mm256_shuffle_epi32(rhs_vec_4567_41, 177), 170), _mm256_shuffle_epi32(lhs_vec_4, 170)));
+                    iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_41, 177) ,rhs_vec_4567_41, 170), _mm256_shuffle_epi32(lhs_vec_4, 255)));
 
-                    const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221);  //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23)
-                    const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221);  //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23)
+                    iacc_4 = _mm256_madd_epi16(iacc_4, scales_4);
 
-                    const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221);  //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31)
-                    const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221);  //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31)
+                    iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_50 ,_mm256_shuffle_epi32(rhs_vec_4567_50, 177), 170), _mm256_shuffle_epi32(lhs_vec_5, 0)));
+                    iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_50, 177) ,rhs_vec_4567_50, 170), _mm256_shuffle_epi32(lhs_vec_5, 85)));
 
-                    // Scale values - Load the wight scale values of block_q4_0x8
-                    const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
+                    iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_51 ,_mm256_shuffle_epi32(rhs_vec_4567_51, 177), 170), _mm256_shuffle_epi32(lhs_vec_5, 170)));
+                    iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_51, 177) ,rhs_vec_4567_51, 170), _mm256_shuffle_epi32(lhs_vec_5, 255)));
 
-                    // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
-                    // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
-                    __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs)));
-                    __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0);
-                    __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17);
-                    __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32)));
-                    __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0);
-                    __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17);
-                    __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64)));
-                    __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0);
-                    __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17);
-                    __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96)));
-                    __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0);
-                    __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17);
+                    iacc_5 = _mm256_madd_epi16(iacc_5, scales_5);
 
-                    // Shuffle pattern one - left side input
+                    iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_60 ,_mm256_shuffle_epi32(rhs_vec_4567_60, 177), 170), _mm256_shuffle_epi32(lhs_vec_6, 0)));
+                    iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_60, 177) ,rhs_vec_4567_60, 170), _mm256_shuffle_epi32(lhs_vec_6, 85)));
 
-                    const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
-                    const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+                    iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_61 ,_mm256_shuffle_epi32(rhs_vec_4567_61, 177), 170), _mm256_shuffle_epi32(lhs_vec_6, 170)));
+                    iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_61, 177) ,rhs_vec_4567_61, 170), _mm256_shuffle_epi32(lhs_vec_6, 255)));
 
-                    const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
-                    const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+                    iacc_6 = _mm256_madd_epi16(iacc_6, scales_6);
 
-                    const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
-                    const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+                    iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_70 ,_mm256_shuffle_epi32(rhs_vec_4567_70, 177), 170), _mm256_shuffle_epi32(lhs_vec_7, 0)));
+                    iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_70, 177) ,rhs_vec_4567_70, 170), _mm256_shuffle_epi32(lhs_vec_7, 85)));
 
-                    const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
-                    const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+                    iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_71 ,_mm256_shuffle_epi32(rhs_vec_4567_71, 177), 170), _mm256_shuffle_epi32(lhs_vec_7, 170)));
+                    iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_71, 177) ,rhs_vec_4567_71, 170), _mm256_shuffle_epi32(lhs_vec_7, 255)));
 
-                    // Shuffle pattern two - left side input
+                    iacc_7 = _mm256_madd_epi16(iacc_7, scales_7);
 
-                    const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
-                    const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+                    // Accumulate the iacc value for one sb
+                    __m256i iacc_sb = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_0, iacc_1), _mm256_add_epi32(iacc_2, iacc_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_4, iacc_5), _mm256_add_epi32(iacc_6, iacc_7)));
 
-                    const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
-                    const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+                    __m128i q8sums = _mm_loadu_si128((const __m128i *)(a_ptr[b].bsums + sb * 8));
+                    __m256i q8s = _mm256_castsi128_si256(q8sums);
+                    q8s= _mm256_permute2f128_si256(q8s, q8s, 0);
 
-                    const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
-                    const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+                    // Broadcast the bsums of the two corresponding subblocks of q8_k
+                    // Multiply-Add with corresponding mins of Q2_Kx8 with bsums
+                    __m256i iacc_min_sb_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 0), mins_01);
+                    __m256i iacc_min_sb_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 85), mins_23);
+                    __m256i iacc_min_sb_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 170), mins_45);
+                    __m256i iacc_min_sb_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 255), mins_67);
 
-                    const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
-                    const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+                    __m256i iacc_min_sb = _mm256_add_epi32(_mm256_add_epi32(iacc_min_sb_01, iacc_min_sb_23), _mm256_add_epi32(iacc_min_sb_45,iacc_min_sb_67));
 
-                    // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
-                    // Resembles MMLAs into 2x2 matrices in ARM Version
-                    const __m256i zero = _mm256_setzero_si256();
-                    __m256i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1);
-                    __m256i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1);
-                    __m256i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1);
-                    __m256i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1);
-                    __m256i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2);
-                    __m256i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2);
-                    __m256i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2);
-                    __m256i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2);
+                    // Accumulate for the complete block
+                    iacc_b = _mm256_add_epi32(iacc_b, iacc_sb);
+                    iacc_min_b = _mm256_add_epi32(iacc_min_b, iacc_min_sb);
+                }
 
-                    // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
-                    __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
-                    __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
-                    __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
-                    __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+                //Multiply-Add with scale values for complete super block
+                acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_b), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row);
+                acc_min_rows = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_min_b), _mm256_mul_ps(col_dmin_f32, row_scale_f32), acc_min_rows);
+            }
+            // Accumulated output values permuted so as to be stored in appropriate order post accumulation
+            acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask);
+            _mm256_storeu_ps(s + (y * nr + x * 8), _mm256_sub_ps(acc_row, acc_min_rows));
+        }
+    }
+#else
 
+    ggml_gemv_q2_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
 
-                    // Straighten out to make 4 row vectors
-                    __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
-                    __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
-                    __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
-                    __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
+#endif
+}
 
-                    // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
-                    const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptr[b].d, loadMask);
+void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+#if defined(__AVX2__) || defined(__AVX512F__)
+    {
+        // Lookup table to convert signed nibbles to signed bytes
+        __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
+        signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
 
-                    // Multiply with appropiate scales and accumulate
-                    acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
-                    acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
-                    acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
-                    acc_rows[3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]);
-                }
+        gemm_q4_b32_8x8_q8_0_lut_avx<block_q4_0x8>(n, s, bs, vx, vy, nr, nc, signextendlut);
 
-                // Store the accumulated values
-                for (int i = 0; i < 4; i++) {
-                    _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
-                }
-            }
-        }
         return;
     }
+#endif // defined(__AVX2__) || defined(__AVX512F__)
 
-#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
     ggml_gemm_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
@@ -3364,6 +3408,21 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 #endif
 }
 
+void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+#if defined(__AVX2__) || defined(__AVX512F__)
+    {
+        __m256i signextendlut = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i*)kvalues_iq4nl));
+        signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
+
+        gemm_q4_b32_8x8_q8_0_lut_avx<block_iq4_nlx8>(n, s, bs, vx, vy, nr, nc, signextendlut);
+
+        return;
+    }
+#endif // defined(__AVX2__) || defined(__AVX512F__)
+
+    ggml_gemm_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
 void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK_K;
     const int nb = n / qk;
diff --git a/ggml/src/ggml-cpu/repack.cpp b/ggml/src/ggml-cpu/repack.cpp
index 2583aefae47..f531d21e232 100644
--- a/ggml/src/ggml-cpu/repack.cpp
+++ b/ggml/src/ggml-cpu/repack.cpp
@@ -206,8 +206,9 @@ void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
     const int ncols_interleaved = 4;
     const int blocklen = 4;
 
-    assert (n % qk == 0);
-    assert (nc % ncols_interleaved == 0);
+    assert(nr == 1);
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
 
     UNUSED(s);
     UNUSED(bs);
@@ -307,30 +308,28 @@ void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
-    {
-        float sumf[8];
-        int sumi;
+    float sumf[8];
+    int sumi;
 
-        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
 
-            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int j = 0; j < ncols_interleaved; j++) {
-                        sumi = 0;
-                        for (int i = 0; i < blocklen; ++i) {
-                            const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
-                        }
-                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                     }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                 }
             }
-            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
         }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
     }
 }
 
@@ -494,43 +493,73 @@ void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
     const int ncols_interleaved = 4;
     const int blocklen = 4;
 
-    assert (n % qk == 0);
-    assert (nc % ncols_interleaved == 0);
+    assert(nr == 1);
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
 
-    UNUSED(s);
     UNUSED(bs);
-    UNUSED(vx);
-    UNUSED(vy);
     UNUSED(nr);
-    UNUSED(nc);
-    UNUSED(nb);
-    UNUSED(ncols_interleaved);
-    UNUSED(blocklen);
 
-    {
-        float sumf[4];
-        int sumi;
+    float sumf[4];
+    int sumi;
 
-        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
 
-            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int j = 0; j < ncols_interleaved; j++) {
-                        sumi = 0;
-                        for (int i = 0; i < blocklen; ++i) {
-                            const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
-                            const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
-                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
-                        }
-                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                        const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
                     }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                 }
             }
-            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
         }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+}
+
+void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert(nr == 1);
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float sumf[8];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_iq4_nlx8 * b_ptr = (const block_iq4_nlx8 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                        const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
+                    }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
     }
 }
 
@@ -934,6 +963,50 @@ void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
     }
 }
 
+void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    float sumf[4][8];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_iq4_nlx8 * b_ptr = (const block_iq4_nlx8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                                const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
+                            }
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+}
+
 } // extern "C"
 
 static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) {
@@ -1285,15 +1358,16 @@ static block_iq4_nlx4 make_block_iq4_nlx4(block_iq4_nl * in, unsigned int blck_s
 
 static int repack_iq4_nl_to_iq4_nl_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
     GGML_ASSERT(t->type == GGML_TYPE_IQ4_NL);
-    //GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
     GGML_ASSERT(interleave_block == 4);
 
-    block_iq4_nlx4 * dst = (block_iq4_nlx4 *)t->data;
-    const block_iq4_nl * src = (const block_iq4_nl *)data;
+    const block_iq4_nl   * src = (const block_iq4_nl   *)data;
+          block_iq4_nlx4 * dst = (      block_iq4_nlx4 *)t->data;
+
     block_iq4_nl dst_tmp[4];
+
     int nrow = ggml_nrows(t);
     int nrows_interleaved = 4;
-    int nblocks = t->ne[0] / QK4_0;
+    int nblocks = t->ne[0] / QK4_NL;
 
     GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_iq4_nl));
 
@@ -1315,6 +1389,63 @@ static int repack_iq4_nl_to_iq4_nl_4_bl(struct ggml_tensor * t, int interleave_b
     GGML_UNUSED(data_size);
 }
 
+static block_iq4_nlx8 make_block_iq4_nlx8(block_iq4_nl * in, unsigned int blck_size_interleave) {
+    block_iq4_nlx8 out;
+
+    for (int i = 0; i < 8; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    const int end = QK4_NL * 4 / blck_size_interleave;
+
+    if (blck_size_interleave == 8) {
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 8;
+            int src_offset = (i / 8) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint64_t));
+        }
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    return out;
+}
+
+static int repack_iq4_nl_to_iq4_nl_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_IQ4_NL);
+    GGML_ASSERT(interleave_block == 8);
+
+    const block_iq4_nl   * src = (const block_iq4_nl   *)data;
+          block_iq4_nlx8 * dst = (      block_iq4_nlx8 *)t->data;
+
+    block_iq4_nl dst_tmp[8];
+
+    int nrow = ggml_nrows(t);
+    int nrows_interleaved = 8;
+    int nblocks = t->ne[0] / QK4_NL;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_iq4_nl));
+
+    if (t->ne[1] % nrows_interleaved != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i = 0; i < nrows_interleaved; i++) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_iq4_nlx8(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+
 namespace ggml::cpu::repack {
 // repack
 template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS>
@@ -1350,6 +1481,10 @@ template <> int repack<block_iq4_nl, 4, 4>(struct ggml_tensor * t, const void *
 //    return repack_iq4_nl_to_iq4_nl_4_bl(t, 8, data, data_size);
 //}
 
+template <> int repack<block_iq4_nl, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_iq4_nl_to_iq4_nl_8_bl(t, 8, data, data_size);
+}
+
 // gemv
 template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE>
 void gemv(int, float *, size_t, const void *, const void *, int, int);
@@ -1378,6 +1513,10 @@ template <> void gemv<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size
     ggml_gemv_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemv<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_iq4_nl_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
 // gemm
 template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE>
 void gemm(int, float *, size_t, const void *, const void *, int, int);
@@ -1406,6 +1545,10 @@ template <> void gemm<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size
     ggml_gemm_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemm<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_iq4_nl_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
 class tensor_traits_base : public ggml::cpu::tensor_traits {
   public:
     virtual int repack(struct ggml_tensor * t, const void * data, size_t data_size) = 0;
@@ -1680,6 +1823,7 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
 
     // instance for IQ4
     static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0> iq4_nl_4x4_q8_0;
+    static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0> iq4_nl_8x8_q8_0;
 
     if (cur->type == GGML_TYPE_Q4_0) {
         if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)) {
@@ -1710,6 +1854,11 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
             }
         }
     } else if (cur->type == GGML_TYPE_IQ4_NL) {
+        if (ggml_cpu_has_avx2()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &iq4_nl_8x8_q8_0;
+            }
+        }
         if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
             if (cur->ne[1] % 4 == 0) {
                 return &iq4_nl_4x4_q8_0;
diff --git a/ggml/src/ggml-cpu/repack.h b/ggml/src/ggml-cpu/repack.h
index cd322e74389..cb32b503d3a 100644
--- a/ggml/src/ggml-cpu/repack.h
+++ b/ggml/src/ggml-cpu/repack.h
@@ -67,6 +67,13 @@ struct block_iq4_nlx4 {
 
 static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding");
 
+struct block_iq4_nlx8 {
+    ggml_half d[8];            // deltas for 8 iq4_nl blocks
+    uint8_t   qs[QK4_NL * 4];  // nibbles / quants for 8 iq4_nl blocks
+};
+
+static_assert(sizeof(block_iq4_nlx8) == 8 * sizeof(ggml_half) + QK4_NL * 4, "wrong iq4_nlx8 block size/padding");
+
 #if defined(__cplusplus)
 extern "C" {
 #endif
@@ -80,12 +87,14 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
 void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 
 // Native implementations
 void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
@@ -97,12 +106,14 @@ void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
 void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 
 #if defined(__cplusplus)
 } // extern "C"

From 37102a5848e2c315dfc5d07d8252c15ff7553b28 Mon Sep 17 00:00:00 2001
From: Judd <4046440+foldl@users.noreply.github.com>
Date: Wed, 13 Aug 2025 18:45:15 +0800
Subject: [PATCH 61/86] ggml : update `ggml_rope_multi` (llama/12665)

* update `rope_multi`:

1. add `ggml_rope_multi_inplace`;
1. use `GGML_MROPE_SECTIONS` instead of 4.

* Apply suggestions from code review

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---
 ggml/include/ggml.h | 20 ++++++++++-
 ggml/src/ggml.c     | 84 ++++++++++++++++++++++++---------------------
 2 files changed, 63 insertions(+), 41 deletions(-)

diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
index 2f06e1e39b2..c09d82a91cf 100644
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -241,6 +241,8 @@
 #define GGML_ROPE_TYPE_MROPE  8
 #define GGML_ROPE_TYPE_VISION 24
 
+#define GGML_MROPE_SECTIONS   4
+
 #define GGML_UNUSED(x) (void)(x)
 
 #define GGML_PAD(x, n) (((x) + (n) - 1) & ~((n) - 1))
@@ -1660,7 +1662,7 @@ extern "C" {
             struct ggml_tensor  * b,
             struct ggml_tensor  * c,
             int                   n_dims,
-            int                   sections[4],
+            int                   sections[GGML_MROPE_SECTIONS],
             int                   mode,
             int                   n_ctx_orig,
             float                 freq_base,
@@ -1686,6 +1688,22 @@ extern "C" {
             float                 beta_fast,
             float                 beta_slow);
 
+    GGML_API struct ggml_tensor * ggml_rope_multi_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            struct ggml_tensor  * c,
+            int                   n_dims,
+            int                   sections[GGML_MROPE_SECTIONS],
+            int                   mode,
+            int                   n_ctx_orig,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow);
+
     GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_rope_custom(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 55a76f8248c..aff0b47cb26 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -3885,6 +3885,7 @@ static struct ggml_tensor * ggml_rope_impl(
         struct ggml_tensor  * b,
         struct ggml_tensor  * c,
         int                   n_dims,
+        int                   sections[GGML_MROPE_SECTIONS],
         int                   mode,
         int                   n_ctx_orig,
         float                 freq_base,
@@ -3898,15 +3899,19 @@ static struct ggml_tensor * ggml_rope_impl(
 
     GGML_ASSERT(ggml_is_vector(b));
     GGML_ASSERT(b->type == GGML_TYPE_I32);
-    GGML_ASSERT(a->ne[2] == b->ne[0]);
+
+    bool mrope_used = mode & GGML_ROPE_TYPE_MROPE;
+    if (mrope_used) {
+        GGML_ASSERT(a->ne[2] * 4 == b->ne[0]); // mrope expecting 4 position ids per token
+    } else {
+        GGML_ASSERT(a->ne[2] == b->ne[0]);
+    }
 
     if (c) {
         GGML_ASSERT(c->type == GGML_TYPE_F32);
         GGML_ASSERT(c->ne[0] >= n_dims / 2);
     }
 
-    int sections[4] = {0, 0, 0, 0};
-
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     int32_t params[15] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig };
@@ -3916,7 +3921,11 @@ static struct ggml_tensor * ggml_rope_impl(
     memcpy(params +  8, &attn_factor,  sizeof(float));
     memcpy(params +  9, &beta_fast,    sizeof(float));
     memcpy(params + 10, &beta_slow,    sizeof(float));
-    memcpy(params + 11, &sections,     sizeof(int)*4);
+    if (mrope_used) {
+        memcpy(params + 11, sections,  sizeof(int32_t) * GGML_MROPE_SECTIONS);
+    } else {
+        memset(params + 11, 0,         sizeof(int32_t) * GGML_MROPE_SECTIONS);
+    }
     ggml_set_op_params(result, params, sizeof(params));
 
     result->op     = GGML_OP_ROPE;
@@ -3934,7 +3943,7 @@ struct ggml_tensor * ggml_rope(
         int                   n_dims,
         int                   mode) {
     return ggml_rope_impl(
-        ctx, a, b, NULL, n_dims, mode, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, false
+        ctx, a, b, NULL, n_dims, NULL, mode, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, false
     );
 }
 
@@ -3944,7 +3953,7 @@ struct ggml_tensor * ggml_rope_multi(
         struct ggml_tensor  * b,
         struct ggml_tensor  * c,
         int                   n_dims,
-        int                   sections[4],
+        int                   sections[GGML_MROPE_SECTIONS],
         int                   mode,
         int                   n_ctx_orig,
         float                 freq_base,
@@ -3953,36 +3962,31 @@ struct ggml_tensor * ggml_rope_multi(
         float                 attn_factor,
         float                 beta_fast,
         float                 beta_slow) {
-    // Multimodal Rotary Position Embedding
-    GGML_ASSERT((mode & 1) == 0 && "mode & 1 == 1 is no longer supported");
-
-    GGML_ASSERT(ggml_is_vector(b));
-    GGML_ASSERT(b->type == GGML_TYPE_I32);
-    GGML_ASSERT(a->ne[2] * 4 == b->ne[0]); // mrope expecting 4 position ids per token
-
-    if (c) {
-        GGML_ASSERT(c->type == GGML_TYPE_F32);
-        GGML_ASSERT(c->ne[0] >= n_dims / 2);
-    }
-
-    struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
-
-    int32_t params[11 + 4] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig };
-    memcpy(params +  5, &freq_base,    sizeof(float));
-    memcpy(params +  6, &freq_scale,   sizeof(float));
-    memcpy(params +  7, &ext_factor,   sizeof(float));
-    memcpy(params +  8, &attn_factor,  sizeof(float));
-    memcpy(params +  9, &beta_fast,    sizeof(float));
-    memcpy(params + 10, &beta_slow,    sizeof(float));
-    memcpy(&params[11], sections,      sizeof(int)*4);
-    ggml_set_op_params(result, params, sizeof(params));
-
-    result->op   = GGML_OP_ROPE;
-    result->src[0] = a;
-    result->src[1] = b;
-    result->src[2] = c;
+    return ggml_rope_impl(
+        ctx, a, b, c, n_dims, sections, mode, n_ctx_orig, freq_base, freq_scale,
+        ext_factor, attn_factor, beta_fast, beta_slow, false
+    );
+}
 
-    return result;
+struct ggml_tensor * ggml_rope_multi_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        struct ggml_tensor  * c,
+        int                   n_dims,
+        int                   sections[GGML_MROPE_SECTIONS],
+        int                   mode,
+        int                   n_ctx_orig,
+        float                 freq_base,
+        float                 freq_scale,
+        float                 ext_factor,
+        float                 attn_factor,
+        float                 beta_fast,
+        float                 beta_slow) {
+    return ggml_rope_impl(
+        ctx, a, b, c, n_dims, sections, mode, n_ctx_orig, freq_base, freq_scale,
+        ext_factor, attn_factor, beta_fast, beta_slow, true
+    );
 }
 
 struct ggml_tensor * ggml_rope_inplace(
@@ -3992,7 +3996,7 @@ struct ggml_tensor * ggml_rope_inplace(
         int                   n_dims,
         int                   mode) {
     return ggml_rope_impl(
-        ctx, a, b, NULL, n_dims, mode, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, true
+        ctx, a, b, NULL, n_dims, NULL, mode, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, true
     );
 }
 
@@ -4011,7 +4015,7 @@ struct ggml_tensor * ggml_rope_ext(
         float                 beta_fast,
         float                 beta_slow) {
     return ggml_rope_impl(
-        ctx, a, b, c, n_dims, mode, n_ctx_orig, freq_base, freq_scale,
+        ctx, a, b, c, n_dims, NULL, mode, n_ctx_orig, freq_base, freq_scale,
         ext_factor, attn_factor, beta_fast, beta_slow, false
     );
 }
@@ -4031,7 +4035,7 @@ struct ggml_tensor * ggml_rope_ext_inplace(
         float                 beta_fast,
         float                 beta_slow) {
     return ggml_rope_impl(
-        ctx, a, b, c, n_dims, mode, n_ctx_orig, freq_base, freq_scale,
+        ctx, a, b, c, n_dims, NULL, mode, n_ctx_orig, freq_base, freq_scale,
         ext_factor, attn_factor, beta_fast, beta_slow, true
     );
 }
@@ -4050,7 +4054,7 @@ struct ggml_tensor * ggml_rope_custom(
         float                 beta_fast,
         float                 beta_slow) {
     return ggml_rope_impl(
-        ctx, a, b, NULL, n_dims, mode, n_ctx_orig, freq_base, freq_scale,
+        ctx, a, b, NULL, n_dims, NULL, mode, n_ctx_orig, freq_base, freq_scale,
         ext_factor, attn_factor, beta_fast, beta_slow, false
     );
 }
@@ -4069,7 +4073,7 @@ struct ggml_tensor * ggml_rope_custom_inplace(
         float                 beta_fast,
         float                 beta_slow) {
     return ggml_rope_impl(
-        ctx, a, b, NULL, n_dims, mode, n_ctx_orig, freq_base, freq_scale,
+        ctx, a, b, NULL, n_dims, NULL, mode, n_ctx_orig, freq_base, freq_scale,
         ext_factor, attn_factor, beta_fast, beta_slow, true
     );
 }

From 843095fb53fdb22e8de41c8965d0c03e1462ce25 Mon Sep 17 00:00:00 2001
From: uvos <carl@uvos.xyz>
Date: Wed, 13 Aug 2025 20:44:30 +0200
Subject: [PATCH 62/86] HIP: bump requirement to rocm 6.1 (llama/15296)

---
 ggml/src/ggml-cuda/common.cuh    | 14 +++++---------
 ggml/src/ggml-cuda/ggml-cuda.cu  | 24 ------------------------
 ggml/src/ggml-cuda/vendors/hip.h | 16 ----------------
 ggml/src/ggml-hip/CMakeLists.txt |  4 ++--
 4 files changed, 7 insertions(+), 51 deletions(-)

diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
index 5a2a3478d26..2b14b30ac90 100644
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -464,25 +464,21 @@ static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b
 }
 
 static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const half2 b) {
-#if defined(GGML_USE_HIP) && HIP_VERSION >= 50700000
+#if defined(GGML_USE_HIP)
     return half2(__hmax(a.x, b.x), __hmax(a.y, b.y));
-#elif !defined(GGML_USE_HIP) && CUDART_VERSION >= CUDART_HMAX
+#elif CUDART_VERSION >= CUDART_HMAX
     return __hmax2(a, b);
-#elif !defined(GGML_USE_HIP)
+#else
     half2 ret;
     reinterpret_cast<half&>(ret.x) = __float2half(fmaxf( __low2float(a),  __low2float(b)));
     reinterpret_cast<half&>(ret.y) = __float2half(fmaxf(__high2float(a), __high2float(b)));
     return ret;
-#else
-    GGML_UNUSED(a);
-    GGML_UNUSED(b);
-    NO_DEVICE_CODE;
 #endif
 }
 
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
-#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || (defined(GGML_USE_HIP) && HIP_VERSION >= 50700000)
+#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || defined(GGML_USE_HIP)
 #pragma unroll
    for (int offset = width/2; offset > 0; offset >>= 1) {
        x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, offset, width));
@@ -491,7 +487,7 @@ static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
 #else
    GGML_UNUSED(x);
    NO_DEVICE_CODE;
-#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || (defined(GGML_USE_HIP) && HIP_VERSION >= 50700000)
+#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || defined(GGML_USE_HIP)
 }
 
 #if CUDART_VERSION < CUDART_HMASK
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index d9110491ec7..0d92901cb21 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -180,30 +180,6 @@ static int ggml_cuda_parse_id(char devName[]) {
 #endif // defined(GGML_USE_HIP)
 
 static ggml_cuda_device_info ggml_cuda_init() {
-#if defined(GGML_USE_HIP)
-    // Workaround for a rocBLAS bug when using multiple graphics cards:
-    // https://github.com/ROCmSoftwarePlatform/rocBLAS/issues/1346
-    {
-        int major_version = 0;
-        size_t version_length = 0;
-        if (rocblas_get_version_string_size(&version_length) == rocblas_status_success) {
-            std::vector<char> version(version_length+1, '\0');
-            if (rocblas_get_version_string(version.data(), version.size()) == rocblas_status_success) {
-                version.resize(::strlen(version.data()));
-                int parsed_value = 0;
-                if (std::from_chars(version.data(), version.data() + version.size(), parsed_value).ec == std::errc()) {
-                    major_version = parsed_value;
-                }
-            }
-        }
-        if (major_version < 4) {
-            GGML_LOG_DEBUG(GGML_CUDA_NAME " calling rocblas_initialize as a workaround for a rocBLAS bug\n");
-            rocblas_initialize();
-            CUDA_CHECK(cudaDeviceSynchronize());
-        }
-    }
-#endif
-
     ggml_cuda_device_info info = {};
 
     cudaError_t err = cudaGetDeviceCount(&info.device_count);
diff --git a/ggml/src/ggml-cuda/vendors/hip.h b/ggml/src/ggml-cuda/vendors/hip.h
index 96f8bc75e96..ec1b59caafc 100644
--- a/ggml/src/ggml-cuda/vendors/hip.h
+++ b/ggml/src/ggml-cuda/vendors/hip.h
@@ -5,8 +5,6 @@
 #include <hipblas/hipblas.h>
 #include <hip/hip_fp16.h>
 #include <hip/hip_bfloat16.h>
-// for rocblas_initialize()
-#include "rocblas/rocblas.h"
 
 #define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
 #define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
@@ -251,17 +249,3 @@ static __device__ __forceinline__ unsigned int __vcmpne4(unsigned int a, unsigne
     }
     return c;
 }
-
-#if HIP_VERSION < 50600000
-// __shfl_xor() for half2 was added in ROCm 5.6
-static __device__ __forceinline__ half2 __shfl_xor(half2 var, int laneMask, int width) {
-    typedef union half2_b32 {
-        half2 val;
-        int   b32;
-    } half2_b32_t;
-    half2_b32_t tmp;
-    tmp.val = var;
-    tmp.b32 = __shfl_xor(tmp.b32, laneMask, width);
-    return tmp.val;
-}
-#endif // HIP_VERSION < 50600000
diff --git a/ggml/src/ggml-hip/CMakeLists.txt b/ggml/src/ggml-hip/CMakeLists.txt
index 852de973460..d327b90cceb 100644
--- a/ggml/src/ggml-hip/CMakeLists.txt
+++ b/ggml/src/ggml-hip/CMakeLists.txt
@@ -46,8 +46,8 @@ if (GGML_HIP_ROCWMMA_FATTN)
     endif()
 endif()
 
-if (${hip_VERSION} VERSION_LESS 5.5)
-    message(FATAL_ERROR "At least ROCM/HIP V5.5 is required")
+if (${hip_VERSION} VERSION_LESS 6.1)
+    message(FATAL_ERROR "At least ROCM/HIP V6.1 is required")
 endif()
 
 message(STATUS "HIP and hipBLAS found")

From 164545a2cae49018f767fb69c72843bf9d8bb00a Mon Sep 17 00:00:00 2001
From: Jonathan Graehl <99024+graehl@users.noreply.github.com>
Date: Thu, 14 Aug 2025 03:03:57 -0700
Subject: [PATCH 63/86] finetune: SGD optimizer, more CLI args (llama/13873)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

* examples/finetune -opt SGD (stochastic gradient descent) memory opt

add unit tested GGML_OPT_OPTIMIZER_SGD to ggml - avoids allocating
m, v tensors.

support finetune.cpp arg -opt SGD (or sgd). (default adamw as before)

llama 3.2-1b-F32 result: observed 11gb gpu ram (41 sec/epoch)
when using SGD instead of 19gb (55 sec/epoch) using adamw.
(wikipedia 100 lines finetune)

(
using the same GPU memory, adamw can only do before OOM 512
batch/context, reaching:
train: [███████▉] data=0000140/0000140 loss=0.02575±0.00099 acc=99.52±0.03% t=00:00:47 ETA=00:00:00
val:   [███████▉] data=0000008/0000008 loss=4.76565±0.28810 acc=41.46±0.77% t=00:00:00 ETA=00:00:00

SGD is superior, though it converges slower, with max before OOM 1728
batch/context (esp see the better validation perf):
train: [███████▉] data=0000039/0000039 loss=0.00371±0.00010 acc=99.96±0.01% t=00:00:41 ETA=00:00:00
val:   [███████▉] data=0000003/0000003 loss=5.11406±0.76034 acc=48.01±0.69% t=00:00:01 ETA=00:00:00
)

note: when finetuning long enough (or w/ enough -lr),
validation accuracy *eventually* drops ('catastrophic forgetting')

-lr-half (halflife) option useful for SGD to avoid oscillation or
super slow underdamped learning (makes setting -lr more forgiving).
terminal -lr for now is set by lr-halvings i.e. if you want at most
1/8 the inital -lr you set -lr-halvings 3.

note: objective loss not directly comparable between adamw, sgd? -
check perplexity or accuracy or consider relative improvements
for convergence

new finetune args -wd 1e-9 to enable weight decay in sgd or adamw,
and max -epochs N (default 2 as before)

cache (1 - wd*alpha) in 'adamw' opt struct -
no noticeable perf benefit, disabled (still done
for new SGD though)

since opt. memory is pre-allocated, the ggml_opt_get_optimizer_params
would probably be able to change between SGD and AdamW with each epoch
but would need to use adamw for the first (unconfirmed - no cmdline arg
to set such a policy yet)

test-opt checks adamw as before and now sgd (except for a few disabled
tests for sgd only; probably just needs logging values and adding
alternate reference values);  tolerance on the 'regression'
test is broader for sgd (so we don't need many more epochs)

* Vulkan: Implement GGML_OP_OPT_STEP_SGD

* tests: Fix OPT_STEP_SGD test-backend-ops

* SGD op param store weight-decay and not 1-alpha*wd

* minor + cosmetic changes

* fix vulkan sgd

* try CI fix

---------

Co-authored-by: 0cc4m <picard12@live.de>
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
---
 ggml/include/ggml-opt.h                       |  31 +++-
 ggml/include/ggml.h                           |  10 +-
 ggml/src/ggml-cpu/ggml-cpu.c                  |   6 +
 ggml/src/ggml-cpu/ops.cpp                     |  65 ++++++++-
 ggml/src/ggml-cpu/ops.h                       |   2 +-
 ggml/src/ggml-cuda/ggml-cuda.cu               |   5 +
 ggml/src/ggml-cuda/opt-step-sgd.cu            |  49 +++++++
 ggml/src/ggml-cuda/opt-step-sgd.cuh           |   5 +
 ggml/src/ggml-opt.cpp                         | 138 ++++++++++++------
 ggml/src/ggml-vulkan/ggml-vulkan.cpp          |  36 ++++-
 .../vulkan-shaders/opt_step_sgd.comp          |  22 +++
 .../vulkan-shaders/vulkan-shaders-gen.cpp     |   1 +
 ggml/src/ggml.c                               |  29 +++-
 13 files changed, 342 insertions(+), 57 deletions(-)
 create mode 100644 ggml/src/ggml-cuda/opt-step-sgd.cu
 create mode 100644 ggml/src/ggml-cuda/opt-step-sgd.cuh
 create mode 100644 ggml/src/ggml-vulkan/vulkan-shaders/opt_step_sgd.comp

diff --git a/ggml/include/ggml-opt.h b/ggml/include/ggml-opt.h
index 74ec080a055..4703a05afe1 100644
--- a/ggml/include/ggml-opt.h
+++ b/ggml/include/ggml-opt.h
@@ -74,16 +74,26 @@ extern "C" {
         GGML_OPT_BUILD_TYPE_OPT     = 30,
     };
 
+    enum ggml_opt_optimizer_type {
+        GGML_OPT_OPTIMIZER_TYPE_ADAMW,
+        GGML_OPT_OPTIMIZER_TYPE_SGD,
+
+        GGML_OPT_OPTIMIZER_TYPE_COUNT
+    };
+
     // parameters that control which optimizer is used and how said optimizer tries to find the minimal loss
     struct ggml_opt_optimizer_params {
-        // AdamW optimizer parameters
         struct {
             float alpha; // learning rate
-            float beta1;
-            float beta2;
+            float beta1; // first AdamW momentum
+            float beta2; // second AdamW momentum
             float eps;   // epsilon for numerical stability
-            float wd;    // weight decay for AdamW, use 0.0f to disable
+            float wd;    // weight decay - 0.0f to disable
         } adamw;
+        struct {
+            float alpha; // learning rate
+            float wd;    // weight decay
+        } sgd;
     };
 
     // callback to calculate optimizer parameters prior to a backward pass
@@ -112,8 +122,11 @@ extern "C" {
 
         int32_t opt_period; // after how many gradient accumulation steps an optimizer step should be done
 
-        ggml_opt_get_optimizer_params get_opt_pars; // callback for calculating optimizer parameters
-        void * get_opt_pars_ud;                     // userdata for calculating optimizer parameters
+        ggml_opt_get_optimizer_params get_opt_pars;    // callback for calculating optimizer parameters
+        void *                        get_opt_pars_ud; // userdata for calculating optimizer parameters
+
+        // only GGML_OPT_OPTIMIZER_TYPE_ADAMW needs m, v momenta per parameter tensor
+        enum ggml_opt_optimizer_type optimizer;
     };
 
     // get parameters for an optimization context with defaults set where possible
@@ -142,6 +155,10 @@ extern "C" {
     // get the gradient accumulator for a node from the forward graph
     GGML_API struct ggml_tensor * ggml_opt_grad_acc(ggml_opt_context_t opt_ctx, struct ggml_tensor * node);
 
+    GGML_API enum ggml_opt_optimizer_type ggml_opt_context_optimizer_type(ggml_opt_context_t); //TODO consistent naming scheme
+
+    GGML_API const char * ggml_opt_optimizer_name(enum ggml_opt_optimizer_type);
+
     // ====== Optimization Result ======
 
     GGML_API ggml_opt_result_t ggml_opt_result_init(void);
@@ -226,12 +243,14 @@ extern "C" {
             struct ggml_tensor            * outputs,        // output tensor, must have shape [ne_label, ndata_batch] if labels are used
             ggml_opt_dataset_t              dataset,        // dataset with data and optionally also labels
             enum ggml_opt_loss_type         loss_type,      // loss to minimize
+            enum ggml_opt_optimizer_type    optimizer,      // sgd or adamw
             ggml_opt_get_optimizer_params   get_opt_pars,   // callback to get optimizer params, userdata is pointer to epoch (of type int64_t)
             int64_t                         nepoch,         // how many times the dataset should be iterated over
             int64_t                         nbatch_logical, // datapoints optimizer step, must be a multiple of ndata_batch in inputs/outputs
             float                           val_split,      // fraction of the dataset to use for validation, must be in [0.0f, 1.0f)
             bool                            silent);        // whether or not info prints to stderr should be suppressed
 
+
 #ifdef  __cplusplus
 }
 #endif
diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
index c09d82a91cf..da8813fd278 100644
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -542,6 +542,7 @@ extern "C" {
         GGML_OP_CROSS_ENTROPY_LOSS,
         GGML_OP_CROSS_ENTROPY_LOSS_BACK,
         GGML_OP_OPT_STEP_ADAMW,
+        GGML_OP_OPT_STEP_SGD,
 
         GGML_OP_GLU,
 
@@ -2311,7 +2312,14 @@ extern "C" {
             struct ggml_tensor  * grad,
             struct ggml_tensor  * m,
             struct ggml_tensor  * v,
-            struct ggml_tensor  * adamw_params); // parameters such a the learning rate
+            struct ggml_tensor  * adamw_params); // parameters such as the learning rate
+
+    // stochastic gradient descent step (with weight decay)
+    GGML_API struct ggml_tensor * ggml_opt_step_sgd(
+        struct ggml_context * ctx,
+        struct ggml_tensor *  a,
+        struct ggml_tensor *  grad,
+        struct ggml_tensor *  sgd_params); // alpha, weight decay
 
     //
     // automatic differentiation
diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
index d89cd8f4ef6..f6bea3df34a 100644
--- a/ggml/src/ggml-cpu/ggml-cpu.c
+++ b/ggml/src/ggml-cpu/ggml-cpu.c
@@ -2022,6 +2022,11 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
                 ggml_compute_forward_opt_step_adamw(params, tensor);
             }
             break;
+        case GGML_OP_OPT_STEP_SGD:
+            {
+                ggml_compute_forward_opt_step_sgd(params, tensor);
+            }
+            break;
         case GGML_OP_NONE:
             {
                 // nop
@@ -2325,6 +2330,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_CROSS_ENTROPY_LOSS:
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
         case GGML_OP_OPT_STEP_ADAMW:
+        case GGML_OP_OPT_STEP_SGD:
             {
                 n_tasks = n_threads;
             } break;
diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
index 854f1c2b496..b72a2556a5f 100644
--- a/ggml/src/ggml-cpu/ops.cpp
+++ b/ggml/src/ggml-cpu/ops.cpp
@@ -10330,6 +10330,7 @@ static void ggml_compute_forward_opt_step_adamw_f32(
     const int ir1 = MIN(ir0 + dr, nr);
 
     const float * adamw_params_ptr = ggml_get_data_f32(adamw_params);
+
     const float alpha  = adamw_params_ptr[0];
     const float beta1  = adamw_params_ptr[1];
     const float beta2  = adamw_params_ptr[2];
@@ -10337,7 +10338,7 @@ static void ggml_compute_forward_opt_step_adamw_f32(
     const float wd     = adamw_params_ptr[4];
     const float beta1h = adamw_params_ptr[5];
     const float beta2h = adamw_params_ptr[6];
-
+    const float keep   = 1.f - alpha * wd;
     for (int ir = ir0; ir < ir1; ++ir) {
         const int64_t i03 = ir/(ne02*ne01);
         const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
@@ -10360,7 +10361,7 @@ static void ggml_compute_forward_opt_step_adamw_f32(
             // The weight decay is applied independently of the Adam momenta m and v.
             // This is NOT equivalent to l2 regularization that adds w[i00]*w[i00] to the loss.
             // See: https://arxiv.org/pdf/1711.05101v3.pdf
-            w[i00] = w[i00]*(1.0f - alpha*wd) - alpha*mh/vh;
+            w[i00] = w[i00] * keep - alpha * mh / vh;
         }
     }
 }
@@ -10382,3 +10383,63 @@ void ggml_compute_forward_opt_step_adamw(
             }
     }
 }
+
+static void ggml_compute_forward_opt_step_sgd_f32(const ggml_compute_params * params, ggml_tensor * dst) {
+    const ggml_tensor * src0       = dst->src[0];
+    const ggml_tensor * src0_grad  = dst->src[1];
+    const ggml_tensor * sgd_params = dst->src[2];
+
+    GGML_ASSERT(ggml_are_same_shape(src0, src0_grad));
+    GGML_ASSERT(ggml_nelements(sgd_params) == 2);
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nr = ggml_nrows(src0);
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+    GGML_ASSERT(nb00 == sizeof(float));
+
+    // rows per thread
+    const int dr = (nr + nth - 1) / nth;
+
+    // row range for this thread
+    const int ir0 = dr * ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    // using adamw param subset we care about - alpha, wd - could have a separate struct
+    const float * sgd_params_ptr   = ggml_get_data_f32(sgd_params);
+    const float   alpha            = sgd_params_ptr[0];
+    const float   keep             = 1.f - alpha * sgd_params_ptr[1];
+
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int64_t i03 = ir / (ne02 * ne01);
+        const int64_t i02 = (ir - i03 * ne02 * ne01) / ne01;
+        const int64_t i01 = (ir - i03 * ne02 * ne01 - i02 * ne01);
+
+        const size_t offset = i03 * nb03 + i02 * nb02 + i01 * nb01;
+
+        float *       w = (float *) ((char *) src0->data + offset);                   // weight
+        const float * g = (const float *) ((const char *) src0_grad->data + offset);  // grad
+
+        for (int i00 = 0; i00 < ne00; ++i00) {
+            w[i00] = w[i00] * keep - alpha * g[i00];
+        }
+    }
+}
+
+void ggml_compute_forward_opt_step_sgd(const ggml_compute_params * params, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_opt_step_sgd_f32(params, dst);
+            }
+            break;
+        default:
+            {
+                GGML_ABORT("fatal error - sgd is F32 only");
+            }
+    }
+}
diff --git a/ggml/src/ggml-cpu/ops.h b/ggml/src/ggml-cpu/ops.h
index f154afb4624..82ea79eaa51 100644
--- a/ggml/src/ggml-cpu/ops.h
+++ b/ggml/src/ggml-cpu/ops.h
@@ -107,7 +107,7 @@ void ggml_compute_forward_cross_entropy_loss(const struct ggml_compute_params *
 void ggml_compute_forward_cross_entropy_loss_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_opt_step_adamw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-
+void ggml_compute_forward_opt_step_sgd(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 #ifdef __cplusplus
 }
 #endif
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index 0d92901cb21..d6402a8daac 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -28,6 +28,7 @@
 #include "ggml-cuda/mmvq.cuh"
 #include "ggml-cuda/norm.cuh"
 #include "ggml-cuda/opt-step-adamw.cuh"
+#include "ggml-cuda/opt-step-sgd.cuh"
 #include "ggml-cuda/out-prod.cuh"
 #include "ggml-cuda/pad.cuh"
 #include "ggml-cuda/pool2d.cuh"
@@ -2479,6 +2480,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_OPT_STEP_ADAMW:
             ggml_cuda_opt_step_adamw(ctx, dst);
             break;
+        case GGML_OP_OPT_STEP_SGD:
+            ggml_cuda_opt_step_sgd(ctx, dst);
+            break;
         default:
             return false;
     }
@@ -3536,6 +3540,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_CROSS_ENTROPY_LOSS:
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
         case GGML_OP_OPT_STEP_ADAMW:
+        case GGML_OP_OPT_STEP_SGD:
             return true;
         default:
             return false;
diff --git a/ggml/src/ggml-cuda/opt-step-sgd.cu b/ggml/src/ggml-cuda/opt-step-sgd.cu
new file mode 100644
index 00000000000..460b16de447
--- /dev/null
+++ b/ggml/src/ggml-cuda/opt-step-sgd.cu
@@ -0,0 +1,49 @@
+#include "ggml-impl.h"
+#include "opt-step-sgd.cuh"
+
+#include <cstdint>
+
+static __global__ void opt_step_sgd_f32(
+    float * __restrict__ x, const float * __restrict__ g,
+    const float * __restrict__ pars, const int64_t k) {
+
+    const int64_t i = (int64_t) blockIdx.x*blockDim.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    x[i] = x[i] * (1.0f - pars[0] * pars[1]) - pars[0] * g[i];
+}
+
+static void opt_step_sgd_f32_cuda(
+    float * x, const float * g, const float * __restrict__ pars, const int64_t k, cudaStream_t stream) {
+
+    const dim3 block_dims(CUDA_OPT_STEP_SGD_BLOCK_SIZE, 1, 1);
+    const dim3 block_nums((k + CUDA_OPT_STEP_SGD_BLOCK_SIZE - 1) / CUDA_OPT_STEP_SGD_BLOCK_SIZE, 1, 1);
+    opt_step_sgd_f32<<<block_nums, block_dims, 0, stream>>>(x, g, pars, k);
+}
+
+void ggml_cuda_opt_step_sgd(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0      = dst->src[0];
+    const ggml_tensor * src0_grad = dst->src[1];
+    const ggml_tensor * params    = dst->src[2];
+
+    GGML_ASSERT(src0->type      == GGML_TYPE_F32);
+    GGML_ASSERT(src0_grad->type == GGML_TYPE_F32);
+    GGML_ASSERT(params->type    == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src0_grad));
+    GGML_ASSERT(ggml_is_contiguous(params));
+    GGML_ASSERT(ggml_are_same_shape(src0, src0_grad));
+    GGML_ASSERT(ggml_nelements(params) == 2);
+
+    float       * src0_d      = (float       *) src0->data;
+    const float * src0_grad_d = (const float *) src0_grad->data;
+    const float * params_d    = (const float *) params->data;
+
+    cudaStream_t stream = ctx.stream();
+
+    const int64_t ne = ggml_nelements(src0);
+
+    opt_step_sgd_f32_cuda(src0_d, src0_grad_d, params_d, ne, stream);
+}
diff --git a/ggml/src/ggml-cuda/opt-step-sgd.cuh b/ggml/src/ggml-cuda/opt-step-sgd.cuh
new file mode 100644
index 00000000000..f97ab7d9bed
--- /dev/null
+++ b/ggml/src/ggml-cuda/opt-step-sgd.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_OPT_STEP_SGD_BLOCK_SIZE 256
+
+void ggml_cuda_opt_step_sgd(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/ggml/src/ggml-opt.cpp b/ggml/src/ggml-opt.cpp
index a3c82d67577..e078ad14a39 100644
--- a/ggml/src/ggml-opt.cpp
+++ b/ggml/src/ggml-opt.cpp
@@ -64,9 +64,11 @@ struct ggml_opt_context {
     int32_t opt_i              = 0;
     bool    loss_per_datapoint = false;
 
-    ggml_opt_get_optimizer_params get_opt_pars = nullptr;
-    void * get_opt_pars_ud                     = nullptr;
-    struct ggml_tensor * adamw_params          = nullptr;
+    ggml_opt_get_optimizer_params get_opt_pars    = nullptr;
+    void *                        get_opt_pars_ud = nullptr;
+    struct ggml_tensor *          opt_step_params = nullptr; // Stores output of get_opt_pars.
+
+    enum ggml_opt_optimizer_type optimizer = GGML_OPT_OPTIMIZER_TYPE_ADAMW;
 };
 
 struct ggml_opt_result {
@@ -229,9 +231,13 @@ struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * us
     result.adamw.eps   = 1e-8f;
     result.adamw.wd    = 0.0f;
 
+    result.sgd.alpha   = 1e-3f;
+    result.sgd.wd      = 0.0f;
+
     return result;
 }
 
+
 struct ggml_opt_optimizer_params ggml_opt_get_constant_optimizer_params(void * userdata) {
     return *((struct ggml_opt_optimizer_params *) userdata);
 }
@@ -249,6 +255,7 @@ struct ggml_opt_params ggml_opt_default_params(
         /*opt_period      =*/ 1,
         /*get_opt_pars    =*/ ggml_opt_get_default_optimizer_params,
         /*get_opt_pars_ud =*/ nullptr,
+        /*optimizer       =*/ GGML_OPT_OPTIMIZER_TYPE_ADAMW,
     };
 }
 
@@ -316,9 +323,14 @@ static void ggml_opt_build(ggml_opt_context_t opt_ctx) {
     GGML_ASSERT(opt_ctx->ctx_compute && "no compute context set, either use static graphs or set one with ggml_opt_prepare_alloc");
     GGML_ASSERT((!opt_ctx->static_graphs || opt_ctx->inputs->data) && "when using static graphs the inputs must be allocated statically");
 
+    const enum ggml_opt_optimizer_type optimizer = opt_ctx->optimizer;
+
     const bool accumulate = opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_GRAD &&
         !(opt_ctx->static_graphs && opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT && opt_ctx->opt_period == 1);
 
+    const bool need_momenta = opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT &&
+        opt_ctx->optimizer == GGML_OPT_OPTIMIZER_TYPE_ADAMW;
+
     ggml_set_input(opt_ctx->inputs);
     ggml_set_output(opt_ctx->outputs);
 
@@ -340,8 +352,7 @@ static void ggml_opt_build(ggml_opt_context_t opt_ctx) {
         //   - pred (if using static graphs)
         //   - ncorrect (if using static graphs, 2 tensors).
         constexpr size_t n_loss = 1;
-        const size_t tensors_per_param = (accumulate ? 1 : 0) +
-            (opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT ? 2 : 0);
+        const size_t tensors_per_param = (accumulate ? 1 : 0) + (need_momenta ? 2 : 0);
         const size_t tensors_const = opt_ctx->static_graphs ? 9 : 0;
         const size_t size_meta = (n_loss + tensors_per_param*n_param + tensors_const) * ggml_tensor_overhead();
         struct ggml_init_params params = {
@@ -458,7 +469,7 @@ static void ggml_opt_build(ggml_opt_context_t opt_ctx) {
             }
         }
 
-        if (opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_OPT) {
+        if (need_momenta && opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_OPT) {
             opt_ctx->grad_m.resize(n_nodes);
             opt_ctx->grad_v.resize(n_nodes);
             for (int i = 0; i < n_nodes; ++i) {
@@ -492,23 +503,36 @@ static void ggml_opt_build(ggml_opt_context_t opt_ctx) {
     // gb_opt == graph backward optimize, forward pass, then backward pass to calculate gradients, then optimizer step.
     opt_ctx->gb_opt = ggml_graph_dup(opt_ctx->ctx_compute, opt_ctx->gb_grad, /*force_grads =*/ true);
 
-    opt_ctx->adamw_params = ggml_new_tensor_1d(opt_ctx->ctx_cpu, GGML_TYPE_F32, 7);
-    ggml_set_input(opt_ctx->adamw_params);
-    ggml_set_name(opt_ctx->adamw_params, "adamw_params");
-
+    opt_ctx->opt_step_params = ggml_new_tensor_1d(opt_ctx->ctx_cpu, GGML_TYPE_F32, need_momenta ? 7 : 2);
+    ggml_tensor * adamw_params = opt_ctx->opt_step_params;
+    ggml_set_input(adamw_params);
+    const char * optimizer_name = ggml_opt_optimizer_name(opt_ctx->optimizer);
+    ggml_format_name(adamw_params, "%s_params", optimizer_name);
     for (int i = opt_ctx->gf->n_nodes-1; i >= 0; --i) {
         struct ggml_tensor * node = opt_ctx->gb_opt->nodes[i];
         struct ggml_tensor * grad = ggml_graph_get_grad(opt_ctx->gb_opt, node);
 
         if (grad && (node->flags & GGML_TENSOR_FLAG_PARAM)) {
-            struct ggml_tensor * m        = opt_ctx->grad_m[i];
-            struct ggml_tensor * v        = opt_ctx->grad_v[i];
-            struct ggml_tensor * opt_step = ggml_opt_step_adamw(opt_ctx->ctx_compute, node, grad, m, v, opt_ctx->adamw_params);
-
-            ggml_set_name(m,        (std::string("AdamW m for ")    + std::string(node->name)).c_str());
-            ggml_set_name(v,        (std::string("AdamW v for ")    + std::string(node->name)).c_str());
-            ggml_set_name(opt_step, (std::string("AdamW step for ") + std::string(node->name)).c_str());
-
+            struct ggml_tensor * m = nullptr;
+            struct ggml_tensor * v = nullptr;
+            if (need_momenta) {
+                m = opt_ctx->grad_m[i];
+                v = opt_ctx->grad_v[i];
+                ggml_format_name(m, "AdamW m for %s", node->name);
+                ggml_format_name(v, "AdamW v for %s", node->name);
+            }
+            struct ggml_tensor * opt_step;
+            switch (optimizer) {
+                case GGML_OPT_OPTIMIZER_TYPE_ADAMW:
+                    opt_step = ggml_opt_step_adamw(opt_ctx->ctx_compute, node, grad, m, v, adamw_params);
+                    break;
+                case GGML_OPT_OPTIMIZER_TYPE_SGD:
+                    opt_step = ggml_opt_step_sgd(opt_ctx->ctx_compute, node, grad, adamw_params);
+                    break;
+                default:
+                    GGML_ABORT("fatal error");
+            }
+            ggml_format_name(opt_step, "%s step for %s", optimizer_name, node->name);
             ggml_build_forward_expand(opt_ctx->gb_opt, opt_step);
         }
     }
@@ -534,6 +558,7 @@ ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
     result->opt_period       = params.opt_period;
     result->get_opt_pars     = params.get_opt_pars;
     result->get_opt_pars_ud  = params.get_opt_pars_ud;
+    result->optimizer        = params.optimizer;
 
     GGML_ASSERT(result->opt_period >= 1);
 
@@ -756,29 +781,43 @@ void ggml_opt_alloc(ggml_opt_context_t opt_ctx, bool backward) {
 void ggml_opt_eval(ggml_opt_context_t opt_ctx, ggml_opt_result_t result) {
     GGML_ASSERT(opt_ctx->eval_ready);
     if (opt_ctx->allocated_graph == opt_ctx->gb_opt) {
-        struct ggml_opt_optimizer_params opt_pars = opt_ctx->get_opt_pars(opt_ctx->get_opt_pars_ud);
-
-        GGML_ASSERT(opt_pars.adamw.alpha >  0.0f);
-        GGML_ASSERT(opt_pars.adamw.beta1 >= 0.0f);
-        GGML_ASSERT(opt_pars.adamw.beta1 <= 1.0f);
-        GGML_ASSERT(opt_pars.adamw.beta2 >= 0.0f);
-        GGML_ASSERT(opt_pars.adamw.beta2 <= 1.0f);
-        GGML_ASSERT(opt_pars.adamw.eps   >= 0.0f);
-        GGML_ASSERT(opt_pars.adamw.wd    >= 0.0f);
-        GGML_ASSERT(opt_pars.adamw.wd    <= 1.0f);
-
-        // beta1, beta2 after applying warmup
-        const float beta1h = 1.0f/(1.0f - powf(opt_pars.adamw.beta1, opt_ctx->iter));
-        const float beta2h = 1.0f/(1.0f - powf(opt_pars.adamw.beta2, opt_ctx->iter));
-
-        float * adamw_par_data = ggml_get_data_f32(opt_ctx->adamw_params);
-        adamw_par_data[0] = opt_pars.adamw.alpha;
-        adamw_par_data[1] = opt_pars.adamw.beta1;
-        adamw_par_data[2] = opt_pars.adamw.beta2;
-        adamw_par_data[3] = opt_pars.adamw.eps;
-        adamw_par_data[4] = opt_pars.adamw.wd;
-        adamw_par_data[5] = beta1h;
-        adamw_par_data[6] = beta2h;
+        const ggml_opt_optimizer_params & opt_pars = opt_ctx->get_opt_pars(opt_ctx->get_opt_pars_ud);
+
+        switch (opt_ctx->optimizer) {
+            case GGML_OPT_OPTIMIZER_TYPE_ADAMW: {
+                GGML_ASSERT(opt_pars.adamw.alpha > 0.0f);
+                GGML_ASSERT(opt_pars.adamw.beta1 >= 0.0f);
+                GGML_ASSERT(opt_pars.adamw.beta1 <= 1.0f);
+                GGML_ASSERT(opt_pars.adamw.beta2 >= 0.0f);
+                GGML_ASSERT(opt_pars.adamw.beta2 <= 1.0f);
+                GGML_ASSERT(opt_pars.adamw.eps >= 0.0f);
+                GGML_ASSERT(opt_pars.adamw.wd >= 0.0f);
+                GGML_ASSERT(opt_pars.adamw.wd <= 1.0f);
+
+                // beta1, beta2 after applying warmup
+                const float beta1h = 1.0f / (1.0f - powf(opt_pars.adamw.beta1, opt_ctx->iter));
+                const float beta2h = 1.0f / (1.0f - powf(opt_pars.adamw.beta2, opt_ctx->iter));
+
+                float * adamw_par_data = ggml_get_data_f32(opt_ctx->opt_step_params);
+                adamw_par_data[0] = opt_pars.adamw.alpha;
+                adamw_par_data[1] = opt_pars.adamw.beta1;
+                adamw_par_data[2] = opt_pars.adamw.beta2;
+                adamw_par_data[3] = opt_pars.adamw.eps;
+                adamw_par_data[4] = opt_pars.adamw.wd;
+                adamw_par_data[5] = beta1h;
+                adamw_par_data[6] = beta2h;
+            } break;
+            case GGML_OPT_OPTIMIZER_TYPE_SGD: {
+                GGML_ASSERT(opt_pars.sgd.alpha > 0.0f);
+                GGML_ASSERT(opt_pars.sgd.wd >= 0.0f);
+                GGML_ASSERT(opt_pars.sgd.wd <= 1.0f);
+                float * sgd = ggml_get_data_f32(opt_ctx->opt_step_params);
+                sgd[0] = opt_pars.sgd.alpha;
+                sgd[1] = opt_pars.sgd.wd;
+            } break;
+            default:
+                GGML_ABORT("fatal error");
+        }
     }
 
     ggml_backend_sched_graph_compute(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy);
@@ -963,6 +1002,7 @@ void ggml_opt_fit(
         ggml_tensor                   * outputs,
         ggml_opt_dataset_t              dataset,
         enum ggml_opt_loss_type         loss_type,
+        enum ggml_opt_optimizer_type    optimizer,
         ggml_opt_get_optimizer_params   get_opt_pars,
         int64_t                         nepoch,
         int64_t                         nbatch_logical,
@@ -993,6 +1033,7 @@ void ggml_opt_fit(
     params.opt_period      = opt_period;
     params.get_opt_pars    = get_opt_pars;
     params.get_opt_pars_ud = &epoch;
+    params.optimizer       = optimizer;
     ggml_opt_context_t opt_ctx = ggml_opt_init(params);
 
     // Shuffling the data is generally useful but there is only a point if not all data is used in a single batch.
@@ -1035,3 +1076,18 @@ void ggml_opt_fit(
     ggml_opt_result_free(result_train);
     ggml_opt_result_free(result_val);
 }
+
+enum ggml_opt_optimizer_type ggml_opt_context_optimizer_type(ggml_opt_context_t c) {
+    return c->optimizer;
+}
+
+GGML_API const char * ggml_opt_optimizer_name(enum ggml_opt_optimizer_type o) {
+    switch (o) {
+        case GGML_OPT_OPTIMIZER_TYPE_ADAMW:
+            return "adamw";
+        case GGML_OPT_OPTIMIZER_TYPE_SGD:
+            return "sgd";
+        default:
+            return "undefined";
+    };
+}
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 4070e248baa..f50a737f389 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -510,6 +510,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_rwkv_wkv6_f32;
     vk_pipeline pipeline_rwkv_wkv7_f32;
     vk_pipeline pipeline_opt_step_adamw_f32;
+    vk_pipeline pipeline_opt_step_sgd_f32;
     vk_pipeline pipeline_conv2d_f32[CONV_SHAPE_COUNT];
     vk_pipeline pipeline_conv2d_f16_f32[CONV_SHAPE_COUNT];
     vk_pipeline pipeline_conv2d_dw_whcn_f32;
@@ -3123,6 +3124,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_opt_step_adamw_f32, "opt_step_adamw_f32", opt_step_adamw_f32_len, opt_step_adamw_f32_data, "main", 5, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
 
+    ggml_vk_create_pipeline(device, device->pipeline_opt_step_sgd_f32, "opt_step_sgd_f32", opt_step_sgd_f32_len, opt_step_sgd_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
+
     // conv2d
     for (uint32_t s = 0; s < CONV_SHAPE_COUNT; ++s) {
         uint32_t conv2d_WG_SIZE  = 256;
@@ -7193,6 +7196,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_opt_step_adamw_f32;
         }
         return nullptr;
+    case GGML_OP_OPT_STEP_SGD:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_opt_step_sgd_f32;
+        }
+        return nullptr;
     case GGML_OP_LEAKY_RELU:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_leaky_relu_f32;
@@ -7692,6 +7700,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         ggml_vk_buffer_memset_async(subctx, d_D, d_buf_offset, 0, d_sz);
         ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
+    } else if (op == GGML_OP_OPT_STEP_SGD) {
+        // OPT_STEP_SGD works on src0, it does not need dst
+        ggml_vk_sync_buffers(subctx);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz } }, pc, elements);
     } else if (use_src2) {
         ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
@@ -8045,6 +8057,12 @@ static void ggml_vk_opt_step_adamw(ggml_backend_vk_context * ctx, vk_context& su
     );
 }
 
+static void ggml_vk_opt_step_sgd(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool dryrun = false) {
+    const size_t n = ggml_nelements(dst->src[0]);
+
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, src2, dst, GGML_OP_OPT_STEP_SGD, { (uint32_t)n, 0, 0.0f, 0.0f }, dryrun);
+}
+
 static void ggml_vk_concat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
     int * op_params = (int *)dst->op_params;
 
@@ -9598,6 +9616,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_LEAKY_RELU:
     case GGML_OP_FLASH_ATTN_EXT:
     case GGML_OP_OPT_STEP_ADAMW:
+    case GGML_OP_OPT_STEP_SGD:
         break;
     default:
         std::cerr << "ggml_vulkan: Error: Missing op: " << ggml_op_name(node->op) << std::endl;
@@ -9662,6 +9681,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         case GGML_OP_CONV_2D:
         case GGML_OP_CONV_2D_DW:
         case GGML_OP_LEAKY_RELU:
+        case GGML_OP_OPT_STEP_SGD:
             {
                 // These operations all go through ggml_vk_op_f32, so short-circuit and
                 // do the only thing needed for the dryrun.
@@ -9911,6 +9931,11 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_OPT_STEP_ADAMW:
         ggml_vk_opt_step_adamw(ctx, compute_ctx, node, dryrun);
 
+        break;
+
+    case GGML_OP_OPT_STEP_SGD:
+        ggml_vk_opt_step_sgd(ctx, compute_ctx, src0, src1, src2, node, dryrun);
+
         break;
     default:
         return false;
@@ -10014,8 +10039,8 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
     case GGML_OP_REPEAT:
     case GGML_OP_REPEAT_BACK:
     case GGML_OP_OPT_STEP_ADAMW:
+    case GGML_OP_OPT_STEP_SGD:
         buf = tensor->buffer;
-
         break;
     case GGML_OP_UNARY:
         switch (ggml_get_unary_op(tensor)) {
@@ -11154,6 +11179,9 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_SIN:
         case GGML_OP_COS:
         case GGML_OP_CLAMP:
+        case GGML_OP_LEAKY_RELU:
+        case GGML_OP_OPT_STEP_ADAMW:
+        case GGML_OP_OPT_STEP_SGD:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_UPSCALE:
         case GGML_OP_ACC:
@@ -11175,8 +11203,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_POOL_2D:
         case GGML_OP_RWKV_WKV6:
         case GGML_OP_RWKV_WKV7:
-        case GGML_OP_LEAKY_RELU:
-        case GGML_OP_OPT_STEP_ADAMW:
             return true;
         case GGML_OP_CONV_TRANSPOSE_1D:
             return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
@@ -11774,6 +11800,10 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         src_clone[0]->flags = src0->flags;
         tensor_clone = ggml_opt_step_adamw(ggml_ctx, src_clone[0], src_clone[1],
         src_clone[2], src_clone[3], src_clone[4]);
+    } else if (tensor->op == GGML_OP_OPT_STEP_SGD) {
+        src_clone[0]->flags = src0->flags;
+        tensor_clone = ggml_opt_step_sgd(ggml_ctx, src_clone[0], src_clone[1],
+        src_clone[2]);
     }
     else {
         std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_sgd.comp b/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_sgd.comp
new file mode 100644
index 00000000000..6426dedee57
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_sgd.comp
@@ -0,0 +1,22 @@
+#version 450
+
+#include "generic_head.comp"
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) buffer X {A_TYPE data_x[];};
+layout (binding = 1) readonly buffer G {A_TYPE data_grad[];};
+layout (binding = 2) readonly buffer P {float data_params[2];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float alpha = data_params[0];
+    const float keep = 1.f - alpha * data_params[1];
+
+    data_x[i] = data_x[i] * keep - alpha * data_grad[i];
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
index 4cd94c51e3f..68933d19f2e 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -657,6 +657,7 @@ void process_shaders() {
     string_to_spv("rwkv_wkv7_f32", "wkv7.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
 
     string_to_spv("opt_step_adamw_f32", "opt_step_adamw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
+    string_to_spv("opt_step_sgd_f32", "opt_step_sgd.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
 
     string_to_spv("conv2d_f32_unroll", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}});
     string_to_spv("conv2d_f16_f32_unroll", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}});
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index aff0b47cb26..54961213115 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -1012,11 +1012,12 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "CROSS_ENTROPY_LOSS",
     "CROSS_ENTROPY_LOSS_BACK",
     "OPT_STEP_ADAMW",
+    "OPT_STEP_SGD",
 
     "GLU",
 };
 
-static_assert(GGML_OP_COUNT == 87, "GGML_OP_COUNT != 87");
+static_assert(GGML_OP_COUNT == 88, "GGML_OP_COUNT != 88");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -1113,15 +1114,15 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "cross_entropy_loss(x,y)",
     "cross_entropy_loss_back(x,y)",
     "adamw(x)",
+    "sgd(x)",
 
     "glu(x)",
 };
 
-static_assert(GGML_OP_COUNT == 87, "GGML_OP_COUNT != 87");
+static_assert(GGML_OP_COUNT == 88, "GGML_OP_COUNT != 88");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
-
 static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = {
     "ABS",
     "SGN",
@@ -5606,6 +5607,28 @@ struct ggml_tensor * ggml_opt_step_adamw(
     return result;
 }
 
+// opt_step_sgd
+
+struct ggml_tensor * ggml_opt_step_sgd(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * grad,
+        struct ggml_tensor  * params) {
+    GGML_ASSERT(a->flags & GGML_TENSOR_FLAG_PARAM);
+    GGML_ASSERT(ggml_are_same_shape(a, grad));
+    GGML_ASSERT(params->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_nelements(params) == 2);
+
+    struct ggml_tensor * result = ggml_view_tensor(ctx, a);
+
+    result->op     = GGML_OP_OPT_STEP_SGD;
+    result->src[0] = a;
+    result->src[1] = grad;
+    result->src[2] = params;
+
+    return result;
+}
+
 ////////////////////////////////////////////////////////////////////////////////
 
 struct ggml_hash_set ggml_hash_set_new(size_t size) {

From da2d2947ce5d812c2c6da5bfbfde7bd50846a9c1 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= <sigbjorn.skjaeret@scala.com>
Date: Thu, 14 Aug 2025 12:22:07 +0200
Subject: [PATCH 64/86] cuda : fix GGML_CUDA_GRAPHS=OFF (llama/15300)

* fix USE_CUDA_GRAPH=OFF

ggml-ci

* check capture status

* completely disable capturing check instead
---
 ggml/src/ggml-cuda/mean.cu | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/ggml/src/ggml-cuda/mean.cu b/ggml/src/ggml-cuda/mean.cu
index 2ad493239b1..347abc18660 100644
--- a/ggml/src/ggml-cuda/mean.cu
+++ b/ggml/src/ggml-cuda/mean.cu
@@ -25,9 +25,12 @@ void ggml_cuda_op_mean(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
 // Special case for reducing vectors
 #ifdef GGML_CUDA_USE_CUB
+#ifdef USE_CUDA_GRAPH
     cudaStreamCaptureStatus iscapturing;
     CUDA_CHECK(cudaStreamIsCapturing(stream, &iscapturing));
+#endif // USE_CUDA_GRAPH
     if ((nrows == 1) &&
+#ifdef USE_CUDA_GRAPH
             // CUDA_GRAPHS_DISABLED
             ((ncols > 65536) &&
              ((ctx.cuda_graph->instance == nullptr) && (iscapturing == cudaStreamCaptureStatusNone) ||
@@ -38,6 +41,9 @@ void ggml_cuda_op_mean(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
          !((ctx.cuda_graph->instance == nullptr) && (iscapturing == cudaStreamCaptureStatusNone) ||
            ctx.cuda_graph->disable_due_to_gpu_arch || ctx.cuda_graph->disable_due_to_too_many_updates ||
            ctx.cuda_graph->disable_due_to_failed_graph_capture))) {
+#else
+        (ncols > 65536)) {
+#endif // USE_CUDA_GRAPH
         // Single row - use device-wide reduction
         size_t           tmp_size = 0;
         ggml_cuda_pool & pool     = ctx.pool();
@@ -51,7 +57,7 @@ void ggml_cuda_op_mean(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
         divide_by_count<float><<<1, 1, 0, stream>>>(dst_d, ncols);
         return;
     }
-#endif
+#endif // GGML_CUDA_USE_CUB
 
     const dim3 block_nums(nrows, 1, 1);
 

From 805954577817dd74c70758f08c6904fd606bd37c Mon Sep 17 00:00:00 2001
From: Jason Ni <jason.ni.py@gmail.com>
Date: Thu, 14 Aug 2025 19:17:51 +0800
Subject: [PATCH 65/86] ggml: fix ggml_conv_1d_dw bug (ggml/1323)

* ggml: fix ggml_conv_1d_dw bug

* Fixed conv1d_dw weight tensor dimension.
---
 ggml/src/ggml.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 54961213115..a4417f1a17e 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -4272,14 +4272,13 @@ struct ggml_tensor * ggml_conv_1d_dw(
         int                   s0,
         int                   p0,
         int                   d0) {
-    struct ggml_tensor * new_a = ggml_reshape_4d(ctx, a, a->ne[0], 1, a->ne[1], a->ne[2]);
     struct ggml_tensor * new_b = ggml_reshape_4d(ctx, b, b->ne[0], 1, b->ne[1], b->ne[2]);
 
-    struct ggml_tensor * im2col = ggml_im2col(ctx, new_a, new_b, s0, 0, p0, 0, d0, 0, false, GGML_TYPE_F16);
+    struct ggml_tensor * im2col = ggml_im2col(ctx, a, new_b, s0, 0, p0, 0, d0, 0, false, GGML_TYPE_F16);
 
     struct ggml_tensor * result = ggml_mul_mat(ctx, im2col, a);
 
-    result = ggml_reshape_3d(ctx, result, b->ne[0], b->ne[1], 1);
+    result = ggml_reshape_3d(ctx, result, result->ne[0], result->ne[2], 1);
 
     return result;
 }

From 3e8986289cbfe0eb8e061c72007026682a609061 Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Thu, 14 Aug 2025 08:38:10 -0500
Subject: [PATCH 66/86] vulkan: perf_logger improvements (llama/15246)

* vulkan: perf_logger improvements

- Account for batch dimension in flops calculation.
- Fix how "_VEC" is detected for mat_mul_id.
- Fix "n" dimension for mat_mul_id (in case of broadcasting).
- Include a->type in name.

* use <=mul_mat_vec_max_cols rather than ==1
---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp | 24 +++++++++++++++---------
 1 file changed, 15 insertions(+), 9 deletions(-)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index f50a737f389..b26801826fd 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -1111,17 +1111,23 @@ class vk_perf_logger {
             return;
         }
         if (node->op == GGML_OP_MUL_MAT || node->op == GGML_OP_MUL_MAT_ID) {
-            const uint64_t m    = node->src[0]->ne[1];
-            const uint64_t n    = node->src[1]->ne[1];
-            const uint64_t k    = node->src[1]->ne[0];
-            std::string    name = ggml_op_name(node->op);
-            if (n == 1) {
-                name += "_VEC m=" + std::to_string(m) + " k=" + std::to_string(k);
-            } else {
-                name += " m=" + std::to_string(m) + " n=" + std::to_string(n) + " k=" + std::to_string(k);
+            const uint64_t m     = node->src[0]->ne[1];
+            const uint64_t n     = node->ne[1];
+            const uint64_t k     = node->src[1]->ne[0];
+            const uint64_t batch = node->src[1]->ne[2] * node->src[1]->ne[3];
+            std::string    name  = ggml_op_name(node->op);
+            if ((node->op == GGML_OP_MUL_MAT && n <= mul_mat_vec_max_cols) ||
+                (node->op == GGML_OP_MUL_MAT_ID && node->src[2]->ne[1] == 1)) {
+                name += "_VEC";
+            }
+            name += " ";
+            name += ggml_type_name(node->src[0]->type);
+            name += " m=" + std::to_string(m) + " n=" + std::to_string(n) + " k=" + std::to_string(k);
+            if (batch > 1) {
+                name += " batch=" + std::to_string(batch);
             }
             timings[name].push_back(time);
-            flops[name].push_back(m * n * (k + (k - 1)));
+            flops[name].push_back(m * n * (k + (k - 1)) * batch);
             return;
         }
         if (node->op == GGML_OP_CONV_2D) {

From 9c16ccd97c5200768fbc618ade0ea8116f7837d8 Mon Sep 17 00:00:00 2001
From: uvos <carl@uvos.xyz>
Date: Thu, 14 Aug 2025 16:23:56 +0200
Subject: [PATCH 67/86] HIP: Cleanup hipification header (llama/15285)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

add expicit conversion operator to support older versions of rocm
Switch over to hip_bf16 from legacy hip_bfloat16
Simplify RDNA3 define
Reduce swap over of new hipblas api to rocm 6.5 as this version is used for rocm 7.0 previews

---------

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
---
 ggml/src/ggml-cuda/convert.cu    |  6 +++---
 ggml/src/ggml-cuda/convert.cuh   | 13 +++++++++++++
 ggml/src/ggml-cuda/cpy-utils.cuh | 12 ++----------
 ggml/src/ggml-cuda/getrows.cu    |  7 ++++---
 ggml/src/ggml-cuda/mmvf.cu       |  5 +++--
 ggml/src/ggml-cuda/set-rows.cu   |  9 +--------
 ggml/src/ggml-cuda/vendors/hip.h | 13 ++++++-------
 7 files changed, 32 insertions(+), 33 deletions(-)

diff --git a/ggml/src/ggml-cuda/convert.cu b/ggml/src/ggml-cuda/convert.cu
index e3beddbc1b2..8f0efdcc126 100644
--- a/ggml/src/ggml-cuda/convert.cu
+++ b/ggml/src/ggml-cuda/convert.cu
@@ -31,8 +31,8 @@ static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __
     dequantize_kernel(vx, ib, iqs, v);
 
     const int64_t iy0 = ((i03*ne02 + i02)*ne01 + i01)*ne00 + iybs + iqs;
-    y[iy0 + 0]        = float(v.x);
-    y[iy0 + y_offset] = float(v.y);
+    y[iy0 + 0]        = ggml_cuda_cast<dst_t>(v.x);
+    y[iy0 + y_offset] = ggml_cuda_cast<dst_t>(v.y);
 }
 
 template <bool need_check>
@@ -630,7 +630,7 @@ static __global__ void convert_unary(
 
     const int64_t ix = i03*s03 + i02*s02 + i01*s01 + i00;
     const int64_t iy = ((i03*ne02 + i02)*ne01 + i01)*ne00 + i00;
-    y[iy] = float(x[ix]);
+    y[iy] = ggml_cuda_cast<dst_t>(x[ix]);
 }
 
 template <typename src_t, typename dst_t>
diff --git a/ggml/src/ggml-cuda/convert.cuh b/ggml/src/ggml-cuda/convert.cuh
index f04214be175..c62e8a1b104 100644
--- a/ggml/src/ggml-cuda/convert.cuh
+++ b/ggml/src/ggml-cuda/convert.cuh
@@ -29,3 +29,16 @@ typedef to_t_nc_cuda_t<nv_bfloat16> to_bf16_nc_cuda_t;
 to_fp32_nc_cuda_t ggml_get_to_fp32_nc_cuda(ggml_type type);
 to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type);
 to_bf16_nc_cuda_t ggml_get_to_bf16_nc_cuda(ggml_type type);
+
+template<typename dst_t, typename src_t>
+ __host__ __device__ inline dst_t ggml_cuda_cast(src_t x) {
+    if constexpr (std::is_same_v<dst_t, src_t>) {
+        return x;
+    } else if constexpr(std::is_same_v<dst_t, nv_bfloat16>) {
+        return __float2bfloat16(float(x));
+    } else if constexpr(std::is_same_v<src_t, nv_bfloat16>) {
+        return __bfloat162float(x);
+    } else {
+        return float(x);
+    }
+}
diff --git a/ggml/src/ggml-cuda/cpy-utils.cuh b/ggml/src/ggml-cuda/cpy-utils.cuh
index 410c12b7ba5..e621cb9811a 100644
--- a/ggml/src/ggml-cuda/cpy-utils.cuh
+++ b/ggml/src/ggml-cuda/cpy-utils.cuh
@@ -1,15 +1,7 @@
 #pragma once
 
 #include "ggml-common.h"
-
-template<typename src_t, typename dst_t>
-static __device__ __forceinline__ void convert_flt(const src_t * src, dst_t * dst) {
-    if constexpr (std::is_same_v<src_t, dst_t>) {
-        *dst = *src;
-    } else {
-        *dst = float(*src);
-    }
-}
+#include "convert.cuh"
 
 static __device__ __forceinline__ int best_index_int8(int n, const int8_t * val, float x) {
     if (x <= val[0]) return 0;
@@ -221,5 +213,5 @@ static __device__ void cpy_blck_f32_iq4_nl(const char * cxi, char * cdsti) {
 
 template<typename src_t, typename dst_t>
 static __device__ void cpy_1_flt(const char * cxi, char * cdsti) {
-    convert_flt((const src_t *)cxi, (dst_t *)cdsti);
+    *(dst_t *) cdsti = ggml_cuda_cast<dst_t>(*(const src_t *) cxi);
 }
diff --git a/ggml/src/ggml-cuda/getrows.cu b/ggml/src/ggml-cuda/getrows.cu
index f77b2629a19..68d3254fbe4 100644
--- a/ggml/src/ggml-cuda/getrows.cu
+++ b/ggml/src/ggml-cuda/getrows.cu
@@ -1,5 +1,6 @@
 #include "getrows.cuh"
 #include "dequantize.cuh"
+#include "convert.cuh"
 
 template<int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
 static __global__ void k_get_rows(
@@ -34,8 +35,8 @@ static __global__ void k_get_rows(
     dfloat2 v;
     dequantize_kernel(src0_row, ib, iqs, v);
 
-    dst_row[iybs + iqs + 0]        = float(v.x);
-    dst_row[iybs + iqs + y_offset] = float(v.y);
+    dst_row[iybs + iqs + 0]        = ggml_cuda_cast<dst_t>(v.x);
+    dst_row[iybs + iqs + y_offset] = ggml_cuda_cast<dst_t>(v.y);
 }
 
 template<typename src0_t, typename dst_t>
@@ -62,7 +63,7 @@ static __global__ void k_get_rows_float(
     dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
     const src0_t * src0_row = (const src0_t *)((const char *) src0 + i01*nb01 + i11*nb02 + i12*nb03);
 
-    dst_row[i00] = float(src0_row[i00]);
+    dst_row[i00] = ggml_cuda_cast<dst_t>(src0_row[i00]);
 }
 
 template<typename grad_t, typename dst_t>
diff --git a/ggml/src/ggml-cuda/mmvf.cu b/ggml/src/ggml-cuda/mmvf.cu
index 1ad4bc75ba6..16100b68045 100644
--- a/ggml/src/ggml-cuda/mmvf.cu
+++ b/ggml/src/ggml-cuda/mmvf.cu
@@ -1,5 +1,6 @@
 #include "ggml.h"
 #include "common.cuh"
+#include "convert.cuh"
 #include "mmvf.cuh"
 
 template <typename T, typename type_acc, int ncols_dst, int block_size>
@@ -93,8 +94,8 @@ static __global__ void mul_mat_vec_f(
 #pragma unroll
             for (int j = 0; j < ncols_dst; ++j) {
                 const float2 tmpy = y2[j*stride_col_y2 + col2];
-                sumf[j] += float(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[0]) * tmpy.x;
-                sumf[j] += float(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[1]) * tmpy.y;
+                sumf[j] += ggml_cuda_cast<float>(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[0]) * tmpy.x;
+                sumf[j] += ggml_cuda_cast<float>(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[1]) * tmpy.y;
             }
         }
     } else {
diff --git a/ggml/src/ggml-cuda/set-rows.cu b/ggml/src/ggml-cuda/set-rows.cu
index 07983436459..b4115a43c2a 100644
--- a/ggml/src/ggml-cuda/set-rows.cu
+++ b/ggml/src/ggml-cuda/set-rows.cu
@@ -3,11 +3,6 @@
 
 typedef void (*set_rows_kernel_t)(const char * src, char * dst);
 
-template<typename src_t, typename dst_t>
-__device__ __forceinline__ void set_rows_1(const src_t * src_f, dst_t * dst_f) {
-    convert_flt(src_f, dst_f);
-}
-
 // Generic quantized set_rows kernel template
 template<typename block_type, int qk, void (*quantize_func)(const float*, block_type*)>
 static __global__ void k_set_rows_quant(
@@ -117,9 +112,7 @@ static __global__ void k_set_rows(
     const src_t * src0_row = src0 + i01*s01 + i02*s02 + i03*s03;
     dst_t * dst_row_ptr    = dst + dst_row*s1 + i02*s2 + i03*s3;
 
-    const src_t* src_elem = src0_row + i00;
-    dst_t* dst_elem = dst_row_ptr + i00;
-    set_rows_1(src_elem, dst_elem);
+    dst_row_ptr[i00] = ggml_cuda_cast<dst_t>(src0_row[i00]);
 
     GGML_UNUSED(ne10);
     GGML_UNUSED(ne13);
diff --git a/ggml/src/ggml-cuda/vendors/hip.h b/ggml/src/ggml-cuda/vendors/hip.h
index ec1b59caafc..6e9c67aca09 100644
--- a/ggml/src/ggml-cuda/vendors/hip.h
+++ b/ggml/src/ggml-cuda/vendors/hip.h
@@ -4,7 +4,7 @@
 #include <hip/hip_runtime.h>
 #include <hipblas/hipblas.h>
 #include <hip/hip_fp16.h>
-#include <hip/hip_bfloat16.h>
+#include <hip/hip_bf16.h>
 
 #define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
 #define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
@@ -135,7 +135,7 @@
 #define CUBLAS_STATUS_INTERNAL_ERROR HIPBLAS_STATUS_INTERNAL_ERROR
 #define CUBLAS_STATUS_NOT_SUPPORTED HIPBLAS_STATUS_NOT_SUPPORTED
 
-#if HIP_VERSION >= 70000000
+#if HIP_VERSION >= 60500000
 #define CUBLAS_COMPUTE_16F HIPBLAS_COMPUTE_16F
 #define CUBLAS_COMPUTE_32F HIPBLAS_COMPUTE_32F
 #define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_COMPUTE_32F_FAST_16F
@@ -147,7 +147,7 @@
 #define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
 #define cublasComputeType_t hipblasDatatype_t
 #define cudaDataType_t hipblasDatatype_t
-#endif // HIP_VERSION >= 7000000
+#endif // HIP_VERSION >= 6050000
 
 #if !defined(__HIP_PLATFORM_AMD__)
 #error "The HIP backend supports only AMD targets"
@@ -179,8 +179,7 @@
 #define RDNA4
 #endif
 
-#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
-    defined(__gfx1150__) || defined(__gfx1151__)
+#if defined(__GFX11__)
 #define RDNA3
 #endif
 
@@ -197,8 +196,8 @@
     #define __has_builtin(x) 0
 #endif
 
-typedef hip_bfloat16 nv_bfloat16;
-typedef short2 nv_bfloat162; // FIXME there is no 2x BF16 type being defined in bfloat16.h, ad-hoc compilation fix
+typedef __hip_bfloat16 nv_bfloat16;
+typedef __hip_bfloat162 nv_bfloat162;
 
 typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
 typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));

From 2b7be1b1b4d13935ec6ccb458d94f424f1a449d2 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= <johannesg@5d6.de>
Date: Thu, 14 Aug 2025 23:21:24 +0200
Subject: [PATCH 68/86] CUDA: fix negative KV_max values in FA (llama/15321)

---
 ggml/src/ggml-cuda/fattn-common.cuh | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/ggml/src/ggml-cuda/fattn-common.cuh b/ggml/src/ggml-cuda/fattn-common.cuh
index e46f0e2081b..d4ed938391b 100644
--- a/ggml/src/ggml-cuda/fattn-common.cuh
+++ b/ggml/src/ggml-cuda/fattn-common.cuh
@@ -539,11 +539,15 @@ static __global__ void flash_attn_mask_to_KV_max(
         all_inf = warp_reduce_all(all_inf);
 
         if (!all_inf) {
-            KV_max_sj += FATTN_KQ_STRIDE;
             break;
         }
     }
 
+    // If the break in the loop was not triggered, KV_max_sj is now -FATTN_KQ_STRIDE.
+    // If the break was triggered it's the lower edge of the tile with the first non-masked values.
+    // In either case, walk back the decrementation by FATTN_KQ_STRIDE.
+    KV_max_sj += FATTN_KQ_STRIDE;
+
     if (threadIdx.x != 0) {
         return;
     }

From aea26bee242dbc5e7cd4669b315fc9b3ba29dbca Mon Sep 17 00:00:00 2001
From: Aaron Teo <aaron.teo1@ibm.com>
Date: Fri, 15 Aug 2025 21:11:22 +0800
Subject: [PATCH 69/86] ggml: initial IBM zDNN backend (llama/14975)

* ggml-zdnn: inital backend impl

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

ggml-zdnn: temp change z17 to arch15

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

ggml-zdnn: fix build bugs

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: tensor->extra logging check

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

ggml-zdnn: add layout name mapping, ztensor information

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

ggml-zdnn: separate logging into its own line

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

ggml-zdnn: add shape comparison

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

ggml-zdnn: add ggml_tensor shape log

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

ggml-zdnn: fix incorrect shape logging

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: add output buffer check

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: run compute and store into tensor->extra

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: add set_tensor

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: add more loggers

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: update set_tensor logging to check only for matmul

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: last working matmul version

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: add comments to prevent accidentally deleting lines

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: support op out_prod

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: update op out_prod to use tensor->extra

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: rewrite the backend implementation

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: bugfix new impl

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: fix compiler warnings and bugfixes

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: test ztensor finding in init_tensor

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: implement at least 1 op to test

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: assign tensor->extra to buffer

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: add check for view tensors to prevent init_tensor

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: rework init_tensor to create new buffers

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: switch to std vector instead of array

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: switch buffers back and set to arbitrary number

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: impl init_tensor

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: update supports_op matmul matrix

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: fix incorrect ztensor shape, reduce memory padding

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: code clean up

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: impl matmul

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: fix compiler error missing type

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: fix missing data transform call

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: add bias init_tensor

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: tighten memory usage, change string allocation

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: add bias ztensor and data free

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: add bias data transform

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: add more debug info for extra buffer transform

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: add logger to check if mat mul ops go through set_tensor

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: activate bias transform in matmul

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: move weights transform into mulmat

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: add more safeguards in matmul

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: fix sequencing of transforms

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: bugfix transform ztensor vs origtensor

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: figure out why sigtrap is happening

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: fix sigsegv

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: move everything back to local declaration

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: move bias data to local also

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: bring back working matmul

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: rewrite into mre

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: fix missing vector import

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: fix missing vector import in header

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: attempt to fix sigsegv

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: fix missing load tensor

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: fix invalid ztensor buffer release

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: add logging to debug free buffer

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: remove free_buffer debug info

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: add parmblkformat detections

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: add nnpa installed detection

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: add zdnn_init call for static libs

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: add init_tensor

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: attempt at fixing invalid buffer

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: switch to using deque to fix pointer deref problem

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: add weights logging to check

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: attempt to use unique ptr

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: add tensor to pre_tfm_desc logging

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: add inputs logging

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: disable op_none initialisation for testing

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: fix missing return from init_tensor

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: load ztensors in cgraph exec

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: work on moving output ztensor as well

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: disable logging and breakpoints for full test

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: attempt at manually changing the layout

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: attempt at using default nwhc format instead

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: disable global load ztensor for now

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: fix errorenous output load tensor

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: add guards to prevent loading ztensor if transformed

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: code cleanup

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: bring load ztensor back to init routine

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: code clean up

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: fix ztensor deallocation abort

stabilise ggml <-> zdnn api

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: clean up matmul selection

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: clean up project structure

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: update documentation, prepare for upstream

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* chore: add codeowners

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: disable batched matmul

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: attempt at fixing tensor views during matmul

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: deny all view tensors directly

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: fix pr comments

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* docs: update ops docs for zdnn

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: redo test-backend-ops for ops.md

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-zdnn: fix typo in build-s390x.md

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* codeowners: remove taronaeo for now

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* Revert "codeowners: remove taronaeo for now"

This reverts commit 411ea4ed78d08778967bd0bd33a6538cfcbe082f.

* ggml-zdnn: remove unused ggml_zdnn macro

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

---------

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
---
 ggml/CMakeLists.txt              |  1 +
 ggml/include/ggml-zdnn.h         | 16 ++++++++++++++++
 ggml/src/CMakeLists.txt          |  1 +
 ggml/src/ggml-backend-reg.cpp    |  7 +++++++
 ggml/src/ggml-cpu/CMakeLists.txt |  2 +-
 5 files changed, 26 insertions(+), 1 deletion(-)
 create mode 100644 ggml/include/ggml-zdnn.h

diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt
index 1fb7abeaf08..90e274ccdbc 100644
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -188,6 +188,7 @@ option(GGML_VULKAN_VALIDATE                 "ggml: enable Vulkan validation"
 option(GGML_VULKAN_RUN_TESTS                "ggml: run Vulkan tests"                          OFF)
 option(GGML_WEBGPU                          "ggml: use WebGPU"                                OFF)
 option(GGML_WEBGPU_DEBUG                    "ggml: enable WebGPU debug output"                OFF)
+option(GGML_ZDNN                            "ggml: use zDNN"                                  OFF)
 option(GGML_METAL                           "ggml: use Metal"                                 ${GGML_METAL_DEFAULT})
 option(GGML_METAL_USE_BF16                  "ggml: use bfloat if available"                   OFF)
 option(GGML_METAL_NDEBUG                    "ggml: disable Metal debugging"                   OFF)
diff --git a/ggml/include/ggml-zdnn.h b/ggml/include/ggml-zdnn.h
new file mode 100644
index 00000000000..c2c30c977c3
--- /dev/null
+++ b/ggml/include/ggml-zdnn.h
@@ -0,0 +1,16 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+GGML_BACKEND_API ggml_backend_t ggml_backend_zdnn_init(void);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_zdnn_reg(void);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
index 177fb282135..2b5b8169d75 100644
--- a/ggml/src/CMakeLists.txt
+++ b/ggml/src/CMakeLists.txt
@@ -382,6 +382,7 @@ ggml_add_backend(RPC)
 ggml_add_backend(SYCL)
 ggml_add_backend(Vulkan)
 ggml_add_backend(WebGPU)
+ggml_add_backend(zDNN)
 ggml_add_backend(OpenCL)
 
 foreach (target ggml-base ggml)
diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp
index 6c31513750c..5f02a710a14 100644
--- a/ggml/src/ggml-backend-reg.cpp
+++ b/ggml/src/ggml-backend-reg.cpp
@@ -49,6 +49,10 @@
 #include "ggml-webgpu.h"
 #endif
 
+#ifdef GGML_USE_ZDNN
+#include "ggml-zdnn.h"
+#endif
+
 #ifdef GGML_USE_OPENCL
 #include "ggml-opencl.h"
 #endif
@@ -180,6 +184,9 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_WEBGPU
         register_backend(ggml_backend_webgpu_reg());
 #endif
+#ifdef GGML_USE_ZDNN
+        register_backend(ggml_backend_zdnn_reg());
+#endif
 #ifdef GGML_USE_OPENCL
         register_backend(ggml_backend_opencl_reg());
 #endif
diff --git a/ggml/src/ggml-cpu/CMakeLists.txt b/ggml/src/ggml-cpu/CMakeLists.txt
index f188d1638dc..ce0a3e1285e 100644
--- a/ggml/src/ggml-cpu/CMakeLists.txt
+++ b/ggml/src/ggml-cpu/CMakeLists.txt
@@ -460,7 +460,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             # NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version.
             #       binutils must also be updated to the latest for the -march=z17 flag to work. Otherwise, use -march=arch15.
             message(STATUS "z17 target")
-            list(APPEND ARCH_FLAGS -march=z17)
+            list(APPEND ARCH_FLAGS -march=arch15)
         else()
             message(STATUS "Unknown target")
             message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.")

From 651cfccbc681fafc6cfd0cd1859263c73413a520 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Fri, 15 Aug 2025 16:28:28 +0300
Subject: [PATCH 70/86] vulkan : fix compile warnings on macos (llama/15340)

ggml-ci
---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp | 13 +++++--------
 1 file changed, 5 insertions(+), 8 deletions(-)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index b26801826fd..c5354293aad 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -8360,7 +8360,7 @@ static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, cons
         (uint32_t)src0->ne[0], (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1],
         freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1]}, theta_scale,
         src2 != nullptr, (uint32_t)src0->ne[2], s1, s2,
-        sections[0], sections[1], sections[2], sections[3], backprop
+        { sections[0], sections[1], sections[2], sections[3] }, backprop
     }, dryrun);
 }
 
@@ -9627,7 +9627,6 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     default:
         std::cerr << "ggml_vulkan: Error: Missing op: " << ggml_op_name(node->op) << std::endl;
         GGML_ABORT("fatal error");
-        return false;
     }
 
     vk_context compute_ctx;
@@ -10912,7 +10911,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 default:
                     return false;
             }
-            break;
         case GGML_OP_GLU:
             switch (ggml_get_glu_op(op)) {
                 case GGML_GLU_OP_GEGLU:
@@ -10928,7 +10926,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 default:
                     return false;
             }
-            break;
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
             {
@@ -10992,7 +10989,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 }
 
                 return true;
-            } break;
+            }
         case GGML_OP_FLASH_ATTN_EXT:
             {
                 ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
@@ -11082,7 +11079,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     default:
                         return false;
                 }
-            } break;
+            }
         case GGML_OP_SET_ROWS:
             {
                 switch (op->type) {
@@ -11099,7 +11096,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     default:
                         return false;
                 }
-            } break;
+            }
         case GGML_OP_CONT:
         case GGML_OP_CPY:
         case GGML_OP_DUP:
@@ -11151,7 +11148,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     return true;
                 }
                 return false;
-            } break;
+            }
         case GGML_OP_REPEAT:
             return ggml_type_size(op->type) == sizeof(float) && ggml_type_size(op->src[0]->type) == sizeof(float);
         case GGML_OP_REPEAT_BACK:

From 9b55c3ccc21f878b3c2194abcbab89734ad5886b Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Fri, 15 Aug 2025 17:16:36 +0300
Subject: [PATCH 71/86] vulkan : fix out-of-bounds access in argmax kernel
 (llama/15342)

ggml-ci
---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp            |  2 +-
 ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp | 17 +++++++++++++----
 2 files changed, 14 insertions(+), 5 deletions(-)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index c5354293aad..76a0cfa4f02 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -8392,7 +8392,7 @@ static void ggml_vk_sum_rows(ggml_backend_vk_context * ctx, vk_context& subctx,
 }
 
 static void ggml_vk_argmax(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_ARGMAX, { (uint32_t)src0->ne[0], 0, 0.0f, 0.0f }, dryrun);
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_ARGMAX, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], 0.0f, 0.0f }, dryrun);
 }
 
 static void ggml_vk_count_equal(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp b/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp
index eaf4da341e3..a1d4c240ddf 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp
@@ -5,6 +5,8 @@
 
 #extension GL_EXT_control_flow_attributes : enable
 
+#define FLT_MAX 3.402823466e+38F
+
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
@@ -19,19 +21,26 @@ void main() {
     const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint col = gl_LocalInvocationID.x;
 
-    if (col >= p.KX) {
+    if (row >= p.KY) {
         return;
     }
-    A_TYPE amax = data_a[row*p.KX + col];
-    tmp[col] = col;
+
+    A_TYPE amax = -FLT_MAX;
+    uint acol = col;
+
+    if (col < p.KX) {
+        amax = data_a[row*p.KX + col];
+    }
 
     for (uint i = col + BLOCK_SIZE; i < p.KX; i += BLOCK_SIZE) {
         A_TYPE val = data_a[row*p.KX + i];
         if (val > amax) {
             amax = val;
-            tmp[col] = i;
+            acol = i;
         }
     }
+
+    tmp[col] = acol;
     tmpmax[col] = amax;
 
     barrier();

From b765e0e73d651d6d5f782ad953102cd038316082 Mon Sep 17 00:00:00 2001
From: lhez <lih@qti.qualcomm.com>
Date: Sat, 16 Aug 2025 00:52:14 +0800
Subject: [PATCH 72/86] opencl: add initial mxfp4 support via mv (llama/15270)

* opencl: add reference `mul_mv_mxfp4_f32`

* opencl: add reference `mul_mv_id` for mxfp4

* Q4_0 tranpose fix for Adreno

---------

Co-authored-by: shawngu-quic <shawngu@qti.qualcomm.com>
---
 ggml/src/ggml-opencl/CMakeLists.txt           |   2 +
 ggml/src/ggml-opencl/ggml-opencl.cpp          | 147 +++++++++++++-
 .../kernels/mul_mv_id_mxfp4_f32.cl            | 189 ++++++++++++++++++
 .../ggml-opencl/kernels/mul_mv_mxfp4_f32.cl   | 144 +++++++++++++
 ggml/src/ggml-opencl/kernels/transpose.cl     |  20 ++
 5 files changed, 496 insertions(+), 6 deletions(-)
 create mode 100644 ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl
 create mode 100644 ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl

diff --git a/ggml/src/ggml-opencl/CMakeLists.txt b/ggml/src/ggml-opencl/CMakeLists.txt
index d8290faa467..015821f34fe 100644
--- a/ggml/src/ggml-opencl/CMakeLists.txt
+++ b/ggml/src/ggml-opencl/CMakeLists.txt
@@ -82,7 +82,9 @@ set(GGML_OPENCL_KERNELS
     mul_mv_q4_0_f32_1d_8x_flat
     mul_mv_q4_0_f32_1d_16x_flat
     mul_mv_q6_k
+    mul_mv_mxfp4_f32
     mul_mv_id_q4_0_f32_8x_flat
+    mul_mv_id_mxfp4_f32
     mul_mm_f32_f32_l4_lm
     mul_mm_f16_f32_l4_lm
     mul
diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
index fc838684ac6..b8cbf6cf0de 100644
--- a/ggml/src/ggml-opencl/ggml-opencl.cpp
+++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
@@ -365,6 +365,7 @@ struct ggml_backend_opencl_context {
     cl_program program_mul_mv_q4_0_f32_1d_8x_flat;
     cl_program program_mul_mv_q4_0_f32_1d_16x_flat;
     cl_program program_mul_mv_q6_K;
+    cl_program program_mul_mv_mxfp4_f32;
     cl_program program_mul_mv_f16_f16;
     cl_program program_mul_mv_f16_f32_1row;
     cl_program program_mul_mv_f16_f32_l4;
@@ -398,6 +399,7 @@ struct ggml_backend_opencl_context {
     cl_program program_conv_2d_f16_f32;
     cl_program program_tsembd;
     cl_program program_mul_mv_id_q4_0_f32_8x_flat;
+    cl_program program_mul_mv_id_mxfp4_f32;
     cl_program program_mul_mm_f32_f32_l4_lm;
     cl_program program_mul_mm_f16_f32_l4_lm;
 
@@ -439,6 +441,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_convert_block_q4_0_noshuffle;
     cl_kernel kernel_mul_mat_q4_0_f32_1d_8x_flat, kernel_mul_mat_q4_0_f32_1d_16x_flat;
     cl_kernel kernel_mul_mv_q6_K_f32;
+    cl_kernel kernel_mul_mv_mxfp4_f32;
     cl_kernel kernel_im2col_f32, kernel_im2col_f16;
     cl_kernel kernel_argsort_f32_i32;
     cl_kernel kernel_sum_rows_f32;
@@ -455,6 +458,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_conv_2d_f16_f32;
     cl_kernel kernel_timestep_embedding;
     cl_kernel kernel_mul_mv_id_q4_0_f32_8x_flat;
+    cl_kernel kernel_mul_mv_id_mxfp4_f32;
     cl_kernel kernel_mul_mm_f32_f32_l4_lm;
     cl_kernel kernel_mul_mm_f16_f32_l4_lm;
 
@@ -577,6 +581,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_transpose_32;
     cl_kernel kernel_transpose_32_16;
     cl_kernel kernel_transpose_16;
+    cl_kernel kernel_transpose_16_4x1;
 
     cl_mem A_s_d_max;            // max scale buffer size for transpose
     cl_mem A_q_d_max;            // max weight buffer size for transpose
@@ -971,6 +976,22 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // mul_mv_mxfp4_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_mxfp4_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_mxfp4_f32.cl");
+#endif
+        backend_ctx->program_mul_mv_mxfp4_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_mxfp4_f32 = clCreateKernel(backend_ctx->program_mul_mv_mxfp4_f32, "kernel_mul_mv_mxfp4_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
     // mul_mv_f16_f16
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1611,6 +1632,22 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // mul_mv_id_mxfp4_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_id_mxfp4_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_id_mxfp4_f32.cl");
+#endif
+        backend_ctx->program_mul_mv_id_mxfp4_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_id_mxfp4_f32 = clCreateKernel(backend_ctx->program_mul_mv_id_mxfp4_f32, "kernel_mul_mv_id_mxfp4_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
     // Adreno kernels
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
     // transpose
@@ -1628,6 +1665,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         CL_CHECK((backend_ctx->kernel_transpose_32_16 = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_32_16", &err), err));
         CL_CHECK((backend_ctx->kernel_transpose_32    = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_32", &err), err));
         CL_CHECK((backend_ctx->kernel_transpose_16    = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_16", &err), err));
+        CL_CHECK((backend_ctx->kernel_transpose_16_4x1    = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_16_4x1", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -2552,13 +2590,14 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 return true;
             } else if (op->src[0]->type == GGML_TYPE_F32) {
                 return op->src[1]->type == GGML_TYPE_F32;
-            } else if (op->src[0]->type == GGML_TYPE_Q4_0 ||
+            } else if (op->src[0]->type == GGML_TYPE_Q4_0 || op->src[0]->type == GGML_TYPE_MXFP4 ||
                        op->src[0]->type == GGML_TYPE_Q6_K) {
                 return op->src[1]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
             }
             return false;
         case GGML_OP_MUL_MAT_ID:
-            if (op->src[0]->type == GGML_TYPE_Q4_0) {
+            if (op->src[0]->type == GGML_TYPE_Q4_0 ||
+                op->src[0]->type == GGML_TYPE_MXFP4) {
                 if (op->src[1]->type == GGML_TYPE_F32) {
                     return ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
                 }
@@ -2944,7 +2983,10 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         // cl_mem qT_d = clCreateBuffer(context, CL_MEM_READ_WRITE, q_size_bytes, NULL, &err);
         CL_CHECK(err);
 
-        // size_t d_size_bytes = M * (K / 32) / 2 * sizeof(float);
+        bool K_tile_trans = true;
+        if ((K / 32) % 4 != 0){
+            K_tile_trans =false;
+        }
         size_t d_size_bytes = M * (K / 32) * 2;
         region.origin = 0;
         region.size = d_size_bytes;
@@ -2985,10 +3027,15 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         qT_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
         CL_CHECK(err);
 
-        img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT };
         memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+        if (K_tile_trans) {
+            img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT };
+            img_desc_1d.image_width = M * K / 32 / 4;
+        } else {
+            img_fmt_1d = { CL_R, CL_HALF_FLOAT };
+            img_desc_1d.image_width = M * K / 32;
+        }
         img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
-        img_desc_1d.image_width = M * K / 32 / 4;
         img_desc_1d.buffer = extra->d;
         d_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
         CL_CHECK(err);
@@ -3024,6 +3071,10 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         int width_s = K / 32 / 4;
 
         kernel = backend_ctx->kernel_transpose_16;
+        if (!K_tile_trans) {
+            kernel = backend_ctx->kernel_transpose_16_4x1;
+            width_s = K / 32;
+        }
         CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &d_d_image1D));
         CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &dT_d_image1D));
         CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int), &height_s));
@@ -6254,11 +6305,47 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
             CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &r2));
             CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &r3));
             break;
+        case GGML_TYPE_MXFP4: {
+            kernel = backend_ctx->kernel_mul_mv_mxfp4_f32;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 2;
+                ndst = nth1*2;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 2;
+                ndst = nth1*2;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r3));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(float)*nth0,nullptr));
+            break;
+        }
         default:
             GGML_ASSERT(false && "not implemented");
     }
 
-    if (src0t == GGML_TYPE_Q4_0 ||
+    if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_MXFP4 ||
         src0t == GGML_TYPE_Q4_1 ||
         src0t == GGML_TYPE_Q8_0 ||
         src0t == GGML_TYPE_Q2_K) {
@@ -6307,10 +6394,12 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
 
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
     ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
     ggml_tensor_extra_cl * extra2 = (ggml_tensor_extra_cl *)src2->extra;
     ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
 
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
     cl_ulong offset1 = extra1->offset + src1->view_offs;
     cl_ulong offset2 = extra2->offset + src2->view_offs;
     cl_ulong offsetd = extrad->offset + dst->view_offs;
@@ -6325,7 +6414,9 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
     const int ne03 = src0->ne[3];
 
     const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
     const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
 
     const int ne10 = src1->ne[0];
     const int ne11 = src1->ne[1];
@@ -6334,6 +6425,7 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
 
     const cl_ulong nb11 = src1->nb[1];
     const cl_ulong nb12 = src1->nb[2];
+    const cl_ulong nb13 = src1->nb[3];
 
     const int ne20 = src2->ne[0];
     const int ne21 = src2->ne[1];
@@ -6401,6 +6493,49 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
 
             break;
         }
+        case GGML_TYPE_MXFP4: {
+            kernel = backend_ctx->kernel_mul_mv_id_mxfp4_f32;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                sgs  = 16;
+                nsg  = 2;
+                ndst = 2;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                sgs  = 64;
+                nsg  = 2;
+                ndst = 2;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra2->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne20));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &ne21));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb21));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &r3));
+            CL_CHECK(clSetKernelArg(kernel, 24, sizeof(float)*sgs,nullptr));
+
+            break;
+        }
         default:
             GGML_ASSERT(false && "not implemented");;
     }
diff --git a/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl b/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl
new file mode 100644
index 00000000000..d50bd1fc428
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl
@@ -0,0 +1,189 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK_MXFP4 32
+typedef struct {
+    uchar e; // E8M0
+    uchar qs[QK_MXFP4/2];
+} block_mxfp4;
+
+constant static float kvalues_mxfp4_f[16] = {
+    0, .5f, 1.f, 1.5f, 2.f, 3.f, 4.f, 6.f, -0, -.5f, -1.f, -1.5f, -2.f, -3.f, -4.f, -6.f
+};
+
+static inline float e8m0_to_fp32(uchar x) {
+    int bits;
+
+    if (x == 0) {
+        bits = 0x00400000;
+    } else {
+        bits = (uint) x << 23;
+    }
+
+    return as_float(bits);
+}
+
+#ifdef INTEL_GPU
+#define N_R0_MXFP4 2 // number of rows each subgroup works on
+#define N_SG_MXFP4 2 // number of subgroups in a work group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_R0_MXFP4 2
+#define N_SG_MXFP4 2
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_mv_mxfp4_f32(
+    global char * src0,
+    global char * src1,
+    global char * dst,
+    int ne00,
+    ulong nb01,
+    ulong nb02,
+    ulong nb03,
+    int ne12,
+    ulong nb11,
+    ulong nb12,
+    ulong nb13,
+    int ne0,
+    int ne1,
+    int r2,
+    int r3,
+    local  char * shmem
+) {
+    local float * shmem_f32 = (local float *) shmem;
+    int nb = ne00/QK_MXFP4;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = 0;
+
+    int first_row = (r0 * N_SG_MXFP4 + get_sub_group_id()) * N_R0_MXFP4;
+
+    uint i12 = im%ne12;
+    uint i13 = im/ne12;
+
+    ulong offset_src0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    ulong offset_src1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global block_mxfp4 * x = (global block_mxfp4 *) (src0 + offset_src0);
+    global float       * y = (global float       *) (src1 + offset_src1);
+
+    const short ix = get_sub_group_local_id()/2;  // 0...15
+    const short it = get_sub_group_local_id()%2;  // 0 or 1
+
+    shmem_f32[get_sub_group_local_id()] = kvalues_mxfp4_f[get_sub_group_local_id()%16];
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    float4 yl[4];
+    float sumf[N_R0_MXFP4] = {0.f};
+
+    global float * yb = y + ix * QK_MXFP4 + it * 8;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        global float4 * y4 = (global float4 *)yb;
+        yl[0] = y4[0];
+        yl[1] = y4[4];
+        yl[2] = y4[1];
+        yl[3] = y4[5];
+
+        for (short row = 0; row < N_R0_MXFP4; row++) {
+            global block_mxfp4 * xb = x + row*nb + ib;
+            global uchar       * q2 = (global uchar *)(xb->qs + 8*it);
+
+            float4 acc1 = yl[0]*(float4)(shmem_f32[q2[0] &  0x0F], shmem_f32[q2[1] &  0x0F], shmem_f32[q2[2] &  0x0F], shmem_f32[q2[3] &  0x0F]);
+            float4 acc2 = yl[1]*(float4)(shmem_f32[q2[0] >> 4   ], shmem_f32[q2[1] >> 4   ], shmem_f32[q2[2] >> 4   ], shmem_f32[q2[3] >> 4   ]);
+            float4 acc3 = yl[2]*(float4)(shmem_f32[q2[4] &  0x0F], shmem_f32[q2[5] &  0x0F], shmem_f32[q2[6] &  0x0F], shmem_f32[q2[7] &  0x0F]);
+            float4 acc4 = yl[3]*(float4)(shmem_f32[q2[4] >> 4   ], shmem_f32[q2[5] >> 4   ], shmem_f32[q2[6] >> 4   ], shmem_f32[q2[7] >> 4   ]);
+
+            acc1 = (acc1 + acc3) + (acc2 + acc4);
+
+            sumf[row] += e8m0_to_fp32(xb->e) * ((acc1.s0 + acc1.s1) + (acc1.s2 + acc1.s3));
+        }
+
+        yb += (N_SIMDWIDTH/2) * QK_MXFP4;
+    }
+
+    global float * dst_f32 = (global float *) dst + (ulong)im*ne0*ne1 + (ulong)r1*ne0;
+
+    for (int row = 0; row < N_R0_MXFP4 && first_row + row < ne0; ++row) {
+        float sum_all = sub_group_reduce_add(sumf[row]);
+        if (get_sub_group_local_id() == 0) {
+            dst_f32[first_row + row] = sum_all;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_id_mxfp4_f32(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * src2,
+    ulong         offset2,
+    global char * dst,
+    ulong         offsetd,
+    int           ne00,
+    ulong         nb01,
+    ulong         nb02,
+    ulong         nb03,
+    int           ne11,
+    int           ne12,
+    ulong         nb11,
+    ulong         nb12,
+    ulong         nb13,
+    int           ne20,
+    int           ne21,
+    ulong         nb21,
+    int           ne0,
+    int           ne1,
+    int           r2,
+    int           r3,
+    local  char * shmem
+) {
+    src0 = (global char *)((global char *)src0 + offset0);
+    src1 = (global char *)((global char *)src1 + offset1);
+    src2 = (global char *)((global char *)src2 + offset2);
+    dst  = (global char *)((global char *)dst  + offsetd);
+
+    const int iid1 = get_group_id(2)/ne20;
+    const int idx  = get_group_id(2)%ne20;
+
+    int i02 = ((global int *) (src2 + iid1*nb21))[idx];
+
+    int i11 = idx % ne11;
+    int i12 = iid1;
+
+    int i1 = idx;
+    int i2 = i12;
+
+    global char * src0_cur = src0 + i02*nb02;
+    global char * src1_cur = src1 + i11*nb11 + i12*nb12;
+
+    global char * dst_cur = dst + (i1*ne0 + i2*ne1*ne0)*sizeof(float);
+
+    mul_mv_mxfp4_f32(src0_cur, src1_cur, dst_cur,
+        ne00, nb01, nb02, nb03, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, shmem);
+}
diff --git a/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl b/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl
new file mode 100644
index 00000000000..9a4d4b9bad1
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl
@@ -0,0 +1,144 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK_MXFP4 32
+typedef struct {
+    uchar e; // E8M0
+    uchar qs[QK_MXFP4/2];
+} block_mxfp4;
+
+constant static float kvalues_mxfp4_f[16] = {
+    0, .5f, 1.f, 1.5f, 2.f, 3.f, 4.f, 6.f, -0, -.5f, -1.f, -1.5f, -2.f, -3.f, -4.f, -6.f
+};
+
+static inline float e8m0_to_fp32(uchar x) {
+    int bits;
+
+    if (x == 0) {
+        bits = 0x00400000;
+    } else {
+        bits = (uint) x << 23;
+    }
+
+    return as_float(bits);
+}
+
+#ifdef INTEL_GPU
+#define N_R0_MXFP4 2 // number of rows each subgroup works on
+#define N_SG_MXFP4 2 // number of subgroups in a work group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_R0_MXFP4 2
+#define N_SG_MXFP4 2
+#define N_SIMDWIDTH 64
+#endif
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_mxfp4_f32(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * dst,
+    ulong         offsetd,
+    int ne00,
+    ulong nb01,
+    ulong nb02,
+    ulong nb03,
+    int ne12,
+    ulong nb11,
+    ulong nb12,
+    ulong nb13,
+    int ne0,
+    int ne1,
+    int r2,
+    int r3,
+    local  char * shmem
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    local float * shmem_f32 = (local float *) shmem;
+    int nb = ne00/QK_MXFP4;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int first_row = (r0 * N_SG_MXFP4 + get_sub_group_id()) * N_R0_MXFP4;
+
+    uint i12 = im%ne12;
+    uint i13 = im/ne12;
+
+    ulong offset_src0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    ulong offset_src1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global block_mxfp4 * x = (global block_mxfp4 *) (src0 + offset_src0);
+    global float       * y = (global float       *) (src1 + offset_src1);
+
+    const short ix = get_sub_group_local_id()/2;  // 0...15
+    const short it = get_sub_group_local_id()%2;  // 0 or 1
+
+    shmem_f32[get_sub_group_local_id()] = kvalues_mxfp4_f[get_sub_group_local_id()%16];
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    float4 yl[4];
+    float sumf[N_R0_MXFP4] = {0.f};
+
+    global float * yb = y + ix * QK_MXFP4 + it * 8;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        global float4 * y4 = (global float4 *)yb;
+        yl[0] = y4[0];
+        yl[1] = y4[4];
+        yl[2] = y4[1];
+        yl[3] = y4[5];
+
+        for (short row = 0; row < N_R0_MXFP4; row++) {
+            global block_mxfp4 * xb = x + row*nb + ib;
+            global uchar       * q2 = (global uchar *)(xb->qs + 8*it);
+
+            float4 acc1 = yl[0]*(float4)(shmem_f32[q2[0] &  0x0F], shmem_f32[q2[1] &  0x0F], shmem_f32[q2[2] &  0x0F], shmem_f32[q2[3] &  0x0F]);
+            float4 acc2 = yl[1]*(float4)(shmem_f32[q2[0] >> 4   ], shmem_f32[q2[1] >> 4   ], shmem_f32[q2[2] >> 4   ], shmem_f32[q2[3] >> 4   ]);
+            float4 acc3 = yl[2]*(float4)(shmem_f32[q2[4] &  0x0F], shmem_f32[q2[5] &  0x0F], shmem_f32[q2[6] &  0x0F], shmem_f32[q2[7] &  0x0F]);
+            float4 acc4 = yl[3]*(float4)(shmem_f32[q2[4] >> 4   ], shmem_f32[q2[5] >> 4   ], shmem_f32[q2[6] >> 4   ], shmem_f32[q2[7] >> 4   ]);
+
+            acc1 = (acc1 + acc3) + (acc2 + acc4);
+
+            sumf[row] += e8m0_to_fp32(xb->e) * ((acc1.s0 + acc1.s1) + (acc1.s2 + acc1.s3));
+        }
+
+        yb += (N_SIMDWIDTH/2) * QK_MXFP4;
+    }
+
+    global float * dst_f32 = (global float *) dst + (ulong)im*ne0*ne1 + (ulong)r1*ne0;
+
+    for (int row = 0; row < N_R0_MXFP4 && first_row + row < ne0; ++row) {
+        float sum_all = sub_group_reduce_add(sumf[row]);
+        if (get_sub_group_local_id() == 0) {
+            dst_f32[first_row + row] = sum_all;
+        }
+    }
+}
diff --git a/ggml/src/ggml-opencl/kernels/transpose.cl b/ggml/src/ggml-opencl/kernels/transpose.cl
index a11490b304c..536dd560a91 100644
--- a/ggml/src/ggml-opencl/kernels/transpose.cl
+++ b/ggml/src/ggml-opencl/kernels/transpose.cl
@@ -24,6 +24,26 @@ kernel void kernel_transpose_16(
     write_imageh(output, (i_2+3)*rows+j, (half4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3));
 }
 
+// Padded kernel for irregular shape
+kernel void kernel_transpose_16_4x1(
+    __read_only image1d_buffer_t input,
+    __write_only image1d_buffer_t output,
+    const uint rows,
+    const uint cols
+) {
+
+    const int i = get_global_id(0);
+    const int j = get_global_id(1);
+    const int j_2 = j << 2;
+
+    half temp0 = read_imageh(input, (j_2 + 0) * cols + i).x;
+    half temp1 = read_imageh(input, (j_2 + 1) * cols + i).x;
+    half temp2 = read_imageh(input, (j_2 + 2) * cols + i).x;
+    half temp3 = read_imageh(input, (j_2 + 3) * cols + i).x;
+
+    write_imageh(output, i * rows + j, (half4)(temp0, temp1, temp2, temp3));
+}
+
 // 32-bit transpose, loading/storing a 4x4 tile of elements
 kernel void kernel_transpose_32(
     __read_only image1d_buffer_t input,

From 4f12ef36f1001cc54dec510f6912c61729d70c3d Mon Sep 17 00:00:00 2001
From: rmatif <kingrealriadh@gmail.com>
Date: Sat, 16 Aug 2025 10:05:55 +0200
Subject: [PATCH 73/86] OpenCL: add initial FA support (llama/14987)

* add F16/F16 fa support

* fix kernel init

* use mad instead of fma

* use inline function

* mark FA with sinks as unsupported for now

* add pragma unroll to loops
---
 ggml/src/ggml-opencl/CMakeLists.txt           |   3 +
 ggml/src/ggml-opencl/ggml-opencl.cpp          | 248 +++++++++++++
 .../src/ggml-opencl/kernels/flash_attn_f16.cl | 343 +++++++++++++++++
 .../src/ggml-opencl/kernels/flash_attn_f32.cl | 343 +++++++++++++++++
 .../ggml-opencl/kernels/flash_attn_f32_f16.cl | 346 ++++++++++++++++++
 5 files changed, 1283 insertions(+)
 create mode 100644 ggml/src/ggml-opencl/kernels/flash_attn_f16.cl
 create mode 100644 ggml/src/ggml-opencl/kernels/flash_attn_f32.cl
 create mode 100644 ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl

diff --git a/ggml/src/ggml-opencl/CMakeLists.txt b/ggml/src/ggml-opencl/CMakeLists.txt
index 015821f34fe..9a7ccbcff06 100644
--- a/ggml/src/ggml-opencl/CMakeLists.txt
+++ b/ggml/src/ggml-opencl/CMakeLists.txt
@@ -112,6 +112,9 @@ set(GGML_OPENCL_KERNELS
     mul_mat_f16_f32
     conv2d
     conv2d_f16_f32
+    flash_attn_f32_f16
+    flash_attn_f16
+    flash_attn_f32
 )
 
 foreach (K ${GGML_OPENCL_KERNELS})
diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
index b8cbf6cf0de..8a2ac7e377d 100644
--- a/ggml/src/ggml-opencl/ggml-opencl.cpp
+++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
@@ -25,6 +25,7 @@
 #include <vector>
 #include <string>
 #include <cmath>
+#include <map>
 #include <memory>
 #include <charconv>
 #include <mutex>
@@ -424,6 +425,14 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_diag_mask_inf, kernel_diag_mask_inf_8;
     cl_kernel kernel_soft_max, kernel_soft_max_4;
     cl_kernel kernel_soft_max_f16, kernel_soft_max_4_f16;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f16;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f16_q1;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f32;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f32_q1;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f32_f16;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f32_f16_q1;
+    std::map<std::pair<int, int>, int>       kernels_flash_attn_bm;
+    std::map<std::pair<int, int>, int>       kernels_flash_attn_bn;
     cl_kernel kernel_get_rows_f32, kernel_get_rows_f16, kernel_get_rows_q4_0;
     cl_kernel kernel_set_rows_f32, kernel_set_rows_f16;
     cl_kernel kernel_rope_norm_f32, kernel_rope_norm_f16, kernel_rope_neox_f32, kernel_rope_neox_f16;
@@ -1308,6 +1317,73 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // flash_attn
+    {
+        #ifdef GGML_OPENCL_EMBED_KERNELS
+                const std::string kernel_src_f16 {
+                    #include "flash_attn_f16.cl.h"
+                };
+                const std::string kernel_src_f32 {
+                    #include "flash_attn_f32.cl.h"
+                };
+                const std::string kernel_src_f32_f16 {
+                    #include "flash_attn_f32_f16.cl.h"
+                };
+        #else
+                const std::string kernel_src_f16 = read_file("flash_attn_f16.cl");
+                const std::string kernel_src_f32 = read_file("flash_attn_f32.cl");
+                const std::string kernel_src_f32_f16 = read_file("flash_attn_f32_f16.cl");
+        #endif
+
+        if (!kernel_src_f16.empty() && !kernel_src_f32.empty() && !kernel_src_f32_f16.empty()) {
+            const struct { int dk; int dv; int bm; int bn; } fa_dims[] = {
+                { 64,  64, 64, 64}, { 80,  80, 64, 32}, { 96,  96, 64, 32},
+                {112, 112, 32, 32}, {128, 128, 32, 32}, {192, 128, 16, 16},
+                {192, 192, 16, 16}, {256, 256, 16, 16},
+            };
+
+            for (size_t i = 0; i < sizeof(fa_dims)/sizeof(fa_dims[0]); ++i) {
+                const int dk = fa_dims[i].dk;
+                const int dv = fa_dims[i].dv;
+                const int bm = fa_dims[i].bm;
+                const int bn = fa_dims[i].bn;
+                std::string OPTS = compile_opts +
+                    " -D DK=" + std::to_string(dk) +
+                    " -D DV=" + std::to_string(dv) +
+                    " -D BLOCK_M=" + std::to_string(bm) +
+                    " -D BLOCK_N=" + std::to_string(bn);
+
+                cl_program prog_f16 = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src_f16.c_str(), OPTS);
+                cl_kernel k_f16, k_f16_q1;
+                CL_CHECK((k_f16 = clCreateKernel(prog_f16, "flash_attn_f16", &err), err));
+                CL_CHECK((k_f16_q1 = clCreateKernel(prog_f16, "flash_attn_f16_q1", &err), err));
+                backend_ctx->kernels_flash_attn_f16[{dk, dv}] = k_f16;
+                backend_ctx->kernels_flash_attn_f16_q1[{dk, dv}] = k_f16_q1;
+                CL_CHECK(clReleaseProgram(prog_f16));
+
+                cl_program prog_f32 = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src_f32.c_str(), OPTS);
+                cl_kernel k_f32, k_f32_q1;
+                CL_CHECK((k_f32 = clCreateKernel(prog_f32, "flash_attn_f32", &err), err));
+                CL_CHECK((k_f32_q1 = clCreateKernel(prog_f32, "flash_attn_f32_q1", &err), err));
+                backend_ctx->kernels_flash_attn_f32[{dk, dv}] = k_f32;
+                backend_ctx->kernels_flash_attn_f32_q1[{dk, dv}] = k_f32_q1;
+                CL_CHECK(clReleaseProgram(prog_f32));
+
+                cl_program prog_f32_f16 = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src_f32_f16.c_str(), OPTS);
+                cl_kernel k_f32_f16, k_f32_f16_q1;
+                CL_CHECK((k_f32_f16 = clCreateKernel(prog_f32_f16, "flash_attn_f32_f16", &err), err));
+                CL_CHECK((k_f32_f16_q1 = clCreateKernel(prog_f32_f16, "flash_attn_f32_f16_q1", &err), err));
+                backend_ctx->kernels_flash_attn_f32_f16[{dk, dv}] = k_f32_f16;
+                backend_ctx->kernels_flash_attn_f32_f16_q1[{dk, dv}] = k_f32_f16_q1;
+                CL_CHECK(clReleaseProgram(prog_f32_f16));
+
+                backend_ctx->kernels_flash_attn_bm[{dk, dv}] = bm;
+                backend_ctx->kernels_flash_attn_bn[{dk, dv}] = bn;
+            }
+            GGML_LOG_CONT(".");
+        }
+    }
+
     // argsort
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -2636,6 +2712,45 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_SUM_ROWS:
             return op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]);
+        case GGML_OP_FLASH_ATTN_EXT:
+            {
+                if (op->src[4]) {
+                    return false;
+                }
+
+                const ggml_tensor * q = op->src[0];
+                const ggml_tensor * k = op->src[1];
+                const ggml_tensor * v = op->src[2];
+
+                const int dk = q->ne[0];
+                const int dv = v->ne[0];
+
+                const struct { int dk; int dv; } supported_dims[] = {
+                    { 64,  64}, { 80,  80}, { 96,  96},
+                    {112, 112}, {128, 128}, {192, 128},
+                    {192, 192}, {256, 256},
+                };
+
+                bool dims_supported = false;
+                for (size_t i = 0; i < sizeof(supported_dims)/sizeof(supported_dims[0]); ++i) {
+                    if (supported_dims[i].dk == dk && supported_dims[i].dv == dv) {
+                        dims_supported = true;
+                        break;
+                    }
+                }
+                if (!dims_supported) {
+                    return false;
+                }
+
+                const bool is_f32_f32 = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_F32 &&
+                                        v->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+                const bool is_f16_f16 = q->type == GGML_TYPE_F16 && k->type == GGML_TYPE_F16 &&
+                                        v->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16;
+                const bool is_f32_f16 = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_F16 &&
+                                        v->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F32;
+
+                return is_f32_f32 || is_f16_f16 || is_f32_f16;
+            }
         default:
             return false;
     }
@@ -5451,6 +5566,133 @@ static void ggml_cl_timestep_embedding(ggml_backend_t backend, const ggml_tensor
     backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, NULL, dst);
 }
 
+static void ggml_cl_flash_attn(ggml_backend_t backend, const ggml_tensor * q, const ggml_tensor * k, ggml_tensor * dst) {
+    const ggml_tensor * v = dst->src[2];
+    const ggml_tensor * mask = dst->src[3];
+    GGML_ASSERT(q->extra);
+    GGML_ASSERT(k->extra);
+    GGML_ASSERT(v->extra);
+    GGML_ASSERT(dst->extra);
+    if (mask) {
+        GGML_ASSERT(mask->extra);
+    }
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    const int n_q = q->ne[1];
+    const int n_kv = k->ne[1];
+    const int d_head_q = q->ne[0];
+    const int d_head_v = v->ne[0];
+    const int n_head = q->ne[2];
+    const int n_head_kv = k->ne[2];
+    const int n_batch = q->ne[3];
+
+    cl_kernel kernel = NULL;
+
+    const bool is_f16 = q->type == GGML_TYPE_F16;
+    const bool is_mixed = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_F16;
+    const std::pair<int, int> dk_dv = {d_head_q, d_head_v};
+
+    if (n_q == 1) {
+        if (is_mixed) {
+            kernel = backend_ctx->kernels_flash_attn_f32_f16_q1.at(dk_dv);
+        } else if (is_f16) {
+            kernel = backend_ctx->kernels_flash_attn_f16_q1.at(dk_dv);
+        } else {
+            kernel = backend_ctx->kernels_flash_attn_f32_q1.at(dk_dv);
+        }
+    } else {
+        if (is_mixed) {
+            kernel = backend_ctx->kernels_flash_attn_f32_f16.at(dk_dv);
+        } else if (is_f16) {
+            kernel = backend_ctx->kernels_flash_attn_f16.at(dk_dv);
+        } else {
+            kernel = backend_ctx->kernels_flash_attn_f32.at(dk_dv);
+        }
+    }
+    GGML_ASSERT(kernel != NULL);
+
+    ggml_tensor_extra_cl * extra_q = (ggml_tensor_extra_cl *)q->extra;
+    ggml_tensor_extra_cl * extra_k = (ggml_tensor_extra_cl *)k->extra;
+    ggml_tensor_extra_cl * extra_v = (ggml_tensor_extra_cl *)v->extra;
+    ggml_tensor_extra_cl * extra_o = (ggml_tensor_extra_cl *)dst->extra;
+    ggml_tensor_extra_cl * extra_mask = mask ? (ggml_tensor_extra_cl *)mask->extra : NULL;
+
+    cl_ulong offset_q = extra_q->offset + q->view_offs;
+    cl_ulong offset_k = extra_k->offset + k->view_offs;
+    cl_ulong offset_v = extra_v->offset + v->view_offs;
+    cl_ulong offset_o = extra_o->offset + dst->view_offs;
+    cl_mem   mask_buffer = extra_mask ? extra_mask->data_device : NULL;
+    cl_ulong offset_mask = extra_mask ? extra_mask->offset + mask->view_offs : 0;
+
+    const cl_ulong q_nb1 = q->nb[1], q_nb2 = q->nb[2], q_nb3 = q->nb[3];
+    const cl_ulong k_nb1 = k->nb[1], k_nb2 = k->nb[2], k_nb3 = k->nb[3];
+    const cl_ulong v_nb1 = v->nb[1], v_nb2 = v->nb[2], v_nb3 = v->nb[3];
+    const cl_ulong o_nb1 = dst->nb[1], o_nb2 = dst->nb[2], o_nb3 = dst->nb[3];
+    const cl_ulong mask_nb1 = mask ? mask->nb[1] : 0;
+    const cl_ulong mask_nb2 = mask ? mask->nb[2] : 0;
+    const cl_ulong mask_nb3 = mask ? mask->nb[3] : 0;
+    const int mask_ne2 = mask ? mask->ne[2] : 0;
+    const int mask_ne3 = mask ? mask->ne[3] : 0;
+
+    float scale, max_bias, logit_softcap;
+    const float * params = (const float *)dst->op_params;
+    scale         = params[0];
+    max_bias      = params[1];
+    logit_softcap = params[2];
+
+    const int is_causal = (mask == NULL && n_q > 1 && n_q == n_kv);
+
+    const int n_head_log2_val = n_head > 0 ? 1u << (int)floorf(log2f((float)n_head)) : 0;
+    const float n_head_log2_f = n_head_log2_val > 0 ? (float)n_head_log2_val : 1.0f;
+    const float m0 = powf(2.0f, -(max_bias) / n_head_log2_f);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2_f);
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra_q->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset_q));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra_k->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset_k));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extra_v->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offset_v));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_mem),   &extra_o->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong), &offset_o));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(float),    &scale));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),      &n_q));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),     &n_kv));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),     &is_causal));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),     &n_head));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &q_nb1)); CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &q_nb2)); CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &q_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &k_nb1)); CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &k_nb2)); CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &k_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &v_nb1)); CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &v_nb2)); CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &v_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong), &o_nb1)); CL_CHECK(clSetKernelArg(kernel, 23, sizeof(cl_ulong), &o_nb2)); CL_CHECK(clSetKernelArg(kernel, 24, sizeof(cl_ulong), &o_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 25, sizeof(float),    &max_bias));
+    CL_CHECK(clSetKernelArg(kernel, 26, sizeof(float),    &m0));
+    CL_CHECK(clSetKernelArg(kernel, 27, sizeof(float),    &m1));
+    CL_CHECK(clSetKernelArg(kernel, 28, sizeof(int),      &n_head_log2_val));
+    CL_CHECK(clSetKernelArg(kernel, 29, sizeof(float),    &logit_softcap));
+    CL_CHECK(clSetKernelArg(kernel, 30, sizeof(int),      &n_head_kv));
+    CL_CHECK(clSetKernelArg(kernel, 31, sizeof(cl_mem),   &mask_buffer));
+    CL_CHECK(clSetKernelArg(kernel, 32, sizeof(cl_ulong), &offset_mask));
+    CL_CHECK(clSetKernelArg(kernel, 33, sizeof(cl_ulong), &mask_nb1));
+    CL_CHECK(clSetKernelArg(kernel, 34, sizeof(cl_ulong), &mask_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 35, sizeof(cl_ulong), &mask_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 36, sizeof(int),      &mask_ne2));
+    CL_CHECK(clSetKernelArg(kernel, 37, sizeof(int),      &mask_ne3));
+
+    if (n_q == 1) {
+        const size_t wg_size = 64;
+        size_t local_work_size[] = { wg_size, 1 };
+        size_t global_work_size[] = { wg_size, (size_t)(n_head * n_batch) };
+        backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size, local_work_size, dst);
+    } else {
+        const int block_m = backend_ctx->kernels_flash_attn_bm.at(dk_dv);
+        const size_t wg_size = block_m;
+        size_t local_work_size[] = { wg_size, 1 };
+        size_t global_work_size[] = { (size_t)((n_q + block_m - 1) / block_m) * wg_size, (size_t)(n_head * n_batch) };
+        backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size, local_work_size, dst);
+    }
+}
+
 static void ggml_cl_mul_mat_f16_f32_tiled(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
@@ -7607,6 +7849,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_sum_rows;
             break;
+        case GGML_OP_FLASH_ATTN_EXT:
+            if (!any_on_device) {
+                return false;
+            }
+            ggml_cl_flash_attn(backend, tensor->src[0], tensor->src[1], tensor);
+            return true;
         default:
             return false;
     }
diff --git a/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl b/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl
new file mode 100644
index 00000000000..fea06867e10
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl
@@ -0,0 +1,343 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define ACC_TYPE float
+#define ACC_TYPE4 float4
+#define DATA_TYPE half
+#define DATA_TYPE4 half4
+#define CONVERT_ACC4(x) convert_float4(x)
+#define CONVERT_DATA4(x) convert_half4(x)
+
+#define DK_VEC (DK/4)
+#define DV_VEC (DV/4)
+#define WG_SIZE (BLOCK_M)
+#define Q1_WG_SIZE 64
+
+inline float get_alibi_slope(
+    const float max_bias, const uint h, const uint n_head_log2, const float m0, const float m1
+) {
+    if (max_bias <= 0.0f) {
+        return 1.0f;
+    }
+    const float base = h < n_head_log2 ? m0 : m1;
+    const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+    return pow(base, exph);
+}
+__kernel void flash_attn_f16(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3
+) {
+    const int tid = get_local_id(0);
+    const int block_q_idx = get_group_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int my_query_row = block_q_idx * BLOCK_M + tid;
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    if (my_query_row < n_q) {
+        const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2 + my_query_row * q_nb1;
+        const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset);
+        #pragma unroll
+        for (int i = 0; i < DK_VEC; ++i) {
+            q_priv[i] = CONVERT_ACC4(q_ptr[i]);
+        }
+    }
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) {
+        o_acc[i] = (ACC_TYPE4)(0.0f);
+    }
+    ACC_TYPE m_i = -INFINITY;
+    ACC_TYPE l_i = 0.0f;
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    __local DATA_TYPE4 l_k[BLOCK_N][DK_VEC];
+    __local DATA_TYPE4 l_v[BLOCK_N][DV_VEC];
+
+    for (int k_start = 0; k_start < n_kv; k_start += BLOCK_N) {
+        for (int i = tid; i < BLOCK_N * DK_VEC; i += WG_SIZE) {
+            const int row = i / DK_VEC;
+            const int col = i % DK_VEC;
+            const int k_row_idx = k_start + row;
+            if (k_row_idx < n_kv) {
+                const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_row_idx * k_nb1;
+                l_k[row][col] = ((__global DATA_TYPE4*)(k_base + k_row_offset))[col];
+            }
+        }
+        for (int i = tid; i < BLOCK_N * DV_VEC; i += WG_SIZE) {
+            const int row = i / DV_VEC;
+            const int col = i % DV_VEC;
+            const int v_row_idx = k_start + row;
+            if (v_row_idx < n_kv) {
+                const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + v_row_idx * v_nb1;
+                l_v[row][col] = ((__global DATA_TYPE4*)(v_base + v_row_offset))[col];
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        if (my_query_row >= n_q) {
+            continue;
+        }
+
+        for (int j = 0; j < BLOCK_N; j += 2) {
+            const int k_row0 = k_start + j;
+            const int k_row1 = k_start + j + 1;
+
+            ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f);
+            ACC_TYPE4 dot_acc1 = (ACC_TYPE4)(0.0f);
+            #pragma unroll
+            for (int k = 0; k < DK_VEC; k++) {
+                dot_acc0 = mad(q_priv[k], CONVERT_ACC4(l_k[j][k]), dot_acc0);
+                dot_acc1 = mad(q_priv[k], CONVERT_ACC4(l_k[j+1][k]), dot_acc1);
+            }
+            ACC_TYPE score0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale;
+            ACC_TYPE score1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale;
+
+            if (is_causal) {
+                if (k_row0 > (n_kv - n_q + my_query_row)) score0 = -INFINITY;
+                if (k_row1 > (n_kv - n_q + my_query_row)) score1 = -INFINITY;
+            }
+
+            if (k_row0 >= n_kv) score0 = -INFINITY;
+            if (k_row1 >= n_kv) score1 = -INFINITY;
+
+            if (mask_base != NULL) {
+                const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base + my_query_row * mask_nb1);
+                if (k_row0 < n_kv) score0 += slope * (ACC_TYPE)mask_ptr[k_row0];
+                if (k_row1 < n_kv) score1 += slope * (ACC_TYPE)mask_ptr[k_row1];
+            }
+
+            if (logit_softcap > 0.0f) {
+                score0 = logit_softcap * tanh(score0 / logit_softcap);
+                score1 = logit_softcap * tanh(score1 / logit_softcap);
+            }
+
+            const ACC_TYPE m_new = max(m_i, max(score0, score1));
+            const ACC_TYPE p0 = exp(score0 - m_new);
+            const ACC_TYPE p1 = exp(score1 - m_new);
+            const ACC_TYPE scale_prev = exp(m_i - m_new);
+
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_acc[i] = o_acc[i] * scale_prev + p0 * CONVERT_ACC4(l_v[j][i]) + p1 * CONVERT_ACC4(l_v[j+1][i]);
+            }
+            l_i = l_i * scale_prev + p0 + p1;
+            m_i = m_new;
+        }
+    }
+
+    if (my_query_row < n_q) {
+        const ulong o_row_offset = batch_idx * o_nb3 + my_query_row * o_nb2 + head_idx * o_nb1;
+        global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
+        if (l_i > 0.0f) {
+            const ACC_TYPE l_inv = 1.0f / l_i;
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = CONVERT_DATA4(o_acc[i] * l_inv);
+            }
+        } else {
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = (DATA_TYPE4)(0.0f);
+            }
+        }
+    }
+}
+
+__kernel void flash_attn_f16_q1(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3
+) {
+    const int tid = get_local_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2;
+    const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset);
+    #pragma unroll
+    for (int i = 0; i < DK_VEC; ++i) {
+        q_priv[i] = CONVERT_ACC4(q_ptr[i]);
+    }
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    ACC_TYPE m_i = -INFINITY;
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        m_i = max(m_i, score);
+    }
+
+    __local ACC_TYPE local_m[Q1_WG_SIZE];
+    local_m[tid] = m_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_m[tid] = max(local_m[tid], local_m[tid + s]);
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    const ACC_TYPE m_final = local_m[0];
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) o_acc[i] = (ACC_TYPE4)(0.0f);
+    ACC_TYPE l_i = 0.0f;
+
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + k_idx * v_nb1;
+        const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
+        const global DATA_TYPE4* v_ptr = (const global DATA_TYPE4*)(v_base + v_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        const ACC_TYPE p = exp(score - m_final);
+        l_i += p;
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; i++) {
+            o_acc[i] = mad(p, CONVERT_ACC4(v_ptr[i]), o_acc[i]);
+        }
+    }
+
+    __local ACC_TYPE local_l[Q1_WG_SIZE];
+    __local ACC_TYPE4 local_o_comp[Q1_WG_SIZE];
+    local_l[tid] = l_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_l[tid] += local_l[tid + s];
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const ulong o_row_offset = batch_idx * o_nb3 + head_idx * o_nb1;
+    global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
+    const ACC_TYPE l_final = local_l[0];
+
+    if (l_final > 0.0f) {
+        const ACC_TYPE l_inv = 1.0f / l_final;
+        for (int i = 0; i < DV_VEC; i++) {
+            local_o_comp[tid] = o_acc[i];
+            barrier(CLK_LOCAL_MEM_FENCE);
+            #pragma unroll
+            for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+                if (tid < s) local_o_comp[tid] += local_o_comp[tid + s];
+                barrier(CLK_LOCAL_MEM_FENCE);
+            }
+            if (tid == 0) {
+                o_row[i] = CONVERT_DATA4(local_o_comp[0] * l_inv);
+            }
+        }
+    } else if (tid == 0) {
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; ++i) o_row[i] = (DATA_TYPE4)(0.0f);
+    }
+}
diff --git a/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl b/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl
new file mode 100644
index 00000000000..2d657327d64
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl
@@ -0,0 +1,343 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define ACC_TYPE float
+#define ACC_TYPE4 float4
+#define DATA_TYPE float
+#define DATA_TYPE4 float4
+#define CONVERT_ACC4(x) (x)
+#define CONVERT_DATA4(x) (x)
+
+#define DK_VEC (DK/4)
+#define DV_VEC (DV/4)
+#define WG_SIZE (BLOCK_M)
+#define Q1_WG_SIZE 64
+
+inline float get_alibi_slope(
+    const float max_bias, const uint h, const uint n_head_log2, const float m0, const float m1
+) {
+    if (max_bias <= 0.0f) {
+        return 1.0f;
+    }
+    const float base = h < n_head_log2 ? m0 : m1;
+    const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+    return pow(base, exph);
+}
+__kernel void flash_attn_f32(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3
+) {
+    const int tid = get_local_id(0);
+    const int block_q_idx = get_group_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int my_query_row = block_q_idx * BLOCK_M + tid;
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    if (my_query_row < n_q) {
+        const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2 + my_query_row * q_nb1;
+        const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset);
+        #pragma unroll
+        for (int i = 0; i < DK_VEC; ++i) {
+            q_priv[i] = CONVERT_ACC4(q_ptr[i]);
+        }
+    }
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) {
+        o_acc[i] = (ACC_TYPE4)(0.0f);
+    }
+    ACC_TYPE m_i = -INFINITY;
+    ACC_TYPE l_i = 0.0f;
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    __local DATA_TYPE4 l_k[BLOCK_N][DK_VEC];
+    __local DATA_TYPE4 l_v[BLOCK_N][DV_VEC];
+
+    for (int k_start = 0; k_start < n_kv; k_start += BLOCK_N) {
+        for (int i = tid; i < BLOCK_N * DK_VEC; i += WG_SIZE) {
+            const int row = i / DK_VEC;
+            const int col = i % DK_VEC;
+            const int k_row_idx = k_start + row;
+            if (k_row_idx < n_kv) {
+                const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_row_idx * k_nb1;
+                l_k[row][col] = ((__global DATA_TYPE4*)(k_base + k_row_offset))[col];
+            }
+        }
+        for (int i = tid; i < BLOCK_N * DV_VEC; i += WG_SIZE) {
+            const int row = i / DV_VEC;
+            const int col = i % DV_VEC;
+            const int v_row_idx = k_start + row;
+            if (v_row_idx < n_kv) {
+                const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + v_row_idx * v_nb1;
+                l_v[row][col] = ((__global DATA_TYPE4*)(v_base + v_row_offset))[col];
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        if (my_query_row >= n_q) {
+            continue;
+        }
+
+        for (int j = 0; j < BLOCK_N; j += 2) {
+            const int k_row0 = k_start + j;
+            const int k_row1 = k_start + j + 1;
+
+            ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f);
+            ACC_TYPE4 dot_acc1 = (ACC_TYPE4)(0.0f);
+            #pragma unroll
+            for (int k = 0; k < DK_VEC; k++) {
+                dot_acc0 = mad(q_priv[k], CONVERT_ACC4(l_k[j][k]), dot_acc0);
+                dot_acc1 = mad(q_priv[k], CONVERT_ACC4(l_k[j+1][k]), dot_acc1);
+            }
+            ACC_TYPE score0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale;
+            ACC_TYPE score1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale;
+
+            if (is_causal) {
+                if (k_row0 > (n_kv - n_q + my_query_row)) score0 = -INFINITY;
+                if (k_row1 > (n_kv - n_q + my_query_row)) score1 = -INFINITY;
+            }
+
+            if (k_row0 >= n_kv) score0 = -INFINITY;
+            if (k_row1 >= n_kv) score1 = -INFINITY;
+
+            if (mask_base != NULL) {
+                const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base + my_query_row * mask_nb1);
+                if (k_row0 < n_kv) score0 += slope * (ACC_TYPE)mask_ptr[k_row0];
+                if (k_row1 < n_kv) score1 += slope * (ACC_TYPE)mask_ptr[k_row1];
+            }
+
+            if (logit_softcap > 0.0f) {
+                score0 = logit_softcap * tanh(score0 / logit_softcap);
+                score1 = logit_softcap * tanh(score1 / logit_softcap);
+            }
+
+            const ACC_TYPE m_new = max(m_i, max(score0, score1));
+            const ACC_TYPE p0 = exp(score0 - m_new);
+            const ACC_TYPE p1 = exp(score1 - m_new);
+            const ACC_TYPE scale_prev = exp(m_i - m_new);
+
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_acc[i] = o_acc[i] * scale_prev + p0 * CONVERT_ACC4(l_v[j][i]) + p1 * CONVERT_ACC4(l_v[j+1][i]);
+            }
+            l_i = l_i * scale_prev + p0 + p1;
+            m_i = m_new;
+        }
+    }
+
+    if (my_query_row < n_q) {
+        const ulong o_row_offset = batch_idx * o_nb3 + my_query_row * o_nb2 + head_idx * o_nb1;
+        global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
+        if (l_i > 0.0f) {
+            const ACC_TYPE l_inv = 1.0f / l_i;
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = CONVERT_DATA4(o_acc[i] * l_inv);
+            }
+        } else {
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = (DATA_TYPE4)(0.0f);
+            }
+        }
+    }
+}
+
+__kernel void flash_attn_f32_q1(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3
+) {
+    const int tid = get_local_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2;
+    const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset);
+    #pragma unroll
+    for (int i = 0; i < DK_VEC; ++i) {
+        q_priv[i] = CONVERT_ACC4(q_ptr[i]);
+    }
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    ACC_TYPE m_i = -INFINITY;
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        m_i = max(m_i, score);
+    }
+
+    __local ACC_TYPE local_m[Q1_WG_SIZE];
+    local_m[tid] = m_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_m[tid] = max(local_m[tid], local_m[tid + s]);
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    const ACC_TYPE m_final = local_m[0];
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) o_acc[i] = (ACC_TYPE4)(0.0f);
+    ACC_TYPE l_i = 0.0f;
+
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + k_idx * v_nb1;
+        const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
+        const global DATA_TYPE4* v_ptr = (const global DATA_TYPE4*)(v_base + v_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        const ACC_TYPE p = exp(score - m_final);
+        l_i += p;
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; i++) {
+            o_acc[i] = mad(p, CONVERT_ACC4(v_ptr[i]), o_acc[i]);
+        }
+    }
+
+    __local ACC_TYPE local_l[Q1_WG_SIZE];
+    __local ACC_TYPE4 local_o_comp[Q1_WG_SIZE];
+    local_l[tid] = l_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_l[tid] += local_l[tid + s];
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const ulong o_row_offset = batch_idx * o_nb3 + head_idx * o_nb1;
+    global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
+    const ACC_TYPE l_final = local_l[0];
+
+    if (l_final > 0.0f) {
+        const ACC_TYPE l_inv = 1.0f / l_final;
+        for (int i = 0; i < DV_VEC; i++) {
+            local_o_comp[tid] = o_acc[i];
+            barrier(CLK_LOCAL_MEM_FENCE);
+            #pragma unroll
+            for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+                if (tid < s) local_o_comp[tid] += local_o_comp[tid + s];
+                barrier(CLK_LOCAL_MEM_FENCE);
+            }
+            if (tid == 0) {
+                o_row[i] = CONVERT_DATA4(local_o_comp[0] * l_inv);
+            }
+        }
+    } else if (tid == 0) {
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; ++i) o_row[i] = (DATA_TYPE4)(0.0f);
+    }
+}
diff --git a/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl b/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl
new file mode 100644
index 00000000000..7067bd2591f
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl
@@ -0,0 +1,346 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define ACC_TYPE float
+#define ACC_TYPE4 float4
+#define Q_DATA_TYPE4 float4
+#define KV_DATA_TYPE4 half4
+#define O_DATA_TYPE4 float4
+#define MASK_DATA_TYPE half
+#define CONVERT_Q_ACC4(x) (x)
+#define CONVERT_KV_ACC4(x) convert_float4(x)
+#define CONVERT_O_DATA4(x) (x)
+
+#define DK_VEC (DK/4)
+#define DV_VEC (DV/4)
+#define WG_SIZE (BLOCK_M)
+#define Q1_WG_SIZE 64
+
+inline float get_alibi_slope(
+    const float max_bias, const uint h, const uint n_head_log2, const float m0, const float m1
+) {
+    if (max_bias <= 0.0f) {
+        return 1.0f;
+    }
+    const float base = h < n_head_log2 ? m0 : m1;
+    const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+    return pow(base, exph);
+}
+__kernel void flash_attn_f32_f16(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3
+) {
+    const int tid = get_local_id(0);
+    const int block_q_idx = get_group_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int my_query_row = block_q_idx * BLOCK_M + tid;
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    if (my_query_row < n_q) {
+        const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2 + my_query_row * q_nb1;
+        const global Q_DATA_TYPE4* q_ptr = (const global Q_DATA_TYPE4*)(q_base + q_row_offset);
+        #pragma unroll
+        for (int i = 0; i < DK_VEC; ++i) {
+            q_priv[i] = CONVERT_Q_ACC4(q_ptr[i]);
+        }
+    }
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) {
+        o_acc[i] = (ACC_TYPE4)(0.0f);
+    }
+    ACC_TYPE m_i = -INFINITY;
+    ACC_TYPE l_i = 0.0f;
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    __local KV_DATA_TYPE4 l_k[BLOCK_N][DK_VEC];
+    __local KV_DATA_TYPE4 l_v[BLOCK_N][DV_VEC];
+
+    for (int k_start = 0; k_start < n_kv; k_start += BLOCK_N) {
+        for (int i = tid; i < BLOCK_N * DK_VEC; i += WG_SIZE) {
+            const int row = i / DK_VEC;
+            const int col = i % DK_VEC;
+            const int k_row_idx = k_start + row;
+            if (k_row_idx < n_kv) {
+                const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_row_idx * k_nb1;
+                l_k[row][col] = ((__global KV_DATA_TYPE4*)(k_base + k_row_offset))[col];
+            }
+        }
+        for (int i = tid; i < BLOCK_N * DV_VEC; i += WG_SIZE) {
+            const int row = i / DV_VEC;
+            const int col = i % DV_VEC;
+            const int v_row_idx = k_start + row;
+            if (v_row_idx < n_kv) {
+                const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + v_row_idx * v_nb1;
+                l_v[row][col] = ((__global KV_DATA_TYPE4*)(v_base + v_row_offset))[col];
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        if (my_query_row >= n_q) {
+            continue;
+        }
+
+        for (int j = 0; j < BLOCK_N; j += 2) {
+            const int k_row0 = k_start + j;
+            const int k_row1 = k_start + j + 1;
+
+            ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f);
+            ACC_TYPE4 dot_acc1 = (ACC_TYPE4)(0.0f);
+            #pragma unroll
+            for (int k = 0; k < DK_VEC; k++) {
+                dot_acc0 = mad(q_priv[k], CONVERT_KV_ACC4(l_k[j][k]), dot_acc0);
+                dot_acc1 = mad(q_priv[k], CONVERT_KV_ACC4(l_k[j+1][k]), dot_acc1);
+            }
+            ACC_TYPE score0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale;
+            ACC_TYPE score1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale;
+
+            if (is_causal) {
+                if (k_row0 > (n_kv - n_q + my_query_row)) score0 = -INFINITY;
+                if (k_row1 > (n_kv - n_q + my_query_row)) score1 = -INFINITY;
+            }
+
+            if (k_row0 >= n_kv) score0 = -INFINITY;
+            if (k_row1 >= n_kv) score1 = -INFINITY;
+
+            if (mask_base != NULL) {
+                const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base + my_query_row * mask_nb1);
+                if (k_row0 < n_kv) score0 += slope * (ACC_TYPE)mask_ptr[k_row0];
+                if (k_row1 < n_kv) score1 += slope * (ACC_TYPE)mask_ptr[k_row1];
+            }
+
+            if (logit_softcap > 0.0f) {
+                score0 = logit_softcap * tanh(score0 / logit_softcap);
+                score1 = logit_softcap * tanh(score1 / logit_softcap);
+            }
+
+            const ACC_TYPE m_new = max(m_i, max(score0, score1));
+            const ACC_TYPE p0 = exp(score0 - m_new);
+            const ACC_TYPE p1 = exp(score1 - m_new);
+            const ACC_TYPE scale_prev = exp(m_i - m_new);
+
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_acc[i] = o_acc[i] * scale_prev + p0 * CONVERT_KV_ACC4(l_v[j][i]) + p1 * CONVERT_KV_ACC4(l_v[j+1][i]);
+            }
+            l_i = l_i * scale_prev + p0 + p1;
+            m_i = m_new;
+        }
+    }
+
+    if (my_query_row < n_q) {
+        const ulong o_row_offset = batch_idx * o_nb3 + my_query_row * o_nb2 + head_idx * o_nb1;
+        global O_DATA_TYPE4 *o_row = (global O_DATA_TYPE4 *)(o_base + o_row_offset);
+        if (l_i > 0.0f) {
+            const ACC_TYPE l_inv = 1.0f / l_i;
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = CONVERT_O_DATA4(o_acc[i] * l_inv);
+            }
+        } else {
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = (O_DATA_TYPE4)(0.0f);
+            }
+        }
+    }
+}
+
+__kernel void flash_attn_f32_f16_q1(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3
+) {
+    const int tid = get_local_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2;
+    const global Q_DATA_TYPE4* q_ptr = (const global Q_DATA_TYPE4*)(q_base + q_row_offset);
+    #pragma unroll
+    for (int i = 0; i < DK_VEC; ++i) {
+        q_priv[i] = CONVERT_Q_ACC4(q_ptr[i]);
+    }
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    ACC_TYPE m_i = -INFINITY;
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const global KV_DATA_TYPE4* k_ptr = (const global KV_DATA_TYPE4*)(k_base + k_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_KV_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        m_i = max(m_i, score);
+    }
+
+    __local ACC_TYPE local_m[Q1_WG_SIZE];
+    local_m[tid] = m_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_m[tid] = max(local_m[tid], local_m[tid + s]);
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    const ACC_TYPE m_final = local_m[0];
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) o_acc[i] = (ACC_TYPE4)(0.0f);
+    ACC_TYPE l_i = 0.0f;
+
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + k_idx * v_nb1;
+        const global KV_DATA_TYPE4* k_ptr = (const global KV_DATA_TYPE4*)(k_base + k_row_offset);
+        const global KV_DATA_TYPE4* v_ptr = (const global KV_DATA_TYPE4*)(v_base + v_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_KV_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        const ACC_TYPE p = exp(score - m_final);
+        l_i += p;
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; i++) {
+            o_acc[i] = mad(p, CONVERT_KV_ACC4(v_ptr[i]), o_acc[i]);
+        }
+    }
+
+    __local ACC_TYPE local_l[Q1_WG_SIZE];
+    __local ACC_TYPE4 local_o_comp[Q1_WG_SIZE];
+    local_l[tid] = l_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_l[tid] += local_l[tid + s];
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const ulong o_row_offset = batch_idx * o_nb3 + head_idx * o_nb1;
+    global O_DATA_TYPE4 *o_row = (global O_DATA_TYPE4 *)(o_base + o_row_offset);
+    const ACC_TYPE l_final = local_l[0];
+
+    if (l_final > 0.0f) {
+        const ACC_TYPE l_inv = 1.0f / l_final;
+        for (int i = 0; i < DV_VEC; i++) {
+            local_o_comp[tid] = o_acc[i];
+            barrier(CLK_LOCAL_MEM_FENCE);
+            #pragma unroll
+            for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+                if (tid < s) local_o_comp[tid] += local_o_comp[tid + s];
+                barrier(CLK_LOCAL_MEM_FENCE);
+            }
+            if (tid == 0) {
+                o_row[i] = CONVERT_O_DATA4(local_o_comp[0] * l_inv);
+            }
+        }
+    } else if (tid == 0) {
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; ++i) o_row[i] = (O_DATA_TYPE4)(0.0f);
+    }
+}

From 812ccabe5e21d0227e2498d68192bf5583461c12 Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Sat, 16 Aug 2025 03:58:38 -0500
Subject: [PATCH 74/86] vulkan: Add missing bounds checking to scalar/coopmat1
 mul_mat_id (llama/15334)

---
 ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
index 8c5114a79d2..a61a464c7be 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
@@ -801,7 +801,7 @@ void main() {
             }
 #else
             const uint row_i = ic * BN + loadc_b + l;
-            if (row_i < _ne1) {
+            if (row_i < _ne1 && block + loadr_b < end_k) {
                 const u16vec2 row_idx = row_ids[row_i];
                 buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = TO_FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]);
             } else {
@@ -875,7 +875,9 @@ void main() {
 
                 const u16vec2 row_idx = row_ids[row_i];
 
-                data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
+                if (dr + cm_row * TM + store_r < p.M) {
+                    data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
+                }
             }
         }
     }
@@ -925,7 +927,9 @@ void main() {
 #endif // MUL_MAT_ID
                 [[unroll]] for (uint cr = 0; cr < TM; cr++) {
 #ifdef MUL_MAT_ID
-                    data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
+                    if (dr_warp + cr < p.M) {
+                        data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
+                    }
 #else
                     if (dr_warp + cr < p.M && dc_warp + cc < p.N) {
                         data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);

From b759bbcd6b08153fd330be97dcb7c1c92ae786fc Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Sat, 16 Aug 2025 04:18:31 -0500
Subject: [PATCH 75/86] vulkan: Support mul_mat_id with f32 accumulators
 (llama/15337)

* vulkan: Add missing bounds checking to scalar/coopmat1 mul_mat_id

* vulkan: Support mul_mat_id with f32 accumulators, but they are not hooked up

- There's no explicit way to request f32 precision for mul_mat_id, but there
probably should be, and this gets the code in place for that.
- A couple fixes to check_results.
- Remove casts to fp16 in coopmat1 FA shader (found by inspection).
---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp          | 164 ++++++++----------
 .../vulkan-shaders/flash_attn_cm1.comp        |   8 +-
 2 files changed, 81 insertions(+), 91 deletions(-)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 76a0cfa4f02..63155849966 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -2387,26 +2387,26 @@ static void ggml_vk_load_shaders(vk_device& device) {
             CREATE_MM(pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4)
         }
 #endif
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f16acc,   matmul_id_iq1_s_f16,   , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f16acc,   matmul_id_iq1_m_f16,   , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc,  matmul_id_iq2_xs_f16,  , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc,   matmul_id_iq2_s_f16,   , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f16acc, matmul_id_iq3_xxs_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc,   matmul_id_iq3_s_f16,   , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc,  matmul_id_iq4_xs_f16,  , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc,  matmul_id_iq4_nl_f16,  , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f16acc,   matmul_id_mxfp4_f16,   , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S],   matmul_id_iq1_s_f16,   mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M],   matmul_id_iq1_m_f16,   mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_iq2_xxs_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS],  matmul_id_iq2_xs_f16,  mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S],   matmul_id_iq2_s_f16,   mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_iq3_xxs_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S],   matmul_id_iq3_s_f16,   mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS],  matmul_id_iq4_xs_f16,  mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL],  matmul_id_iq4_nl_f16,  mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4],   matmul_id_mxfp4_f16,   mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
 #undef CREATE_MM
 #undef CREATE_MM2
     } else
@@ -2502,51 +2502,27 @@ static void ggml_vk_load_shaders(vk_device& device) {
         }
 #endif
 
-        if (device->coopmat_acc_f16_support) {
-            CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-
-            CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ1_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f16acc,   matmul_id_iq1_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ1_M,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f16acc,   matmul_id_iq1_m_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ2_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc,  matmul_id_iq2_xs_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ2_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc,   matmul_id_iq2_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f16acc, matmul_id_iq3_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc,   matmul_id_iq3_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc,  matmul_id_iq4_xs_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc,  matmul_id_iq4_nl_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f16acc,   matmul_id_mxfp4_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        } else {
-            CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-
-            CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ1_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f16acc,   matmul_id_iq1_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ1_M,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f16acc,   matmul_id_iq1_m_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ2_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc,  matmul_id_iq2_xs_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ2_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc,   matmul_id_iq2_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f16acc, matmul_id_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc,   matmul_id_iq3_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc,  matmul_id_iq4_xs_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc,  matmul_id_iq4_nl_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f16acc,   matmul_id_mxfp4_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        }
+        CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+
+        CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ1_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S],   matmul_id_iq1_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ1_M,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M],   matmul_id_iq1_m_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_iq2_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ2_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS],  matmul_id_iq2_xs_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ2_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S],   matmul_id_iq2_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_iq3_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S],   matmul_id_iq3_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS],  matmul_id_iq4_xs_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL],  matmul_id_iq4_nl_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4],   matmul_id_mxfp4_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
 #undef CREATE_MM2
 #undef CREATE_MM
     } else
@@ -2631,27 +2607,27 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
         CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id);
 
-        CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-
-        CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ1_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f16acc,   matmul_id_iq1_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ1_M,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f16acc,   matmul_id_iq1_m_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ2_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc,  matmul_id_iq2_xs_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ2_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc,   matmul_id_iq2_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f16acc, matmul_id_iq3_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc,   matmul_id_iq3_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc,  matmul_id_iq4_xs_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc,  matmul_id_iq4_nl_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f16acc,   matmul_id_mxfp4_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+
+        CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ1_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S],   matmul_id_iq1_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ1_M,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M],   matmul_id_iq1_m_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_iq2_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ2_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS],  matmul_id_iq2_xs_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ2_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S],   matmul_id_iq2_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_iq3_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S],   matmul_id_iq3_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS],  matmul_id_iq4_xs_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL],  matmul_id_iq4_nl_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4],   matmul_id_mxfp4_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
 #undef CREATE_MM2
 #undef CREATE_MMQ
 #undef CREATE_MM
@@ -4470,7 +4446,17 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
             return nullptr;
     }
 
-    return ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f32acc;
+    // XXX TODO 'prec' is not actually allowed in mul_mat_id.
+    bool prefer_fp16acc = ctx->device->fp16 /*&& prec == GGML_PREC_DEFAULT*/;
+    bool support_fp16acc = ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f16acc != nullptr;
+    bool support_fp32acc = ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f32acc != nullptr;
+
+    if (support_fp16acc && (prefer_fp16acc || !support_fp32acc)) {
+        return ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f16acc;
+    } else {
+        GGML_ASSERT(support_fp32acc);
+        return ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f32acc;
+    }
 }
 
 static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) {
@@ -11723,6 +11709,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         } else {
             tensor_clone = ggml_glu_split(ggml_ctx, src_clone[0], src_clone[1], (ggml_glu_op) tensor->op_params[0]);
         }
+        ggml_set_op_params_i32(tensor_clone, 2, ggml_get_op_params_i32(tensor, 2));
+        ggml_set_op_params_i32(tensor_clone, 3, ggml_get_op_params_i32(tensor, 3));
     } else if (tensor->op == GGML_OP_CPY || tensor->op == GGML_OP_DUP) {
         if (src1 == nullptr) {
             tensor_clone = ggml_dup(ggml_ctx, src_clone[0]);
@@ -11807,6 +11795,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         src_clone[0]->flags = src0->flags;
         tensor_clone = ggml_opt_step_sgd(ggml_ctx, src_clone[0], src_clone[1],
         src_clone[2]);
+    } else if (tensor->op == GGML_OP_ADD_ID) {
+        tensor_clone = ggml_add_id(ggml_ctx, src_clone[0], src_clone[1], src_clone[2]);
     }
     else {
         std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
index 230e815f22c..81cc3f81fce 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
@@ -210,7 +210,7 @@ void main() {
 
         [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
             [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-                Of[r][d] = float16_t(eMf[r]) * Of[r][d];
+                Of[r][d] = ACC_TYPE(eMf[r]) * Of[r][d];
             }
         }
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
@@ -233,7 +233,7 @@ void main() {
                 vec4 Vf = vec4(data_vv4[v_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * v_stride / 4 + d * D_split + d_tid]);
 #endif
                 [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-                    Of[r][d] += float16_t(Pf[r]) * ACC_TYPEV4(Vf);
+                    Of[r][d] += ACC_TYPE(Pf[r]) * ACC_TYPEV4(Vf);
                 }
             }
         }
@@ -288,7 +288,7 @@ void main() {
     [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
         [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
 
-            Of[r][d] = float16_t(eMf[r]) * Of[r][d];
+            Of[r][d] = ACC_TYPE(eMf[r]) * Of[r][d];
             tmpshv4[tid] = Of[r][d];
 
             barrier();
@@ -357,7 +357,7 @@ void main() {
 
     [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-            Of[r][d] *= float16_t(Lfrcp[r]);
+            Of[r][d] *= ACC_TYPE(Lfrcp[r]);
         }
     }
 

From 3bb4185ed2f5c2ca4c48f60cbe7625bfad3e3842 Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Sat, 16 Aug 2025 11:48:22 -0500
Subject: [PATCH 76/86] vulkan: fuse adds (llama/15252)

* vulkan: fuse adds

Fuse adds that have the same shape, which are common in MoE models.
It will currently fuse up to 6 adds, because we assume no more than
8 descriptors per dispatch. But this could be changed.

* check runtimeDescriptorArray feature

* disable multi_add for Intel due to likely driver bug
---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp          | 207 +++++++++++++++++-
 .../vulkan-shaders/generic_binary_head.comp   |  19 +-
 .../ggml-vulkan/vulkan-shaders/multi_add.comp |  68 ++++++
 .../src/ggml-vulkan/vulkan-shaders/utils.comp |  25 +++
 .../vulkan-shaders/vulkan-shaders-gen.cpp     |   2 +
 5 files changed, 298 insertions(+), 23 deletions(-)
 create mode 100644 ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp
 create mode 100644 ggml/src/ggml-vulkan/vulkan-shaders/utils.comp

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 63155849966..20674bee588 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -103,6 +103,8 @@ static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; }
 struct ggml_backend_vk_context;
 
 #define MAX_PARAMETER_COUNT 8
+// Max number of adds that can be fused without exceeding MAX_PARAMETER_COUNT.
+#define MAX_FUSED_ADDS (MAX_PARAMETER_COUNT - 2)
 
 struct vk_pipeline_struct {
     std::string name;
@@ -368,6 +370,7 @@ struct vk_device_struct {
     bool float_controls_rte_fp16;
     bool subgroup_add;
     bool subgroup_shuffle;
+    bool multi_add;
 
     bool integer_dot_product;
 
@@ -449,6 +452,9 @@ struct vk_device_struct {
     vk_pipeline pipeline_div[2][2][2];
     vk_pipeline pipeline_div_norepeat[2][2][2];
 
+    // indexed by num_additional_fused_ops == num_adds - 1
+    vk_pipeline pipeline_multi_add[MAX_FUSED_ADDS];
+
     vk_pipeline pipeline_add_id_f32;
 
     vk_pipeline pipeline_concat_f32, pipeline_concat_f16, pipeline_concat_i32;
@@ -801,6 +807,14 @@ struct vk_op_binary_push_constants {
     float param1; float param2; int32_t param3;
 };
 
+struct vk_op_multi_add_push_constants {
+    // shape for dst
+    uint32_t ne20; uint32_t ne21; uint32_t ne22; uint32_t ne23;
+
+    // strides for srcs+dst
+    uint32_t nb[8][4];
+};
+
 struct vk_op_add_id_push_constants {
     uint32_t ne0;
     uint32_t ne1;
@@ -2994,6 +3008,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
     CREATE_BINARY(div, _norepeat, {1})
 #undef CREATE_BINARY
 
+    if (device->multi_add) {
+        for (uint32_t i = 0; i < MAX_FUSED_ADDS; ++i) {
+            ggml_vk_create_pipeline(device, device->pipeline_multi_add[i], "multi_add_f32_" + std::to_string(i+1), multi_add_f32_len, multi_add_f32_data, "main", MAX_PARAMETER_COUNT, sizeof(vk_op_multi_add_push_constants), {512, 1, 1}, {i+2}, 1);
+        }
+    }
+
     ggml_vk_create_pipeline(device, device->pipeline_add_id_f32, "add_id_f32", add_id_f32_len, add_id_f32_data, "main", 4, sizeof(vk_op_add_id_push_constants), {1, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_acc_f32, "acc_f32", acc_f32_len, acc_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
@@ -3533,6 +3553,12 @@ static vk_device ggml_vk_get_device(size_t idx) {
 
         device->pipeline_robustness = pl_robustness_features.pipelineRobustness;
 
+        device->multi_add = vk12_props.shaderRoundingModeRTEFloat16 &&
+                            device->properties.limits.maxPushConstantsSize >= sizeof(vk_op_multi_add_push_constants) &&
+                            vk12_features.runtimeDescriptorArray &&
+                            device->vendor_id != VK_VENDOR_ID_INTEL &&
+                            getenv("GGML_VK_DISABLE_MULTI_ADD") == nullptr;
+
         if (device->subgroup_size_control) {
             device->subgroup_min_size = subgroup_size_control_props.minSubgroupSize;
             device->subgroup_max_size = subgroup_size_control_props.maxSubgroupSize;
@@ -6887,6 +6913,9 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         switch (op) {
         case GGML_OP_ADD:
         {
+            if (ctx->num_additional_fused_ops > 0) {
+                return ctx->device->pipeline_multi_add[ctx->num_additional_fused_ops];
+            }
             auto pipelines = ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_add_norepeat : ctx->device->pipeline_add;
             return pipelines[src0->type == GGML_TYPE_F16][src1->type == GGML_TYPE_F16][dst->type == GGML_TYPE_F16];
         }
@@ -7743,6 +7772,107 @@ static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const
     }, dryrun);
 }
 
+static void ggml_vk_multi_add(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_cgraph * cgraph, int node_idx, bool dryrun = false) {
+    const ggml_tensor *first_node = cgraph->nodes[node_idx];
+    const ggml_tensor *dst = cgraph->nodes[node_idx + ctx->num_additional_fused_ops];
+
+    // Make a list of all the tensors used by the op.
+    // Last element of the list is the dest tensor.
+    const ggml_tensor *tensors[MAX_PARAMETER_COUNT];
+    uint32_t num_srcs = ctx->num_additional_fused_ops + 2;
+    uint32_t num_tensors = num_srcs + 1;
+    GGML_ASSERT(num_tensors <= MAX_PARAMETER_COUNT);
+
+    tensors[0] = first_node->src[0];
+    tensors[1] = first_node->src[1];
+    for (int32_t i = 0; i < ctx->num_additional_fused_ops; ++i) {
+        // check whether the previous result is src[0] or src[1]
+        if (cgraph->nodes[node_idx + i] == cgraph->nodes[node_idx + i + 1]->src[0]) {
+            tensors[i+2] = cgraph->nodes[node_idx + i + 1]->src[1];
+        } else {
+            tensors[i+2] = cgraph->nodes[node_idx + i + 1]->src[0];
+        }
+    }
+    tensors[num_srcs] = dst;
+
+    vk_op_multi_add_push_constants pc;
+    pc.ne20 = (uint32_t)dst->ne[0];
+    pc.ne21 = (uint32_t)dst->ne[1];
+    pc.ne22 = (uint32_t)dst->ne[2];
+    pc.ne23 = (uint32_t)dst->ne[3];
+
+    for (uint32_t i = 0; i < num_tensors; ++i) {
+        const ggml_tensor *t = tensors[i];
+        pc.nb[i][0] = (uint32_t)t->nb[0] / sizeof(float);
+        pc.nb[i][1] = (uint32_t)t->nb[1] / sizeof(float);
+        pc.nb[i][2] = (uint32_t)t->nb[2] / sizeof(float);
+        pc.nb[i][3] = (uint32_t)t->nb[3] / sizeof(float);
+    }
+
+    vk_pipeline pipeline = ctx->device->pipeline_multi_add[ctx->num_additional_fused_ops];
+
+    if (pipeline == nullptr) {
+        std::cerr << "ggml_vulkan: Error: Missing multi_add";
+        GGML_ABORT("fatal error");
+    }
+
+    if (dryrun) {
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
+        return;
+    }
+
+    ggml_backend_vk_buffer_context * buf_ctx[MAX_PARAMETER_COUNT];
+    vk_buffer buf[MAX_PARAMETER_COUNT];
+    size_t offset[MAX_PARAMETER_COUNT];
+    bool uma[MAX_PARAMETER_COUNT];
+
+    for (uint32_t i = 0; i < num_tensors; ++i) {
+        buf_ctx[i] = (ggml_backend_vk_buffer_context *)tensors[i]->buffer->context;
+        buf[i] = nullptr;
+        offset[i] = 0;
+        uma[i] = false;
+
+        if (ctx->device->uma) {
+            ggml_vk_host_get(ctx->device, tensors[i]->data, buf[i], offset[i]);
+            uma[i] = buf[i] != nullptr;
+        }
+        if (!uma[i]) {
+            buf[i] = buf_ctx[i]->dev_buffer;
+            offset[i] = vk_tensor_offset(tensors[i]) + tensors[i]->view_offs;
+        }
+        GGML_ASSERT(buf[i] != nullptr);
+    }
+    // If any remaining descriptors are unused, just point them at src[0]
+    for (uint32_t i = num_tensors; i < MAX_PARAMETER_COUNT; ++i) {
+        buf[i] = buf[0];
+        offset[i] = 0;
+    }
+
+    std::array<uint32_t, 3> elements;
+
+    uint32_t ne = ggml_nelements(dst);
+    if (ne > 262144) {
+        elements = { 512, 512, CEIL_DIV(ne, 262144) };
+    } else if (ne > 512) {
+        elements = { 512, CEIL_DIV(ne, 512), 1 };
+    } else {
+        elements = { ne, 1, 1 };
+    }
+
+    ggml_vk_sync_buffers(subctx);
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
+        {
+            vk_subbuffer{ buf[0], offset[0], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[1], offset[1], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[2], offset[2], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[3], offset[3], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[4], offset[4], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[5], offset[5], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[6], offset[6], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[7], offset[7], VK_WHOLE_SIZE },
+        }, pc, elements);
+}
+
 static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
@@ -9703,8 +9833,11 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
 
         break;
     case GGML_OP_ADD:
-        ggml_vk_add(ctx, compute_ctx, src0, src1, node, dryrun);
-
+        if (ctx->num_additional_fused_ops) {
+            ggml_vk_multi_add(ctx, compute_ctx, cgraph, node_idx, dryrun);
+        } else {
+            ggml_vk_add(ctx, compute_ctx, src0, src1, node, dryrun);
+        }
         break;
     case GGML_OP_SUB:
         ggml_vk_sub(ctx, compute_ctx, src0, src1, node, dryrun);
@@ -10586,6 +10719,58 @@ static bool ggml_vk_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, st
     return true;
 }
 
+static uint32_t ggml_vk_fuse_multi_add(ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph, int node_idx) {
+
+    const ggml_tensor *first_node = cgraph->nodes[node_idx];
+    if (first_node->op != GGML_OP_ADD) {
+        return 0;
+    }
+
+    if (!ctx->device->multi_add) {
+        return 0;
+    }
+
+    int32_t num_adds = 1;
+    while (node_idx + num_adds < cgraph->n_nodes &&
+           cgraph->nodes[node_idx + num_adds]->op == GGML_OP_ADD &&
+           num_adds < MAX_FUSED_ADDS) {
+        num_adds++;
+    }
+
+    // The shader currently requires same shapes (but different strides are allowed),
+    // everything f32, and no misalignment
+    for (int32_t i = 0; i < num_adds; ++i) {
+        const ggml_tensor *next_node = cgraph->nodes[node_idx + i];
+        if (!ggml_are_same_shape(first_node, next_node->src[0]) ||
+            !ggml_are_same_shape(first_node, next_node->src[1]) ||
+            next_node->type != GGML_TYPE_F32 ||
+            next_node->src[0]->type != GGML_TYPE_F32 ||
+            next_node->src[1]->type != GGML_TYPE_F32 ||
+            get_misalign_bytes(ctx, next_node) ||
+            get_misalign_bytes(ctx, next_node->src[0]) ||
+            get_misalign_bytes(ctx, next_node->src[1])) {
+            num_adds = i;
+        }
+    }
+
+    // Verify we can fuse these
+    ggml_op adds[MAX_FUSED_ADDS];
+    for (int32_t i = 0; i < num_adds; ++i) {
+        adds[i] = GGML_OP_ADD;
+    }
+
+    // decrease num_adds if they can't all be fused
+    while (num_adds > 1 && !ggml_can_fuse(cgraph, node_idx, adds, num_adds)) {
+        num_adds--;
+    }
+
+    // a single add is not "fused", so just return zero
+    if (num_adds == 1) {
+        return 0;
+    }
+    return num_adds;
+}
+
 static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
     VK_LOG_DEBUG("ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)");
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
@@ -10599,8 +10784,13 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
 
     uint64_t total_mat_mul_bytes = 0;
     for (int i = 0; i < cgraph->n_nodes; i++) {
-        if (!ctx->device->disable_fusion && ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
-            ctx->num_additional_fused_ops = 1;
+        if (!ctx->device->disable_fusion) {
+            uint32_t num_adds = ggml_vk_fuse_multi_add(ctx, cgraph, i);
+            if (num_adds) {
+                ctx->num_additional_fused_ops = num_adds - 1;
+            } else if (ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
+                ctx->num_additional_fused_ops = 1;
+            }
         }
         ggml_vk_build_graph(ctx, cgraph, i, nullptr, 0, true, false, false, false);
         if (cgraph->nodes[i]->op == GGML_OP_MUL_MAT || cgraph->nodes[i]->op == GGML_OP_MUL_MAT_ID) {
@@ -10675,8 +10865,13 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
             mul_mat_bytes += ggml_nbytes(cgraph->nodes[i]->src[0]);
         }
 
-        if (!ctx->device->disable_fusion && ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
-            ctx->num_additional_fused_ops = 1;
+        if (!ctx->device->disable_fusion) {
+            uint32_t num_adds = ggml_vk_fuse_multi_add(ctx, cgraph, i);
+            if (num_adds) {
+                ctx->num_additional_fused_ops = num_adds - 1;
+            } else if (ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
+                ctx->num_additional_fused_ops = 1;
+            }
         }
 
         // Signal the almost_ready fence when the graph is mostly complete (< 20% remaining)
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp b/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp
index 4b4316cf3d9..750e7857539 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp
@@ -2,6 +2,7 @@
 #extension GL_EXT_control_flow_attributes : require
 
 #include "rte.comp"
+#include "utils.comp"
 
 layout (push_constant) uniform parameter
 {
@@ -28,25 +29,9 @@ uint get_aoffset() { return p.misalign_offsets >> 16; }
 uint get_boffset() { return (p.misalign_offsets >> 8) & 0xFF; }
 uint get_doffset() { return p.misalign_offsets & 0xFF; }
 
-// mod and div are expensive and coordinates/dimensions are often power of 2 or equal to 1
-uint fastmod(uint a, uint b) {
-    if ((b & (b-1)) == 0) {
-        return a & (b-1);
-    }
-    return a % b;
-}
-
-uint fastdiv(uint a, uint b) {
-    return (a < b) ? 0 : (a / b);
-}
 
 void get_indices(uint idx, out uint i00, out uint i01, out uint i02, out uint i03) {
-    i03 = fastdiv(idx, (p.ne02*p.ne01*p.ne00));
-    const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
-    i02 = fastdiv((idx - i03_offset), (p.ne01*p.ne00));
-    const uint i02_offset = i02*p.ne01*p.ne00;
-    i01 = (idx - i03_offset - i02_offset) / p.ne00;
-    i00 = idx - i03_offset - i02_offset - i01*p.ne00;
+    get_indices(idx, i00, i01, i02, i03, p.ne00, p.ne01, p.ne02, p.ne03);
 }
 
 uint src0_idx(uint i00, uint i01, uint i02, uint i03) {
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp b/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp
new file mode 100644
index 00000000000..0c7acb7060f
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp
@@ -0,0 +1,68 @@
+#version 450
+
+#extension GL_EXT_shader_16bit_storage : require
+#extension GL_EXT_nonuniform_qualifier : enable
+#extension GL_EXT_control_flow_attributes : require
+
+#include "rte.comp"
+#include "types.comp"
+#include "utils.comp"
+
+layout (push_constant) uniform parameter2
+{
+    // shape for dst
+    uint ne20; uint ne21; uint ne22; uint ne23;
+
+    // strides for srcs+dst
+    uint nb[8][4];
+} p;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];} a[];
+layout (binding = 0) writeonly buffer D {D_TYPE data_d[];} d[];
+
+layout(constant_id = 0) const uint num_srcs = 2;
+
+uint src_idx(uint s, uint i00, uint i01, uint i02, uint i03) {
+    return i03*p.nb[s][3] + i02*p.nb[s][2] + i01*p.nb[s][1] + i00*p.nb[s][0];
+}
+
+uint dst_idx(uint i00, uint i01, uint i02, uint i03) {
+    uint nb20 = p.nb[num_srcs][0];
+    uint nb21 = p.nb[num_srcs][1];
+    uint nb22 = p.nb[num_srcs][2];
+    uint nb23 = p.nb[num_srcs][3];
+    return i03*nb23 + i02*nb22 + i01*nb21 + i00*nb20;
+}
+
+uint get_idx() {
+    return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+}
+
+const uint num_threads = 256;
+
+layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+    uint idx = get_idx();
+
+    uint ne = p.ne20 * p.ne21 * p.ne22 * p.ne23;
+
+    // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
+    const uint num_iter = 2;
+
+    [[unroll]] for (uint i = 0; i < num_iter; ++i) {
+        if (idx >= ne) {
+            continue;
+        }
+        uint i00, i01, i02, i03;
+        get_indices(idx, i00, i01, i02, i03, p.ne20, p.ne21, p.ne22, p.ne23);
+
+        FLOAT_TYPE sum = FLOAT_TYPE(0);
+        [[unroll]] for (uint s = 0; s < num_srcs; ++s) {
+            sum += FLOAT_TYPE(a[s].data_a[src_idx(s, i00, i01, i02, i03)]);
+        }
+        d[num_srcs].data_d[dst_idx(i00, i01, i02, i03)] = D_TYPE(sum);
+
+        idx += num_threads;
+    }
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/utils.comp b/ggml/src/ggml-vulkan/vulkan-shaders/utils.comp
new file mode 100644
index 00000000000..dc4a1e6d96b
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/utils.comp
@@ -0,0 +1,25 @@
+#ifndef UTILS_COMP
+#define UTILS_COMP
+
+// mod and div are expensive and coordinates/dimensions are often power of 2 or equal to 1
+uint fastmod(uint a, uint b) {
+    if ((b & (b-1)) == 0) {
+        return a & (b-1);
+    }
+    return a % b;
+}
+
+uint fastdiv(uint a, uint b) {
+    return (a < b) ? 0 : (a / b);
+}
+
+void get_indices(uint idx, out uint i00, out uint i01, out uint i02, out uint i03, uint ne00, uint ne01, uint ne02, uint ne03) {
+    i03 = fastdiv(idx, (ne02*ne01*ne00));
+    const uint i03_offset = i03 * ne02*ne01*ne00;
+    i02 = fastdiv((idx - i03_offset), (ne01*ne00));
+    const uint i02_offset = i02*ne01*ne00;
+    i01 = (idx - i03_offset - i02_offset) / ne00;
+    i00 = idx - i03_offset - i02_offset - i01*ne00;
+}
+
+#endif // UTILS_COMP
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
index 68933d19f2e..67a9ee422bf 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -677,6 +677,8 @@ void process_shaders() {
 
     string_to_spv("add_id_f32", "add_id.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
 
+    string_to_spv("multi_add_f32", "multi_add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}});
+
     for (auto &c : compiles) {
         c.wait();
     }

From 786d86b64793916fcfacbb5c6ff25ef03c8833e6 Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Sun, 17 Aug 2025 03:41:45 -0500
Subject: [PATCH 77/86] vulkan: Optimize argsort (llama/15354)

- Launch an appropriate number of invocations (next larger power of two).
32 invocations is common and the barrier is much cheaper there.
- Specialize for "needs bounds checking" vs not.
- Make the code less branchy and [[unroll]] the loops. In the final code,
I see no branches inside the main loop (only predicated stores) when
needs_bounds_check is false.
- Always sort ascending, then apply the ascending vs descending option when
doing the final stores to memory.
- Copy the values into shared memory, makes them slightly cheaper to access.
---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp          | 24 +++----
 .../ggml-vulkan/vulkan-shaders/argsort.comp   | 68 +++++++++++--------
 2 files changed, 50 insertions(+), 42 deletions(-)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 20674bee588..4ecfb825665 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -345,6 +345,9 @@ enum vk_conv_shapes {
     CONV_SHAPE_COUNT,
 };
 
+static constexpr uint32_t num_argsort_pipelines = 11;
+static constexpr uint32_t max_argsort_cols = 1 << (num_argsort_pipelines-1);
+
 struct vk_device_struct {
     std::recursive_mutex mutex;
 
@@ -505,7 +508,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16;
     vk_pipeline pipeline_rope_multi_f32, pipeline_rope_multi_f16;
     vk_pipeline pipeline_rope_vision_f32, pipeline_rope_vision_f16;
-    vk_pipeline pipeline_argsort_f32;
+    vk_pipeline pipeline_argsort_f32[num_argsort_pipelines];
     vk_pipeline pipeline_sum_rows_f32;
     vk_pipeline pipeline_argmax_f32;
     vk_pipeline pipeline_count_equal_i32;
@@ -870,7 +873,6 @@ struct vk_op_soft_max_push_constants {
 
 struct vk_op_argsort_push_constants {
     uint32_t ncols;
-    uint32_t ncols_pad;
     int32_t order;
 };
 
@@ -3099,7 +3101,9 @@ static void ggml_vk_load_shaders(vk_device& device) {
         ggml_vk_create_pipeline(device, device->pipeline_rope_vision_f16, "rope_vision_f16", rope_vision_f16_len, rope_vision_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
     }
 
-    ggml_vk_create_pipeline(device, device->pipeline_argsort_f32, "argsort_f32", argsort_f32_len, argsort_f32_data, "main", 2, sizeof(vk_op_argsort_push_constants), {1024, 1, 1}, {}, 1);
+    for (uint32_t i = 0; i < num_argsort_pipelines; ++i) {
+        ggml_vk_create_pipeline(device, device->pipeline_argsort_f32[i], "argsort_f32_"+std::to_string(i), argsort_f32_len, argsort_f32_data, "main", 2, sizeof(vk_op_argsort_push_constants), {1u<<i, 1, 1}, {1u<<i, i}, 1, true);
+    }
 
     ggml_vk_create_pipeline(device, device->pipeline_argmax_f32, "argmax_f32", argmax_f32_len, argmax_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
 
@@ -7160,7 +7164,8 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         }
     case GGML_OP_ARGSORT:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_I32) {
-            return ctx->device->pipeline_argsort_f32;
+            uint32_t idx = (uint32_t)ceilf(log2f(float(dst->ne[0])));
+            return ctx->device->pipeline_argsort_f32[idx];
         }
         return nullptr;
     case GGML_OP_SUM:
@@ -8485,16 +8490,8 @@ static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context& subctx, c
 
     uint32_t ncols = src0->ne[0];
 
-    uint32_t ncols_pad = 1;
-    while (ncols_pad < ncols) {
-        ncols_pad *= 2;
-    }
-
-    GGML_ASSERT(ncols_pad <= 1024);
-
     ggml_vk_op_f32<vk_op_argsort_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_ARGSORT, {
         ncols,
-        ncols_pad,
         op_params[0],
     }, dryrun);
 }
@@ -11367,6 +11364,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_OPT_STEP_ADAMW:
         case GGML_OP_OPT_STEP_SGD:
             return op->src[0]->type == GGML_TYPE_F32;
+        case GGML_OP_ARGSORT:
+            return op->ne[0] <= max_argsort_cols;
         case GGML_OP_UPSCALE:
         case GGML_OP_ACC:
         case GGML_OP_CONCAT:
@@ -11376,7 +11375,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_SOFT_MAX_BACK:
-        case GGML_OP_ARGSORT:
         case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:
         case GGML_OP_ARGMAX:
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp b/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp
index d4fa45b1e10..dc53a401e0b 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp
@@ -1,22 +1,24 @@
 #version 450
+#extension GL_EXT_control_flow_attributes : enable
 
 #include "types.comp"
 
-#define BLOCK_SIZE 1024
+layout(constant_id = 0) const int BLOCK_SIZE = 1024;
+layout(constant_id = 1) const int BLOCK_SIZE_LOG2 = 10;
 #define ASC 0
 
-layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 layout (binding = 1)          buffer D {int data_d[];};
 
 layout (push_constant) uniform parameter {
     uint ncols;
-    uint ncols_pad;
     uint order;
 } p;
 
 shared int dst_row[BLOCK_SIZE];
+shared A_TYPE a_sh[BLOCK_SIZE];
 
 void swap(uint idx0, uint idx1) {
     int tmp = dst_row[idx0];
@@ -24,7 +26,7 @@ void swap(uint idx0, uint idx1) {
     dst_row[idx1] = tmp;
 }
 
-void main() {
+void argsort(bool needs_bounds_check) {
     // bitonic sort
     const int col = int(gl_LocalInvocationID.x);
     const uint row = gl_WorkGroupID.y;
@@ -32,38 +34,46 @@ void main() {
     const uint row_offset = row * p.ncols;
 
     // initialize indices
-    if (col < p.ncols_pad) {
-        dst_row[col] = col;
-    }
+    dst_row[col] = col;
+    a_sh[col] = data_a[row_offset + col];
     barrier();
 
-    for (uint k = 2; k <= p.ncols_pad; k *= 2) {
-        for (uint j = k / 2; j > 0; j /= 2) {
-            const uint ixj = col ^ j;
-            if (col < p.ncols_pad && ixj > col) {
-                if ((col & k) == 0) {
-                    if (dst_row[col] >= p.ncols ||
-                        (dst_row[ixj] < p.ncols && (p.order == ASC ?
-                            data_a[row_offset + dst_row[col]] > data_a[row_offset + dst_row[ixj]] :
-                            data_a[row_offset + dst_row[col]] < data_a[row_offset + dst_row[ixj]]))
-                    ) {
-                        swap(col, ixj);
-                    }
-                } else {
-                    if (dst_row[ixj] >= p.ncols ||
-                        (dst_row[col] < p.ncols && (p.order == ASC ?
-                            data_a[row_offset + dst_row[col]] < data_a[row_offset + dst_row[ixj]] :
-                            data_a[row_offset + dst_row[col]] > data_a[row_offset + dst_row[ixj]]))
-                    ) {
-                        swap(col, ixj);
-                    }
-                }
+    uint num_outer_loop_iters = BLOCK_SIZE_LOG2;
+    [[unroll]] for (uint k = 2, outer_idx = 0; outer_idx < num_outer_loop_iters; k *= 2, outer_idx++) {
+        uint num_inner_loop_iters = outer_idx + 1;
+        [[unroll]] for (uint j = k / 2, inner_idx = 0; inner_idx < num_inner_loop_iters; j /= 2, inner_idx++) {
+            const int ixj = int(col ^ j);
+
+            int idx_0 = (col & k) == 0 ? col : ixj;
+            int idx_1 = (col & k) == 0 ? ixj : col;
+
+            int sh_idx_0 = dst_row[idx_0];
+            int sh_idx_1 = dst_row[idx_1];
+            bool idx_0_oob = needs_bounds_check ? sh_idx_0 >= p.ncols : false;
+            bool idx_1_oob = needs_bounds_check ? sh_idx_1 >= p.ncols : false;
+
+            if ((idx_0_oob ||
+                (!idx_1_oob && a_sh[sh_idx_0] > a_sh[sh_idx_1])) && (ixj > col)) {
+                swap(idx_0, idx_1);
             }
+
             barrier();
         }
     }
 
     if (col < p.ncols) {
-        data_d[row_offset + col] = dst_row[col];
+        if (p.order == ASC) {
+            data_d[row_offset + col] = dst_row[col];
+        } else {
+            data_d[row_offset + p.ncols - col - 1] = dst_row[col];
+        }
+    }
+}
+
+void main() {
+    if (p.ncols == BLOCK_SIZE) {
+        argsort(false);
+    } else {
+        argsort(true);
     }
 }

From 185db9b0f8774b1c21bbfd11da80dfa4781e3ee4 Mon Sep 17 00:00:00 2001
From: Dong Won Kim <63934649+ddwkim@users.noreply.github.com>
Date: Sun, 17 Aug 2025 23:03:09 +0900
Subject: [PATCH 78/86] vulkan: support sqrt (llama/15370)

---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp          | 23 +++++++++++++++++++
 ggml/src/ggml-vulkan/vulkan-shaders/sqrt.comp | 17 ++++++++++++++
 .../vulkan-shaders/vulkan-shaders-gen.cpp     |  2 ++
 3 files changed, 42 insertions(+)
 create mode 100644 ggml/src/ggml-vulkan/vulkan-shaders/sqrt.comp

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 4ecfb825665..bc227ffb9f0 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -464,6 +464,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_upscale_nearest_f32, pipeline_upscale_bilinear_f32, pipeline_upscale_bilinear_ac_f32;
     vk_pipeline pipeline_scale_f32;
     vk_pipeline pipeline_sqr_f32;
+    vk_pipeline pipeline_sqrt_f32;
     vk_pipeline pipeline_sin_f32;
     vk_pipeline pipeline_cos_f32;
     vk_pipeline pipeline_clamp_f32;
@@ -3031,6 +3032,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_scale_f32, "scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_sqr_f32, "sqr_f32", sqr_f32_len, sqr_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_sqrt_f32, "sqrt_f32", sqrt_f32_len, sqrt_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_sin_f32, "sin_f32", sin_f32_len, sin_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_cos_f32, "cos_f32", cos_f32_len, cos_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
@@ -6981,6 +6983,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_sqr_f32;
         }
         return nullptr;
+    case GGML_OP_SQRT:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_sqrt_f32;
+        }
+        return nullptr;
     case GGML_OP_SIN:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_sin_f32;
@@ -7290,6 +7297,7 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
     case GGML_OP_CONCAT:
     case GGML_OP_UPSCALE:
     case GGML_OP_SQR:
+    case GGML_OP_SQRT:
     case GGML_OP_SIN:
     case GGML_OP_COS:
     case GGML_OP_CLAMP:
@@ -7595,6 +7603,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_MUL:
     case GGML_OP_SCALE:
     case GGML_OP_SQR:
+    case GGML_OP_SQRT:
     case GGML_OP_SIN:
     case GGML_OP_COS:
     case GGML_OP_CLAMP:
@@ -8242,6 +8251,10 @@ static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context& subctx, const
     ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SQR, vk_op_unary_push_constants_init(src0, dst), dryrun);
 }
 
+static void ggml_vk_sqrt(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SQRT, vk_op_unary_push_constants_init(src0, dst), dryrun);
+}
+
 static void ggml_vk_sin(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SIN, vk_op_unary_push_constants_init(src0, dst), dryrun);
 }
@@ -9697,6 +9710,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_UPSCALE:
     case GGML_OP_SCALE:
     case GGML_OP_SQR:
+    case GGML_OP_SQRT:
     case GGML_OP_SIN:
     case GGML_OP_COS:
     case GGML_OP_CLAMP:
@@ -9766,6 +9780,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         case GGML_OP_UPSCALE:
         case GGML_OP_SCALE:
         case GGML_OP_SQR:
+        case GGML_OP_SQRT:
         case GGML_OP_SIN:
         case GGML_OP_COS:
         case GGML_OP_CLAMP:
@@ -9867,6 +9882,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_SQR:
         ggml_vk_sqr(ctx, compute_ctx, src0, node, dryrun);
 
+        break;
+    case GGML_OP_SQRT:
+        ggml_vk_sqrt(ctx, compute_ctx, src0, node, dryrun);
+
         break;
     case GGML_OP_SIN:
         ggml_vk_sin(ctx, compute_ctx, src0, node, dryrun);
@@ -10118,6 +10137,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
     case GGML_OP_UPSCALE:
     case GGML_OP_SCALE:
     case GGML_OP_SQR:
+    case GGML_OP_SQRT:
     case GGML_OP_SIN:
     case GGML_OP_COS:
     case GGML_OP_CLAMP:
@@ -11357,6 +11377,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_SILU_BACK:
         case GGML_OP_RMS_NORM_BACK:
         case GGML_OP_SQR:
+        case GGML_OP_SQRT:
         case GGML_OP_SIN:
         case GGML_OP_COS:
         case GGML_OP_CLAMP:
@@ -11801,6 +11822,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         tensor_clone = ggml_scale_bias(ggml_ctx, src_clone[0], params[0], params[1]);
     } else if (tensor->op == GGML_OP_SQR) {
         tensor_clone = ggml_sqr(ggml_ctx, src_clone[0]);
+    } else if (tensor->op == GGML_OP_SQRT) {
+        tensor_clone = ggml_sqrt(ggml_ctx, src_clone[0]);
     } else if (tensor->op == GGML_OP_SIN) {
         tensor_clone = ggml_sin(ggml_ctx, src_clone[0]);
     } else if (tensor->op == GGML_OP_COS) {
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/sqrt.comp b/ggml/src/ggml-vulkan/vulkan-shaders/sqrt.comp
new file mode 100644
index 00000000000..4bc697b9b9c
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/sqrt.comp
@@ -0,0 +1,17 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(sqrt(val));
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
index 67a9ee422bf..8c6627dee43 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -566,6 +566,8 @@ void process_shaders() {
 
     string_to_spv("sqr_f32", "square.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
 
+    string_to_spv("sqrt_f32", "sqrt.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+
     string_to_spv("sin_f32", "sin.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
 
     string_to_spv("cos_f32", "cos.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});

From 2eefbe12a88bfe15c508611cdce9589e66818c30 Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Sun, 17 Aug 2025 11:08:57 -0500
Subject: [PATCH 79/86] vulkan: Use larger workgroups for mul_mat_vec when M is
 small (llama/15355)

* vulkan: Use larger workgroups for mul_mat_vec when M is small

Also use subgroup instructions for (part of) the reduction when supported.
Without this, the more expensive reductions would eat into the benefits of
the larger workgroups.

* update heuristic for amd/intel

Co-authored-by: 0cc4m <picard12@live.de>

---------

Co-authored-by: 0cc4m <picard12@live.de>
---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp          | 136 +++++++++++-------
 .../vulkan-shaders/mul_mat_vec_base.comp      |  38 ++++-
 .../vulkan-shaders/vulkan-shaders-gen.cpp     |  15 ++
 3 files changed, 135 insertions(+), 54 deletions(-)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index bc227ffb9f0..7ef938066b1 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -345,6 +345,12 @@ enum vk_conv_shapes {
     CONV_SHAPE_COUNT,
 };
 
+enum dmmv_wg_sizes {
+    DMMV_WG_SIZE_SUBGROUP,
+    DMMV_WG_SIZE_LARGE,
+    DMMV_WG_SIZE_COUNT,
+};
+
 static constexpr uint32_t num_argsort_pipelines = 11;
 static constexpr uint32_t max_argsort_cols = 1 << (num_argsort_pipelines-1);
 
@@ -435,8 +441,8 @@ struct vk_device_struct {
     vk_pipeline pipeline_quantize_q8_1;
 
     vk_pipeline pipeline_dequant[GGML_TYPE_COUNT];
-    vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_COUNT][mul_mat_vec_max_cols];
-    vk_pipeline pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_COUNT][mul_mat_vec_max_cols];
+    vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[DMMV_WG_SIZE_COUNT][GGML_TYPE_COUNT][mul_mat_vec_max_cols];
+    vk_pipeline pipeline_dequant_mul_mat_vec_f16_f32[DMMV_WG_SIZE_COUNT][GGML_TYPE_COUNT][mul_mat_vec_max_cols];
     vk_pipeline pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_COUNT];
 
     vk_pipeline pipeline_mul_mat_vec_p021_f16_f32[p021_max_gqa_ratio];
@@ -2773,54 +2779,61 @@ static void ggml_vk_load_shaders(vk_device& device) {
         rm_stdq = 2;
     uint32_t rm_iq = 2 * rm_kq;
 
-    for (uint32_t i = 0; i < mul_mat_vec_max_cols; ++i) {
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f32_f32_"+std::to_string(i+1),  mul_mat_vec_f32_f32_f32_len,  mul_mat_vec_f32_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f32_f32_"+std::to_string(i+1),  mul_mat_vec_f16_f32_f32_len,  mul_mat_vec_f16_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f32_f32_"+std::to_string(i+1), mul_mat_vec_bf16_f32_f32_len, mul_mat_vec_bf16_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ1_S][i],   "mul_mat_vec_iq1_s_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_s_f32_f32_len,   mul_mat_vec_iq1_s_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ1_M][i],   "mul_mat_vec_iq1_m_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_m_f32_f32_len,   mul_mat_vec_iq1_m_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xxs_f32_f32_len, mul_mat_vec_iq2_xxs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f32_f32_"+std::to_string(i+1),  mul_mat_vec_iq2_xs_f32_f32_len,  mul_mat_vec_iq2_xs_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq2_s_f32_f32_len,   mul_mat_vec_iq2_s_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq3_xxs_f32_f32_len, mul_mat_vec_iq3_xxs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq3_s_f32_f32_len,   mul_mat_vec_iq3_s_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f32_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_xs_f32_f32_len,  mul_mat_vec_iq4_xs_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f32_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_nl_f32_f32_len,  mul_mat_vec_iq4_nl_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_MXFP4][i],   "mul_mat_vec_mxfp4_f32_f32_"+std::to_string(i+1),   mul_mat_vec_mxfp4_f32_f32_len,   mul_mat_vec_mxfp4_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f16_f32_"+std::to_string(i+1),  mul_mat_vec_f32_f16_f32_len,  mul_mat_vec_f32_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f16_f32_"+std::to_string(i+1),  mul_mat_vec_f16_f16_f32_len,  mul_mat_vec_f16_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f16_f32_"+std::to_string(i+1), mul_mat_vec_bf16_f16_f32_len, mul_mat_vec_bf16_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ1_S][i],   "mul_mat_vec_iq1_s_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_s_f16_f32_len,   mul_mat_vec_iq1_s_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ1_M][i],   "mul_mat_vec_iq1_m_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_m_f16_f32_len,   mul_mat_vec_iq1_m_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xxs_f16_f32_len, mul_mat_vec_iq2_xxs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f16_f32_"+std::to_string(i+1),  mul_mat_vec_iq2_xs_f16_f32_len,  mul_mat_vec_iq2_xs_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq2_s_f16_f32_len,   mul_mat_vec_iq2_s_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq3_xxs_f16_f32_len, mul_mat_vec_iq3_xxs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq3_s_f16_f32_len,   mul_mat_vec_iq3_s_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f16_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_xs_f16_f32_len,  mul_mat_vec_iq4_xs_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f16_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_nl_f16_f32_len,  mul_mat_vec_iq4_nl_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_MXFP4][i],   "mul_mat_vec_mxfp4_f16_f32_"+std::to_string(i+1),   mul_mat_vec_mxfp4_f16_f32_len,   mul_mat_vec_mxfp4_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+    for (uint32_t w = 0; w < DMMV_WG_SIZE_COUNT; ++w) {
+        uint32_t wg_size_subgroup16 = (w == DMMV_WG_SIZE_SUBGROUP) ? subgroup_size_16 : (subgroup_size_16 * 4);
+        uint32_t wg_size_subgroup   = (w == DMMV_WG_SIZE_SUBGROUP) ? device->subgroup_size : (device->subgroup_size * 4);
+
+        const bool s = device->subgroup_add && device->architecture != vk_device_architecture::AMD_GCN;
+
+        for (uint32_t i = 0; i < mul_mat_vec_max_cols; ++i) {
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f32_f32_"+std::to_string(w)+"_"+std::to_string(i+1),  arr_dmmv_f32_f32_f32_len[s],  arr_dmmv_f32_f32_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f32_f32_"+std::to_string(w)+"_"+std::to_string(i+1),  arr_dmmv_f16_f32_f32_len[s],  arr_dmmv_f16_f32_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f32_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_bf16_f32_f32_len[s], arr_dmmv_bf16_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f32_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_q4_0_f32_f32_len[s], arr_dmmv_q4_0_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f32_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_q4_1_f32_f32_len[s], arr_dmmv_q4_1_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f32_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_q5_0_f32_f32_len[s], arr_dmmv_q5_0_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f32_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_q5_1_f32_f32_len[s], arr_dmmv_q5_1_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f32_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_q8_0_f32_f32_len[s], arr_dmmv_q8_0_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup, 1*rm_stdq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f32_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_q2_k_f32_f32_len[s], arr_dmmv_q2_k_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f32_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_q3_k_f32_f32_len[s], arr_dmmv_q3_k_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f32_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_q4_k_f32_f32_len[s], arr_dmmv_q4_k_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f32_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_q5_k_f32_f32_len[s], arr_dmmv_q5_k_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f32_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_q6_k_f32_f32_len[s], arr_dmmv_q6_k_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ1_S][i],   "mul_mat_vec_iq1_s_f32_f32_"+std::to_string(w)+"_"+std::to_string(i+1),   arr_dmmv_iq1_s_f32_f32_len[s],   arr_dmmv_iq1_s_f32_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ1_M][i],   "mul_mat_vec_iq1_m_f32_f32_"+std::to_string(w)+"_"+std::to_string(i+1),   arr_dmmv_iq1_m_f32_f32_len[s],   arr_dmmv_iq1_m_f32_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f32_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_iq2_xxs_f32_f32_len[s], arr_dmmv_iq2_xxs_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f32_f32_"+std::to_string(w)+"_"+std::to_string(i+1),  arr_dmmv_iq2_xs_f32_f32_len[s],  arr_dmmv_iq2_xs_f32_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f32_f32_"+std::to_string(w)+"_"+std::to_string(i+1),   arr_dmmv_iq2_s_f32_f32_len[s],   arr_dmmv_iq2_s_f32_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f32_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_iq3_xxs_f32_f32_len[s], arr_dmmv_iq3_xxs_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f32_f32_"+std::to_string(w)+"_"+std::to_string(i+1),   arr_dmmv_iq3_s_f32_f32_len[s],   arr_dmmv_iq3_s_f32_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f32_f32_"+std::to_string(w)+"_"+std::to_string(i+1),  arr_dmmv_iq4_xs_f32_f32_len[s],  arr_dmmv_iq4_xs_f32_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f32_f32_"+std::to_string(w)+"_"+std::to_string(i+1),  arr_dmmv_iq4_nl_f32_f32_len[s],  arr_dmmv_iq4_nl_f32_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_MXFP4][i],   "mul_mat_vec_mxfp4_f32_f32_"+std::to_string(w)+"_"+std::to_string(i+1),   arr_dmmv_mxfp4_f32_f32_len[s],   arr_dmmv_mxfp4_f32_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f16_f32_"+std::to_string(w)+"_"+std::to_string(i+1),  arr_dmmv_f32_f16_f32_len[s],  arr_dmmv_f32_f16_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f16_f32_"+std::to_string(w)+"_"+std::to_string(i+1),  arr_dmmv_f16_f16_f32_len[s],  arr_dmmv_f16_f16_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f16_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_bf16_f16_f32_len[s], arr_dmmv_bf16_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f16_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_q4_0_f16_f32_len[s], arr_dmmv_q4_0_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f16_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_q4_1_f16_f32_len[s], arr_dmmv_q4_1_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f16_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_q5_0_f16_f32_len[s], arr_dmmv_q5_0_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f16_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_q5_1_f16_f32_len[s], arr_dmmv_q5_1_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f16_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_q8_0_f16_f32_len[s], arr_dmmv_q8_0_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup, 1*rm_stdq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f16_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_q2_k_f16_f32_len[s], arr_dmmv_q2_k_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f16_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_q3_k_f16_f32_len[s], arr_dmmv_q3_k_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f16_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_q4_k_f16_f32_len[s], arr_dmmv_q4_k_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f16_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_q5_k_f16_f32_len[s], arr_dmmv_q5_k_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f16_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_q6_k_f16_f32_len[s], arr_dmmv_q6_k_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ1_S][i],   "mul_mat_vec_iq1_s_f16_f32_"+std::to_string(w)+"_"+std::to_string(i+1),   arr_dmmv_iq1_s_f16_f32_len[s],   arr_dmmv_iq1_s_f16_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ1_M][i],   "mul_mat_vec_iq1_m_f16_f32_"+std::to_string(w)+"_"+std::to_string(i+1),   arr_dmmv_iq1_m_f16_f32_len[s],   arr_dmmv_iq1_m_f16_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f16_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_iq2_xxs_f16_f32_len[s], arr_dmmv_iq2_xxs_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f16_f32_"+std::to_string(w)+"_"+std::to_string(i+1),  arr_dmmv_iq2_xs_f16_f32_len[s],  arr_dmmv_iq2_xs_f16_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f16_f32_"+std::to_string(w)+"_"+std::to_string(i+1),   arr_dmmv_iq2_s_f16_f32_len[s],   arr_dmmv_iq2_s_f16_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f16_f32_"+std::to_string(w)+"_"+std::to_string(i+1), arr_dmmv_iq3_xxs_f16_f32_len[s], arr_dmmv_iq3_xxs_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f16_f32_"+std::to_string(w)+"_"+std::to_string(i+1),   arr_dmmv_iq3_s_f16_f32_len[s],   arr_dmmv_iq3_s_f16_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f16_f32_"+std::to_string(w)+"_"+std::to_string(i+1),  arr_dmmv_iq4_xs_f16_f32_len[s],  arr_dmmv_iq4_xs_f16_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f16_f32_"+std::to_string(w)+"_"+std::to_string(i+1),  arr_dmmv_iq4_nl_f16_f32_len[s],  arr_dmmv_iq4_nl_f16_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_MXFP4][i],   "mul_mat_vec_mxfp4_f16_f32_"+std::to_string(w)+"_"+std::to_string(i+1),   arr_dmmv_mxfp4_f16_f32_len[s],   arr_dmmv_mxfp4_f16_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+        }
     }
 
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32",  mul_mat_vec_id_f32_f32_len,  mul_mat_vec_id_f32_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
@@ -4389,7 +4402,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
     return (ctx->device->fp16 && prec == GGML_PREC_DEFAULT) ? ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f32acc;
 }
 
-static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type, uint32_t num_cols) {
+static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type, uint32_t num_cols, uint32_t m, uint32_t k) {
     VK_LOG_DEBUG("ggml_vk_get_dequantize_mul_mat_vec()");
     GGML_ASSERT(b_type == GGML_TYPE_F32 || b_type == GGML_TYPE_F16);
     GGML_ASSERT(num_cols >= 1 && num_cols <= mul_mat_vec_max_cols);
@@ -4423,7 +4436,24 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context *
             return nullptr;
     }
 
-    return b_type == GGML_TYPE_F32 ? ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[a_type][num_cols-1] : ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[a_type][num_cols-1];
+    // heuristic to choose workgroup size
+    uint32_t dmmv_wg = DMMV_WG_SIZE_SUBGROUP;
+    if (ctx->device->vendor_id == VK_VENDOR_ID_NVIDIA || ctx->device->vendor_id == VK_VENDOR_ID_INTEL) {
+        // Prefer larger workgroups when M is small, to spread the work out more
+        // and keep more SMs busy.
+        // q6_k seems to prefer small workgroup size even for "medium" values of M.
+        if (a_type == GGML_TYPE_Q6_K) {
+            if (m < 4096 && k >= 1024) {
+                dmmv_wg = DMMV_WG_SIZE_LARGE;
+            }
+        } else {
+            if (m <= 8192 && k >= 1024) {
+                dmmv_wg = DMMV_WG_SIZE_LARGE;
+            }
+        }
+    }
+
+    return b_type == GGML_TYPE_F32 ? ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[dmmv_wg][a_type][num_cols-1] : ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[dmmv_wg][a_type][num_cols-1];
 }
 
 static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_context * ctx, ggml_type src0_type, ggml_type src1_type, ggml_prec prec) {
@@ -5730,7 +5760,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     } else {
         to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
     }
-    vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type, src1->type, ne11);
+    vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type, src1->type, ne11, ne20, ne00);
     GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr);  // NOLINT
     GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr);  // NOLINT
     GGML_ASSERT(dmmv != nullptr);
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp
index 903753c7e2e..b93e9948f7f 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp
@@ -1,6 +1,10 @@
 #extension GL_EXT_control_flow_attributes : enable
 #extension GL_EXT_shader_16bit_storage : require
 #extension GL_EXT_shader_8bit_storage : require
+#if USE_SUBGROUP_ADD
+#extension GL_KHR_shader_subgroup_basic : require
+#extension GL_KHR_shader_subgroup_arithmetic : require
+#endif
 
 #ifdef MUL_MAT_ID
 #define EXPERT_COUNT 8
@@ -90,7 +94,38 @@ layout (constant_id = 2) const uint NUM_COLS = 1;
 
 shared FLOAT_TYPE tmpsh[NUM_COLS][NUM_ROWS][BLOCK_SIZE];
 
-void reduce_result(const in FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offset, const in uint32_t first_row, const in uint32_t num_rows, const in uint32_t tid) {
+void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offset, const in uint32_t first_row, const in uint32_t num_rows, const in uint32_t tid) {
+    // subgroupAdd is probably faster on devices that support it,
+    // particularly when the workgroup has more than one subgroup
+#if USE_SUBGROUP_ADD
+    // sum up partial sums within a subgroup
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            temp[j][n] = subgroupAdd(temp[j][n]);
+        }
+    }
+
+    // Go through shared memory to sum partials across subgroups
+    if (gl_SubgroupInvocationID == 0) {
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                tmpsh[j][n][gl_SubgroupID] = temp[j][n];
+            }
+        }
+    }
+    barrier();
+    if (tid == 0) {
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                temp[j][n] = FLOAT_TYPE(0);
+                [[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
+                    temp[j][n] += tmpsh[j][n][s];
+                }
+                data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(temp[j][n]);
+            }
+        }
+    }
+#else
     // sum up partial sums and write back result
     [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
         [[unroll]] for (uint n = 0; n < num_rows; ++n) {
@@ -115,4 +150,5 @@ void reduce_result(const in FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32
             }
         }
     }
+#endif
 }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
index 8c6627dee43..6cc474b90a3 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -472,6 +472,9 @@ void process_shaders() {
         string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
         string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}}));
 
+        string_to_spv("mul_mat_vec_" + tname + "_f32_f32_subgroup", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
+        string_to_spv("mul_mat_vec_" + tname + "_f16_f32_subgroup", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
+
         string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
 
         // Dequant shaders
@@ -789,6 +792,18 @@ void write_output_files() {
         fputs(data.c_str(), src);
         fputs(len.c_str(), src);
     }
+
+    for (const std::string& btype : {"f16", "f32"}) {
+    for (const auto& tname : type_names) {
+        fprintf(hdr, "extern unsigned char *arr_dmmv_%s_%s_f32_data[2];\n", tname.c_str(), btype.c_str());
+        fprintf(hdr, "extern uint64_t arr_dmmv_%s_%s_f32_len[2];\n", tname.c_str(), btype.c_str());
+        std::string data = "unsigned char *arr_dmmv_" + tname + "_" + btype + "_f32_data[2] = {mul_mat_vec_" + tname + "_" + btype + "_f32_data, mul_mat_vec_" + tname + "_" + btype + "_f32_subgroup_data};\n";
+        std::string len =  "uint64_t arr_dmmv_"       + tname + "_" + btype + "_f32_len[2] =  {mul_mat_vec_" + tname + "_" + btype + "_f32_len,  mul_mat_vec_" + tname + "_" + btype + "_f32_subgroup_len};\n";
+        fputs(data.c_str(), src);
+        fputs(len.c_str(), src);
+    }
+    }
+
     fclose(hdr);
     fclose(src);
 }

From cb061b862ec071ebf8ec5ebdf717a26f5d76f77b Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Mon, 18 Aug 2025 00:56:29 -0500
Subject: [PATCH 80/86] vulkan: disable spirv-opt for bfloat16 shaders
 (llama/15352)

---
 ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
index 6cc474b90a3..75c572d6fc9 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -223,7 +223,8 @@ void string_to_spv_func(const std::string& _name, const std::string& in_fname, c
     std::string target_env = (name.find("_cm2") != std::string::npos) ? "--target-env=vulkan1.3" : "--target-env=vulkan1.2";
 
     // disable spirv-opt for coopmat shaders for https://github.com/ggerganov/llama.cpp/issues/10734
-    std::string opt_level = coopmat ? "" : "-O";
+    // disable spirv-opt for bf16 shaders for https://github.com/ggml-org/llama.cpp/issues/15344
+    std::string opt_level = (coopmat || name.find("bf16") != std::string::npos) ? "" : "-O";
 
     #ifdef _WIN32
         std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, opt_level, "\"" + in_path + "\"", "-o", "\"" + out_fname + "\""};

From 50a3a0caca85f71b78887a38e1eb2aabbe282813 Mon Sep 17 00:00:00 2001
From: compilade <git@compilade.net>
Date: Mon, 18 Aug 2025 03:23:56 -0400
Subject: [PATCH 81/86] ggml-quants : fix make_qp_quants NANs and IQ1 assertion
 errors (llama/15379)

* ggml-quants : fix make_qp_quants NANs and IQ1 assertion errors

* ggml-quants : avoid division by zero in make_q3_quants
---
 ggml/src/ggml-quants.c | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/ggml/src/ggml-quants.c b/ggml/src/ggml-quants.c
index 94f6405ca1e..727932123e4 100644
--- a/ggml/src/ggml-quants.c
+++ b/ggml/src/ggml-quants.c
@@ -566,7 +566,7 @@ static float make_q3_quants(int n, int nmax, const float * GGML_RESTRICT x, int8
         for (int i = 0; i < n; ++i) {
             L[i] += nmax;
         }
-        return sumlx / suml2;
+        return suml2 > 0.0f ? sumlx / suml2 : 0.0f;
     }
     for (int i = 0; i < n; ++i) {
         int l = nearest_int(iscale * x[i]);
@@ -901,7 +901,7 @@ static float make_qp_quants(int n, int nmax, const float * GGML_RESTRICT x, uint
     for (int i = 0; i < n; ++i) {
         max = MAX(max, x[i]);
     }
-    if (!max) { // all zero
+    if (max < GROUP_MAX_EPS) { // all zero
         for (int i = 0; i < n; ++i) { L[i] = 0; }
         return 0.f;
     }
@@ -966,7 +966,7 @@ static float make_qp_quants(int n, int nmax, const float * GGML_RESTRICT x, uint
             break;
         }
     }
-    return sumlx/suml2;
+    return suml2 > 0.0f ? sumlx / suml2 : 0.0f;
 }
 
 static void quantize_row_q2_K_impl(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int k, const float * GGML_RESTRICT quant_weights) {
@@ -4266,7 +4266,7 @@ static void quantize_row_iq1_s_impl(const float * GGML_RESTRICT x, void * GGML_R
                     sumw[j+1] = sumw[j] + weight[i];
                 }
             }
-            float best_score = -FLT_MIN, scale = max;
+            float best_score = -FLT_MAX, scale = max;
             int besti1 = -1, besti2 = -1, best_shift = 0;
             for (int i1 = 0; i1 <= block_size; ++i1) {
                 for (int i2 = i1; i2 <= block_size; ++i2) {
@@ -4442,7 +4442,7 @@ static void quantize_row_iq1_m_impl(const float * GGML_RESTRICT x, void * GGML_R
                 idx[2*j] = j;
             }
             qsort(pairs, block_size, 2*sizeof(float), iq1_sort_helper);
-            float best_score = -FLT_MIN, scale = max;
+            float best_score = -FLT_MAX, scale = max;
             int besti1 = -1, besti2 = -1, best_k = -1;
             // 0: +, +
             // 1: +, -

From b309fadafad2d62c0035b852c993f8154e9f3ad0 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Mon, 18 Aug 2025 18:57:45 +0300
Subject: [PATCH 82/86] common : handle mxfp4 enum

ggml-ci
---
 examples/common-ggml.cpp | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/examples/common-ggml.cpp b/examples/common-ggml.cpp
index d59b77fce6b..c42b644fedd 100644
--- a/examples/common-ggml.cpp
+++ b/examples/common-ggml.cpp
@@ -72,6 +72,7 @@ bool ggml_common_quantize_0(
         case GGML_FTYPE_MOSTLY_IQ4_XS:
         case GGML_FTYPE_MOSTLY_IQ1_M:
         case GGML_FTYPE_MOSTLY_BF16:
+        case GGML_FTYPE_MOSTLY_MXFP4:
                 {
                     fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
                     return false;
@@ -211,6 +212,7 @@ bool ggml_common_quantize_0(
                 case GGML_TYPE_BF16:
                 case GGML_TYPE_TQ1_0:
                 case GGML_TYPE_TQ2_0:
+                case GGML_TYPE_MXFP4:
                 case GGML_TYPE_COUNT:
                     {
                         fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));

From 1dee2ff73454cbb9f373993d85e0520288a506f3 Mon Sep 17 00:00:00 2001
From: Aaron Teo <aaron.teo1@ibm.com>
Date: Mon, 18 Aug 2025 19:21:15 +0300
Subject: [PATCH 83/86] ggml : initial zDNN backend (llama/14975)

---
 ggml/src/ggml-zdnn/CMakeLists.txt   |  36 ++
 ggml/src/ggml-zdnn/ggml-zdnn-impl.h |  97 ++++
 ggml/src/ggml-zdnn/ggml-zdnn.cpp    | 846 ++++++++++++++++++++++++++++
 3 files changed, 979 insertions(+)
 create mode 100644 ggml/src/ggml-zdnn/CMakeLists.txt
 create mode 100644 ggml/src/ggml-zdnn/ggml-zdnn-impl.h
 create mode 100644 ggml/src/ggml-zdnn/ggml-zdnn.cpp

diff --git a/ggml/src/ggml-zdnn/CMakeLists.txt b/ggml/src/ggml-zdnn/CMakeLists.txt
new file mode 100644
index 00000000000..0a723ce4de2
--- /dev/null
+++ b/ggml/src/ggml-zdnn/CMakeLists.txt
@@ -0,0 +1,36 @@
+if (DEFINED ZDNN_ROOT)
+    message(STATUS "zdnn: using ZDNN_ROOT override: ${ZDNN_ROOT}")
+    set(ZDNN_HINT "${ZDNN_ROOT}")
+else()
+    set(ZDNN_HINT "")
+endif()
+
+find_path(ZDNN_INCLUDE
+            NAMES zdnn.h
+            HINTS ${ZDNN_HINT} /usr /usr/local
+            PATH_SUFFIXES include)
+if (ZDNN_INCLUDE)
+    message(STATUS "zdnn: found include: ${ZDNN_INCLUDE}")
+else()
+    message(FATAL_ERROR "zdnn: include directory not found, please set ZDNN_ROOT to the proper path if necessary")
+endif()
+
+find_library(ZDNN_LIB
+                NAMES zdnn
+                HINTS ${ZDNN_HINT} /usr /usr/local
+                PATH_SUFFIXES lib lib64)
+if (ZDNN_LIB)
+    message(STATUS "zdnn: found library: ${ZDNN_LIB}")
+else()
+    message(FATAL_ERROR "zdnn: library not found, please set ZDNN_ROOT to the proper path if necessary")
+endif()
+
+file(GLOB GGML_SOURCES_ZDNN "*.c" "*.cpp")
+file(GLOB GGML_HEADERS_ZDNN "*.h" "*.hpp")
+
+ggml_add_backend_library(ggml-zdnn ${GGML_HEADERS_ZDNN} ${GGML_SOURCES_ZDNN})
+target_link_libraries(ggml-zdnn PRIVATE ${ZDNN_LIB})
+target_include_directories(ggml-zdnn PRIVATE ${ZDNN_INCLUDE})
+target_link_directories(ggml-zdnn PRIVATE ${ZDNN_LIB})
+
+target_compile_definitions(ggml-zdnn PRIVATE GGML_USE_ZDNN)
diff --git a/ggml/src/ggml-zdnn/ggml-zdnn-impl.h b/ggml/src/ggml-zdnn/ggml-zdnn-impl.h
new file mode 100644
index 00000000000..9dcb040fa89
--- /dev/null
+++ b/ggml/src/ggml-zdnn/ggml-zdnn-impl.h
@@ -0,0 +1,97 @@
+#ifndef GGML_ZDNN_IMPL
+#define GGML_ZDNN_IMPL
+
+#include "zdnn.h"
+#include "ggml.h"
+#include "ggml-zdnn.h"
+
+#include <vector>
+#include <memory>
+#include <vecintrin.h>
+
+#define GGML_ZDNN_NAME    "zDNN"
+#define GGML_ZDNN_VERSION ZDNN_VERNUM
+
+#define vec_neg(a)    (-(a))                // Vector Negate
+#define vec_add(a, b) ((a) + (b))           // Vector Add
+#define vec_sub(a, b) ((a) - (b))           // Vector Subtract
+#define vec_mul(a, b) ((a) * (b))           // Vector Multiply
+#define vec_div(a, b) ((a) / (b))           // Vector Divide
+#define vec_sl(a, b)  ((a) << (b))          // Vector Shift Left
+#define vec_sra(a, b) ((a) >> (b))          // Vector Shift Right
+#define vec_sr(a, b)  ((a) >> (b))          // Vector Shift Right Algebraic
+#define vec_slo(a, b) vec_slb(a, (b) << 64) // Vector Shift Left by Octet
+#define vec_sro(a, b) vec_srb(a, (b) << 64) // Vector Shift Right by Octet
+
+#ifndef vec_and
+#define vec_and(a, b) ((a) & (b)) // Vector AND
+#endif
+
+#ifndef vec_or
+#define vec_or(a, b)  ((a) | (b)) // Vector OR
+#endif
+
+#ifndef vec_xor
+#define vec_xor(a, b) ((a) ^ (b)) // Vector XOR
+#endif
+
+typedef   signed char char8x16_t  __attribute__((vector_size(16)));
+typedef unsigned char uchar8x16_t __attribute__((vector_size(16)));
+
+typedef int8_t   int8x16_t  __attribute__((vector_size(16)));
+typedef int16_t  int16x8_t  __attribute__((vector_size(16)));
+typedef int32_t  int32x4_t  __attribute__((vector_size(16)));
+typedef uint8_t  uint8x16_t __attribute__((vector_size(16)));
+typedef uint16_t uint16x8_t __attribute__((vector_size(16)));
+typedef uint32_t uint32x4_t __attribute__((vector_size(16)));
+
+typedef float float32x4_t   __attribute__((vector_size(16)));
+typedef double double64x2_t __attribute__((vector_size(16)));
+
+typedef   signed long long long64x2_t  __attribute__((vector_size(16)));
+typedef unsigned long long ulong64x2_t __attribute__((vector_size(16)));
+
+#define ZDNN_CHECK(stmt)                \
+    do {                                \
+        zdnn_status status = (stmt);    \
+        GGML_ASSERT(status == ZDNN_OK); \
+    } while (0);
+
+struct ggml_backend_zdnn_device_context {
+    int zdnn_device;
+    int zdnn_device_ref_count;
+
+    bool has_parmblkformat_0;
+    bool has_parmblkformat_1;
+
+    size_t max_size;
+
+    char name[128];
+};
+
+struct ggml_backend_zdnn_context {
+    int device;
+    ggml_cgraph * gf;
+};
+
+struct ggml_backend_zdnn_buffer {
+    void * data;
+    size_t size;
+
+    zdnn_tensor_desc pre_tfm_desc;
+    zdnn_tensor_desc tfm_desc;
+    zdnn_ztensor     ztensor;
+
+    char name[GGML_MAX_NAME];
+};
+
+struct ggml_backend_zdnn_buffer_context {
+    void * all_data;
+    size_t all_size;
+    bool owned;
+
+    int n_buffers;
+    std::vector<std::unique_ptr<ggml_backend_zdnn_buffer>> buffers;
+};
+
+#endif  // GGML_ZDNN_IMPL
diff --git a/ggml/src/ggml-zdnn/ggml-zdnn.cpp b/ggml/src/ggml-zdnn/ggml-zdnn.cpp
new file mode 100644
index 00000000000..7507a52aeae
--- /dev/null
+++ b/ggml/src/ggml-zdnn/ggml-zdnn.cpp
@@ -0,0 +1,846 @@
+#include "zdnn.h"
+#include "ggml-zdnn.h"
+#include "ggml-zdnn-impl.h"
+
+#include "ggml-impl.h"
+#include "ggml-backend-impl.h"
+
+#include <vector>
+#include <memory>
+#include <csignal>
+#include <unistd.h>
+
+inline zdnn_data_types ggml_zdnn_type_mapping(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_F32:
+            return FP32;
+        case GGML_TYPE_F16:
+            return FP16;
+        case GGML_TYPE_BF16:
+            return BFLOAT;
+        case GGML_TYPE_I8:
+            return INT8;
+        case GGML_TYPE_I32:
+            return INT32;
+        case GGML_TYPE_Q8_0:
+            return INT8;
+        default:
+            GGML_ABORT("%s: fatal: unable to determine zTensor data type",
+                       __func__);
+            break;
+    }
+}
+
+inline void ggml_zdnn_create_tensor(zdnn_tensor_desc  & pre_tfm_desc,
+                                    zdnn_tensor_desc  & tfm_desc,
+                                    zdnn_ztensor      & ztensor,
+                              const ggml_tensor       * src,
+                              const int64_t           * ne,
+                              const zdnn_data_layouts   layout) {
+    zdnn_init_pre_transformed_desc(
+        layout,
+        ggml_zdnn_type_mapping(src->type),
+        &pre_tfm_desc,
+        ne[3], ne[2], ne[1], ne[0]
+    );
+
+    ZDNN_CHECK(zdnn_generate_transformed_desc(&pre_tfm_desc, &tfm_desc));
+    ZDNN_CHECK(zdnn_init_ztensor_with_malloc(&pre_tfm_desc, &tfm_desc, &ztensor));
+}
+
+inline void ggml_zdnn_load_tensor(zdnn_ztensor & ztensor,
+                                          void * buffer) {
+    ZDNN_CHECK(zdnn_transform_ztensor(&ztensor, buffer));
+}
+
+inline void ggml_zdnn_init_tensor(ggml_backend_zdnn_buffer * buffer, const ggml_tensor * tensor) {
+    switch (tensor->op) {
+        case GGML_OP_MUL_MAT:
+            {
+                zdnn_init_pre_transformed_desc(
+                    ZDNN_2D,
+                    ggml_zdnn_type_mapping(tensor->type),
+                    &buffer->pre_tfm_desc,
+                    tensor->ne[1], tensor->ne[0]
+                );
+            } break;
+
+        default:
+            {
+                // For 4D tensors, GGML uses NCHW layout. However, because zDNN
+                // automatically transforms everything to NHWC, we will use it
+                // directly to avoid the performance penalty changing the
+                // layout and reshaping the tensor.
+                zdnn_init_pre_transformed_desc(
+                    ZDNN_NHWC,
+                    ggml_zdnn_type_mapping(tensor->type),
+                    &buffer->pre_tfm_desc,
+                    tensor->ne[3], tensor->ne[2], tensor->ne[1], tensor->ne[0]
+                );
+
+                // TODO: Consider adding a ggml check.
+                // TODO: If tensor = 4D, use ZDNN_NCHW by default.
+                // TODO: If tensor = 2D, use ZDNN_NHWC by default.
+            } break;
+    }
+
+    ZDNN_CHECK(zdnn_generate_transformed_desc(&buffer->pre_tfm_desc, &buffer->tfm_desc));
+    ZDNN_CHECK(zdnn_init_ztensor_with_malloc(&buffer->pre_tfm_desc, &buffer->tfm_desc, &buffer->ztensor));
+}
+
+static void ggml_zdnn_mul_mat_op(ggml_backend_zdnn_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_TENSOR_BINARY_OP_LOCALS;
+
+    const enum ggml_type type = src0->type;
+
+    GGML_ASSERT(ne0 == ne01);
+    GGML_ASSERT(ne1 == ne11);
+    GGML_ASSERT(ne2 == ne12);
+    GGML_ASSERT(ne3 == ne13);
+
+    // we don't support permuted src0 or src1
+    GGML_ASSERT(nb00 == ggml_type_size(type));
+    GGML_ASSERT(nb10 == ggml_type_size(src1->type));
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    GGML_ASSERT(nb0 <= nb1);
+    GGML_ASSERT(nb1 <= nb2);
+    GGML_ASSERT(nb2 <= nb3);
+
+    const ggml_tensor * weights = src0;
+    const ggml_tensor * inputs  = src1;
+          ggml_tensor * output  = dst;
+
+    ggml_backend_zdnn_buffer * weights_extra = (ggml_backend_zdnn_buffer *)weights->extra;
+    ggml_backend_zdnn_buffer * inputs_extra  = (ggml_backend_zdnn_buffer *)inputs->extra;
+    ggml_backend_zdnn_buffer * output_extra  = (ggml_backend_zdnn_buffer *)output->extra;
+
+    zdnn_tensor_desc ptd_bias, td_bias;
+    zdnn_ztensor zt_bias;
+
+    const int64_t weights_rows = ne01;
+    const int64_t weights_cols = ne00;
+    const int64_t inputs_rows  = ne11;
+    const int64_t inputs_cols  = ne10;
+
+    assert(inputs_cols == weights_cols);
+
+    const int64_t output_rows = ne1;
+    const int64_t output_cols = ne0;
+
+    const int64_t bias_dim  [GGML_MAX_DIMS]  = { 1, 1, 1, output_cols };
+    ggml_zdnn_create_tensor(ptd_bias, td_bias, zt_bias, output, bias_dim, ZDNN_1D);
+
+    void * bias_data = (void *)calloc(ne0, ggml_element_size(output));
+    if (weights_extra->ztensor.is_transformed == false) ggml_zdnn_load_tensor(weights_extra->ztensor, weights->data);
+    if (inputs_extra->ztensor.is_transformed == false) ggml_zdnn_load_tensor(inputs_extra->ztensor, inputs->data);
+    ggml_zdnn_load_tensor(zt_bias, bias_data);
+
+    // GGML_LOG_INFO("%s: tensor '%s' tensor dimensions: [%ld, %ld, %ld, %ld] pre_tfm_desc dimensions: [%ld, %ld, %ld, %ld]\n",
+    //               __func__, weights_extra->name,
+    //               weights->ne[3], weights->ne[2], weights->ne[1], weights->ne[0],
+    //               weights_extra->pre_tfm_desc.dim1,
+    //               weights_extra->pre_tfm_desc.dim2,
+    //               weights_extra->pre_tfm_desc.dim3,
+    //               weights_extra->pre_tfm_desc.dim4);
+
+    // GGML_LOG_INFO("%s: tensor '%s' tensor dimensions: [%ld, %ld, %ld, %ld] pre_tfm_desc dimensions: [%ld, %ld, %ld, %ld]\n",
+    //               __func__, inputs_extra->name,
+    //               inputs->ne[3], inputs->ne[2], inputs->ne[1], inputs->ne[0],
+    //               inputs_extra->pre_tfm_desc.dim1,
+    //               inputs_extra->pre_tfm_desc.dim2,
+    //               inputs_extra->pre_tfm_desc.dim3,
+    //               inputs_extra->pre_tfm_desc.dim4);
+
+    GGML_ASSERT(weights_extra->pre_tfm_desc.dim1 == weights->ne[0] && "weights_extra->pre_tfm_desc.dim1 must match weights->ne[0]");
+    GGML_ASSERT(weights_extra->pre_tfm_desc.dim2 == weights->ne[1] && "weights_extra->pre_tfm_desc.dim2 must match weights->ne[1]");
+    GGML_ASSERT(inputs_extra->pre_tfm_desc.dim1  == inputs->ne[0]  && "inputs_extra->pre_tfm_desc.dim1 must match inputs->ne[0]");
+    GGML_ASSERT(inputs_extra->pre_tfm_desc.dim2  == inputs->ne[1]  && "inputs_extra->pre_tfm_desc.dim2 must match inputs->ne[1]");
+
+    ZDNN_CHECK(zdnn_matmul_transpose_op(&inputs_extra->ztensor, &weights_extra->ztensor, &zt_bias,
+                                        false, true, MATMUL_OP_ADDITION, &output_extra->ztensor));
+    // TODO: Remove in the future as we are currently DLF16 -> FP32 then in the next op, FP32 -> DLF16 again. Inefficient.
+    ZDNN_CHECK(zdnn_transform_origtensor(&output_extra->ztensor, output->data));
+
+    ZDNN_CHECK(zdnn_free_ztensor_buffer(&zt_bias));
+    free(bias_data);
+}
+
+static void ggml_zdnn_mul_mat_dispatch(ggml_backend_zdnn_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    bool use_mul_mat_vec =
+        (src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_F16)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
+        && src0->ne[0] % 2 == 0 && src1->ne[1] == 1;
+
+    bool use_mul_mat_vec_q =
+        ggml_is_quantized(src0->type)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
+
+    bool use_mul_mat_q =
+        ggml_is_quantized(src0->type)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
+
+    // debug helpers
+    // GGML_LOG_INFO("%s: use_mul_mat_vec   = %d\n", __func__, use_mul_mat_vec);
+    // GGML_LOG_INFO("%s: use_mul_mat_vec_q = %d\n", __func__, use_mul_mat_vec_q);
+    // GGML_LOG_INFO("%s: use_mul_mat_q     = %d\n", __func__, use_mul_mat_q);
+    // GGML_LOG_INFO("%s: src0: %8d %8d %8d %8d\n", __func__, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3]);
+    // GGML_LOG_INFO("%s:       %8d %8d %8d %8d\n", __func__, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3]);
+    // GGML_LOG_INFO("%s: src1: %8d %8d %8d %8d\n", __func__, src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3]);
+    // GGML_LOG_INFO("%s:       %8d %8d %8d %8d\n", __func__, src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3]);
+    // GGML_LOG_INFO("%s: src0 is contiguous %d, transposed %d, type = %s, name = %s\n", __func__, ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
+    // GGML_LOG_INFO("%s: src1 is contiguous %d, transposed %d, type = %s, name = %s\n", __func__, ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
+
+    if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16
+        && !ggml_is_transposed(src0) && !ggml_is_transposed(src1)
+        && src1->ne[2] * src1->ne[3] > 1) {
+        // general KQ + KQV multi-batch
+        GGML_LOG_INFO("%s: using zdnn_mul_mat_batched for KQ + KQV multi-batch\n", __func__);
+        // ggml_zdnn_mul_mat_batched(ctx, src0, src1, dst);
+    } else if (use_mul_mat_vec) {
+        GGML_LOG_INFO("%s: using zdnn_op_mul_mat_vec for vector multiplication\n", __func__);
+        // ggml_zdnn_op_mul_mat(ctx, src0, src1, dst, ggml_zdnn_op_mul_mat_vec, nullptr);
+    } else if (use_mul_mat_vec_q) {
+        GGML_LOG_INFO("%s: using zdnn_op_mul_mat_vec_q for quantized vector multiplication\n", __func__);
+        // ggml_zdnn_op_mul_mat(ctx, src0, src1, dst, ggml_zdnn_op_mul_mat_vec_q, ggml_zdnn_quantize_row_q8_1);
+    } else if (use_mul_mat_q) {
+        GGML_LOG_INFO("%s: using zdnn_op_mul_mat_q for quantized matrix multiplication\n", __func__);
+        // ggml_zdnn_op_mul_mat(ctx, src0, src1, dst, ggml_zdnn_op_mul_mat_q, ggml_zdnn_quantize_mmq_q8_1);
+    } else {
+        // GGML_LOG_INFO("%s: using zdnn_op_mul_mat for general matrix multiplication\n", __func__);
+        ggml_zdnn_mul_mat_op(ctx, src0, src1, dst);
+    }
+}
+
+static bool ggml_zdnn_compute_forward(ggml_backend_zdnn_context * ctx, ggml_tensor * dst) {
+    switch (dst->op) {
+        case GGML_OP_MUL_MAT:
+            ggml_zdnn_mul_mat_dispatch(ctx, dst->src[0], dst->src[1], dst);
+            break;
+
+        default:
+            return false;
+    }
+
+    return true;
+}
+
+static enum ggml_status ggml_zdnn_graph_compute(ggml_backend_t backend, ggml_cgraph * gf) {
+    ggml_backend_zdnn_context        * ctx     = (       ggml_backend_zdnn_context *)backend->context;
+    ggml_backend_zdnn_device_context * ctx_dev = (ggml_backend_zdnn_device_context *)backend->device->context;
+
+    ctx->gf = gf;
+    for (int i = 0; i < gf->n_nodes; i++) {
+        ggml_tensor * node = gf->nodes[i];
+
+        if (ggml_is_empty(node)
+            || node->op == GGML_OP_NONE
+            || node->op == GGML_OP_RESHAPE
+            || node->op == GGML_OP_VIEW
+            || node->op == GGML_OP_PERMUTE
+            || node->op == GGML_OP_TRANSPOSE) {
+            continue;
+        }
+
+        bool ok = ggml_zdnn_compute_forward(ctx, node);
+        if (!ok) {
+            GGML_LOG_ERROR("%s: unsupported op %s (%s)\n",
+                           __func__, node->name, ggml_op_name(node->op));
+        }
+
+        GGML_ASSERT(ok);
+    }
+
+    return GGML_STATUS_SUCCESS;
+}
+
+static bool ggml_zdnn_supports_op(const ggml_backend_zdnn_device_context * ctx_dev, const ggml_tensor * op) {
+    switch (op->op) {
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_PERMUTE:
+            return true;
+
+        case GGML_OP_MUL_MAT:
+            {
+                const ggml_tensor * src0 = op->src[0];
+                const ggml_tensor * src1 = op->src[1];
+
+                const int64_t ne10 = src1->ne[0];
+                const int64_t ne0 = op->ne[0];
+                const int64_t ne1 = op->ne[1];
+
+                const int64_t max_batch = ctx_dev->max_size;
+
+                return ggml_is_matrix(src0) &&
+                       ggml_is_matrix(src1) &&
+                       ggml_is_contiguous(src0) &&
+                       ggml_is_contiguous(src1) &&
+                       src0->view_src == nullptr && src1->view_src == nullptr &&
+                       src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 &&
+                       (ne0 <= max_batch && ne1 <= max_batch && ne10 <= max_batch);
+            } break;
+
+        default:
+            return false;
+    }
+}
+
+////////////////////////////////////////////////////////////////////////////////
+
+//
+// globals
+//
+
+// initialised in ggml_backend_zdnn_reg
+static ggml_backend_reg    g_ggml_backend_zdnn_reg;
+static ggml_backend_device g_ggml_backend_zdnn_device;
+
+static ggml_backend_zdnn_device_context g_ggml_ctx_dev_main = {
+    /* .zdnn_device           = */ 0,
+    /* .zdnn_device_ref_count = */ 0,
+    /* .has_parmblkformat_0   = */ false,
+    /* .has_parmblkformat_1   = */ false,
+    /* .max_size              = */ 0,
+    /* .name                  = */ "",
+};
+
+static int ggml_backend_zdnn_device_acq(ggml_backend_zdnn_device_context * ctx) {
+    assert(ctx != NULL);
+
+    if (ctx->zdnn_device == 0) {
+        ctx->zdnn_device = 1;
+    }
+
+    if (ctx->zdnn_device >= 1) {
+        ctx->has_parmblkformat_0 = zdnn_is_nnpa_parmblk_fmt_installed(1, NNPA_PARMBLKFORMAT_0);
+        ctx->has_parmblkformat_1 = zdnn_is_nnpa_parmblk_fmt_installed(1, NNPA_PARMBLKFORMAT_1);
+        ctx->max_size = zdnn_get_nnpa_max_dim_idx_size();
+        strncpy(ctx->name, GGML_ZDNN_NAME, sizeof(ctx->name) - 1);
+    }
+
+    ctx->zdnn_device_ref_count++;
+    return ctx->zdnn_device;
+}
+
+static void ggml_backend_zdnn_device_rel(ggml_backend_zdnn_device_context * ctx) {
+    assert(ctx != NULL);
+    assert(ctx->zdnn_device_ref_count > 0);
+
+    ctx->zdnn_device_ref_count--;
+    if (ctx->zdnn_device_ref_count == 0) {
+        if (ctx->zdnn_device >= 0) {
+            ctx->zdnn_device = 0;
+        }
+    }
+}
+
+static ggml_backend_zdnn_context * ggml_zdnn_init(ggml_backend_dev_t dev) {
+    GGML_LOG_INFO("%s: allocating\n", __func__);
+    GGML_LOG_INFO("%s: found 1 device\n", __func__);
+
+    #ifdef STATIC_LIB
+    zdnn_init();
+    #endif
+
+    ggml_backend_zdnn_context * ctx = new ggml_backend_zdnn_context();
+    ggml_backend_zdnn_device_context * ctx_dev = (ggml_backend_zdnn_device_context *)dev->context;
+
+    int device = 1;
+    GGML_LOG_INFO("%s: picking default device: %s\n", __func__, ctx_dev->name);
+
+    ctx->device = device;
+    GGML_LOG_INFO("%s: NNPA name: %s\n", __func__, ctx_dev->name);
+    GGML_LOG_INFO("%s: NNPA_PARMBLKFORMAT_0 = %s\n", __func__, ctx_dev->has_parmblkformat_0 ? "true" : "false");
+    GGML_LOG_INFO("%s: NNPA_PARMBLKFORMAT_1 = %s\n", __func__, ctx_dev->has_parmblkformat_1 ? "true" : "false");
+
+    ctx->gf = nullptr;
+
+    return ctx;
+}
+
+static void ggml_zdnn_free(ggml_backend_zdnn_context * ctx) {
+    GGML_LOG_INFO("%s: deallocating\n", __func__);
+    delete ctx;
+}
+
+//
+// backend interface
+//
+
+static void ggml_backend_zdnn_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_backend_zdnn_buffer_context * ctx = (ggml_backend_zdnn_buffer_context *)buffer->context;
+
+    for (int i = 0; i < ctx->n_buffers; i++) {
+        if (ctx->buffers[i]->ztensor.buffer != NULL && ctx->buffers[i]->ztensor.is_transformed) {
+            ZDNN_CHECK(zdnn_free_ztensor_buffer(&ctx->buffers[i]->ztensor));
+        }
+    }
+
+    delete ctx;
+}
+
+static void * ggml_backend_zdnn_buffer_get_base(ggml_backend_buffer_t buffer) {
+    ggml_backend_zdnn_buffer_context * ctx = (ggml_backend_zdnn_buffer_context *)buffer->context;
+    return ctx->all_data;
+}
+
+static enum ggml_status ggml_backend_zdnn_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
+    if (tensor->view_src != NULL) {
+        assert(tensor->view_src->buffer->buft == buffer->buft);
+        return GGML_STATUS_SUCCESS;
+    }
+
+    ggml_backend_zdnn_buffer_context * ctx = (ggml_backend_zdnn_buffer_context *)buffer->context;
+
+    const int64_t tsize = ggml_nbytes(tensor);
+    int buffer_idx = ctx->n_buffers;
+
+    std::unique_ptr<ggml_backend_zdnn_buffer> zdnn_buffer = std::make_unique<ggml_backend_zdnn_buffer>();
+    zdnn_buffer->data = tensor->data;
+    zdnn_buffer->size = tsize;
+    strncpy(zdnn_buffer->name, tensor->name, GGML_MAX_NAME - 1);
+
+    ggml_zdnn_init_tensor(zdnn_buffer.get(), tensor);
+    tensor->extra = zdnn_buffer.get();
+
+    ctx->buffers.push_back(std::move(zdnn_buffer));
+    ctx->n_buffers++;
+
+    // GGML_LOG_INFO("%s: initialised tensor '%s' in buffer %d, size = %8.2f MiB\n",
+    //               __func__, tensor->name, buffer_idx, tsize);
+
+    return GGML_STATUS_SUCCESS;
+}
+
+static void ggml_backend_zdnn_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+    memset((char *)tensor->data + offset, value, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_zdnn_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    memcpy((char *)tensor->data + offset, data, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_zdnn_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    memcpy(data, (const char *)tensor->data + offset, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_zdnn_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    ggml_backend_zdnn_buffer_context * ctx = (ggml_backend_zdnn_buffer_context *)buffer->context;
+
+    memset(ctx->all_data, value, ctx->all_size);
+}
+
+static ggml_backend_buffer_i ggml_backend_zdnn_buffer_i = {
+    /* .free_buffer   = */ ggml_backend_zdnn_buffer_free_buffer,
+    /* .get_base      = */ ggml_backend_zdnn_buffer_get_base,
+    /* .init_tensor   = */ ggml_backend_zdnn_buffer_init_tensor,
+    /* .memset_tensor = */ ggml_backend_zdnn_buffer_memset_tensor,
+    /* .set_tensor    = */ ggml_backend_zdnn_buffer_set_tensor,
+    /* .get_tensor    = */ ggml_backend_zdnn_buffer_get_tensor,
+    /* .cpy_tensor    = */ NULL,
+    /* .clear         = */ ggml_backend_zdnn_buffer_clear,
+    /* .reset         = */ NULL,
+};
+
+//
+// default buffer type
+//
+
+static const char * ggml_backend_zdnn_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return GGML_ZDNN_NAME;
+
+    GGML_UNUSED(buft);
+}
+
+static ggml_backend_buffer_t ggml_backend_zdnn_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    ggml_backend_zdnn_buffer_context * ctx = new ggml_backend_zdnn_buffer_context();
+
+    const size_t size_page = sysconf(_SC_PAGESIZE);
+
+    size_t size_aligned = size;
+    if ((size_aligned % size_page) != 0) {
+        size_aligned += size_page - (size_aligned % size_page);
+    }
+
+    ggml_backend_zdnn_device_context * ctx_dev = (ggml_backend_zdnn_device_context *)buft->device->context;
+
+    GGML_ASSERT(ctx_dev->zdnn_device >= 0);
+    int device = ctx_dev->zdnn_device; GGML_UNUSED(device);
+
+    ctx->all_data  = ggml_aligned_malloc(size_aligned);
+    ctx->all_size  = size_aligned;
+    ctx->owned     = true;
+    ctx->n_buffers = 1;
+
+    if (ctx->all_data != NULL) {
+        std::unique_ptr<ggml_backend_zdnn_buffer> zdnn_buffer = std::make_unique<ggml_backend_zdnn_buffer>();
+        zdnn_buffer->data = ctx->all_data;
+        zdnn_buffer->size = size_aligned;
+        ctx->buffers.push_back(std::move(zdnn_buffer));
+    }
+
+    if (size_aligned > 0 && (ctx->all_data == NULL)) {
+        GGML_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f\n",
+                       __func__, size_aligned / 1024.0 / 1024.0);
+        delete ctx;
+        return NULL;
+    }
+
+    return ggml_backend_buffer_init(buft, ggml_backend_zdnn_buffer_i, ctx, size);
+}
+
+static size_t ggml_backend_zdnn_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return 256;
+
+    GGML_UNUSED(buft);
+}
+
+static bool ggml_backend_zdnn_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return true;
+
+    GGML_UNUSED(buft);
+}
+
+ggml_backend_buffer_type_t ggml_backend_zdnn_buffer_type(void) {
+    static ggml_backend_buffer_type ggml_backend_buffer_type_zdnn = {
+        /* .iface   = */ {
+            /* .get_name       = */ ggml_backend_zdnn_buffer_type_get_name,
+            /* .alloc_buffer   = */ ggml_backend_zdnn_buffer_type_alloc_buffer,
+            /* .get_alignment  = */ ggml_backend_zdnn_buffer_type_get_alignment,
+            /* .get_max_size   = */ NULL,
+            /* .get_alloc_size = */ NULL,  // defaults to ggml_nbytes
+            /* .is_host        = */ ggml_backend_zdnn_buffer_type_is_host,
+        },
+        /* .device  = */ &g_ggml_backend_zdnn_device,
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_buffer_type_zdnn;
+}
+
+static const char * ggml_backend_zdnn_buffer_from_ptr_type_get_name(ggml_backend_buffer_type_t buft) {
+    return GGML_ZDNN_NAME "_Mapped";
+
+    GGML_UNUSED(buft);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_zdnn_buffer_from_ptr_type(void) {
+    static ggml_backend_buffer_type ggml_backend_buffer_from_ptr_type_zdnn = {
+        /* .iface = */ {
+            /* .get_name       = */ ggml_backend_zdnn_buffer_from_ptr_type_get_name,
+            /* .alloc_buffer   = */ ggml_backend_zdnn_buffer_type_alloc_buffer,
+            /* .get_alignment  = */ ggml_backend_zdnn_buffer_type_get_alignment,
+            /* .get_max_size   = */ NULL,
+            /* .get_alloc_size = */ NULL,  // defaults to ggml_nbytes
+            /* .is_host        = */ ggml_backend_zdnn_buffer_type_is_host,
+        },
+        /* .device  = */ &g_ggml_backend_zdnn_device,
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_buffer_from_ptr_type_zdnn;
+}
+
+//
+// backend
+//
+
+static const char * ggml_backend_zdnn_name(ggml_backend_t backend) {
+    return GGML_ZDNN_NAME;
+
+    GGML_UNUSED(backend);
+}
+
+static void ggml_backend_zdnn_free(ggml_backend_t backend) {
+    ggml_backend_zdnn_context * ctx = (ggml_backend_zdnn_context *)backend->context;
+
+    ggml_zdnn_free(ctx);
+    free(backend);
+}
+
+static enum ggml_status ggml_backend_zdnn_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
+    return ggml_zdnn_graph_compute(backend, cgraph);
+}
+
+static ggml_backend_i ggml_backend_zdnn_i = {
+    /* .get_name           = */ ggml_backend_zdnn_name,
+    /* .free               = */ ggml_backend_zdnn_free,
+    /* .set_tensor_async   = */ NULL,
+    /* .get_tensor_async   = */ NULL,
+    /* .cpy_tensor_async   = */ NULL,
+    /* .synchronize        = */ NULL,
+    /* .graph_plan_create  = */ NULL,
+    /* .graph_plan_free    = */ NULL,
+    /* .graph_plan_update  = */ NULL,
+    /* .graph_plan_compute = */ NULL,
+    /* .graph_compute      = */ ggml_backend_zdnn_graph_compute,
+    /* .event_record       = */ NULL,
+    /* .event_wait         = */ NULL,
+};
+
+static ggml_guid_t ggml_backend_zdnn_guid(void) {
+    static const char * guid_str = "IBM-ZDNN-ACCELER";
+    return reinterpret_cast<ggml_guid_t>((void *)guid_str);
+}
+
+// TODO: remove in the future
+ggml_backend_t ggml_backend_zdnn_init(void) {
+    ggml_backend_dev_t dev = ggml_backend_reg_dev_get(ggml_backend_zdnn_reg(), 0);
+
+    ggml_backend_zdnn_context * ctx = ggml_zdnn_init(dev);
+    if (ctx == NULL) {
+        GGML_LOG_ERROR("%s: error: failed to allocate context\n", __func__);
+        return NULL;
+    }
+
+    ggml_backend_t backend = (ggml_backend_t)malloc(sizeof(ggml_backend));
+    *backend = (ggml_backend) {
+        /* .guid       = */ ggml_backend_zdnn_guid(),
+        /* .iface      = */ ggml_backend_zdnn_i,
+        /* .device     = */ dev,
+        /* .context    = */ ctx,
+    };
+
+    return backend;
+}
+
+bool ggml_backend_is_zdnn(ggml_backend_t backend) {
+    return backend != NULL &&
+           ggml_guid_matches(backend->guid, ggml_backend_zdnn_guid());
+
+    GGML_UNUSED(backend);
+}
+
+//
+// backend device
+//
+
+static const char * ggml_backend_zdnn_device_get_name(ggml_backend_dev_t dev) {
+    return GGML_ZDNN_NAME;
+
+    GGML_UNUSED(dev);
+}
+
+static const char * ggml_backend_zdnn_device_get_description(ggml_backend_dev_t dev) {
+    return "IBM Z Neural Network Processing Assist (NNPA)";
+}
+
+static void ggml_backend_zdnn_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    *free  = 0;
+    *total = 0;
+}
+
+static enum ggml_backend_dev_type ggml_backend_zdnn_device_get_type(ggml_backend_dev_t dev) {
+    return GGML_BACKEND_DEVICE_TYPE_ACCEL;
+
+    GGML_UNUSED(dev);
+}
+
+static void ggml_backend_zdnn_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_zdnn_device_get_name(dev);
+    props->description = ggml_backend_zdnn_device_get_description(dev);
+    props->type        = ggml_backend_zdnn_device_get_type(dev);
+    ggml_backend_zdnn_device_get_memory(dev, &props->memory_free, &props->memory_total);
+    props->caps = (ggml_backend_dev_caps) {
+        /* .async                = */ false,
+        /* .host_buffer          = */ false,
+        /* .buffer_from_host_ptr = */ true,
+        /* .events               = */ false,
+    };
+}
+
+static ggml_backend_t ggml_backend_zdnn_device_init(ggml_backend_dev_t dev, const char * params) {
+    ggml_backend_zdnn_context * ctx = ggml_zdnn_init(dev);
+    if (ctx == NULL) {
+        GGML_LOG_ERROR("%s: error: failed to allocate context\n", __func__);
+        return NULL;
+    }
+
+    ggml_backend_t backend = (ggml_backend *)malloc(sizeof(ggml_backend));
+    *backend = (ggml_backend) {
+        /* .guid       = */ ggml_backend_zdnn_guid(),
+        /* .iface      = */ ggml_backend_zdnn_i,
+        /* .device     = */ dev,
+        /* .context    = */ ctx,
+    };
+
+    return backend;
+
+    GGML_UNUSED(params);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_zdnn_device_get_buffer_type(ggml_backend_dev_t dev) {
+    return ggml_backend_zdnn_buffer_type();
+
+    GGML_UNUSED(dev);
+}
+
+static ggml_backend_buffer_t ggml_backend_zdnn_device_buffer_from_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
+    ggml_backend_zdnn_buffer_context * ctx = new ggml_backend_zdnn_buffer_context();
+
+    ctx->all_data  = ptr;
+    ctx->all_size  = size;
+    ctx->owned     = false;
+    ctx->n_buffers = 0;
+
+    const size_t size_page = sysconf(_SC_PAGESIZE);
+
+    // page-align the data ptr
+    {
+        const uintptr_t offs = (uintptr_t) ptr % size_page;
+        ptr  = (void *)((char *)ptr - offs);
+        size += offs;
+    }
+
+    size_t size_aligned = size;
+    if ((size_aligned % size_page) != 0) {
+        size_aligned += size_page - (size_aligned % size_page);
+    }
+
+    ggml_backend_zdnn_device_context * ctx_dev = (ggml_backend_zdnn_device_context *)dev->context;
+
+    GGML_ASSERT(ctx_dev->zdnn_device >= 0);
+    int device = ctx_dev->zdnn_device; GGML_UNUSED(device);
+
+    std::unique_ptr<ggml_backend_zdnn_buffer> zdnn_buffer = std::make_unique<ggml_backend_zdnn_buffer>();
+    zdnn_buffer->data = ptr;
+    zdnn_buffer->size = size;
+    ctx->buffers.push_back(std::move(zdnn_buffer));
+
+    GGML_LOG_INFO("%s: allocated buffer, size = %8.2f MiB\n",
+                  __func__, size_aligned / 1024.0 / 1024.0);
+
+    ++ctx->n_buffers;
+
+    return ggml_backend_buffer_init(ggml_backend_zdnn_buffer_from_ptr_type(), ggml_backend_zdnn_buffer_i, ctx, size);
+}
+
+static bool ggml_backend_zdnn_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
+    ggml_backend_zdnn_device_context * ctx_dev = (ggml_backend_zdnn_device_context *) dev->context;
+
+    return ggml_zdnn_supports_op(ctx_dev, op);
+}
+
+static bool ggml_backend_zdnn_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    return
+        buft->iface.get_name == ggml_backend_zdnn_buffer_type_get_name ||
+        buft->iface.get_name == ggml_backend_zdnn_buffer_from_ptr_type_get_name;
+
+    GGML_UNUSED(dev);
+}
+
+static ggml_backend_device_i ggml_backend_zdnn_device_i = {
+    /* .get_name             = */ ggml_backend_zdnn_device_get_name,
+    /* .get_description      = */ ggml_backend_zdnn_device_get_description,
+    /* .get_memory           = */ ggml_backend_zdnn_device_get_memory,
+    /* .get_type             = */ ggml_backend_zdnn_device_get_type,
+    /* .get_props            = */ ggml_backend_zdnn_device_get_props,
+    /* .init_backend         = */ ggml_backend_zdnn_device_init,
+    /* .get_buffer_type      = */ ggml_backend_zdnn_device_get_buffer_type,
+    /* .get_host_buffer_type = */ NULL,
+    /* .buffer_from_host_ptr = */ ggml_backend_zdnn_device_buffer_from_ptr,
+    /* .supports_op          = */ ggml_backend_zdnn_device_supports_op,
+    /* .supports_buft        = */ ggml_backend_zdnn_device_supports_buft,
+    /* .offload_op           = */ NULL,
+    /* .event_new            = */ NULL,
+    /* .event_free           = */ NULL,
+    /* .event_synchronize    = */ NULL,
+};
+
+//
+// backend registry
+//
+
+static const char * ggml_backend_zdnn_reg_get_name(ggml_backend_reg_t reg) {
+    return GGML_ZDNN_NAME;
+
+    GGML_UNUSED(reg);
+}
+
+static size_t ggml_backend_zdnn_reg_device_count(ggml_backend_reg_t reg) {
+    if (!zdnn_is_nnpa_installed()) {
+        return 0;
+    }
+    return 1;
+
+    GGML_UNUSED(reg);
+}
+
+static ggml_backend_dev_t ggml_backend_zdnn_reg_device_get(ggml_backend_reg_t reg, size_t index) {
+    GGML_ASSERT(index == 0);
+
+    return &g_ggml_backend_zdnn_device;
+
+    GGML_UNUSED(reg);
+    GGML_UNUSED(index);
+}
+
+static ggml_backend_feature g_ggml_backend_zdnn_features[] = {
+    { "NNPA", zdnn_is_nnpa_installed() ? "1" : "0" },
+    { "NNPA_PARMBLKFORMAT_0", zdnn_is_nnpa_parmblk_fmt_installed(1, NNPA_PARMBLKFORMAT_0) ? "1" : "0" },
+    { "NNPA_PARMBLKFORMAT_1", zdnn_is_nnpa_parmblk_fmt_installed(1, NNPA_PARMBLKFORMAT_1) ? "1" : "0" },
+    { NULL, NULL },
+};
+
+static ggml_backend_feature * ggml_backend_zdnn_get_features(ggml_backend_reg_t reg) {
+    return g_ggml_backend_zdnn_features;
+
+    GGML_UNUSED(reg);
+}
+
+static void * ggml_backend_zdnn_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    if (strcmp(name, "ggml_backend_get_features") == 0) {
+        return (void *) ggml_backend_zdnn_get_features;
+    }
+
+    return NULL;
+
+    GGML_UNUSED(reg);
+}
+
+static ggml_backend_reg_i ggml_backend_zdnn_reg_i = {
+    /* .get_name         = */ ggml_backend_zdnn_reg_get_name,
+    /* .get_device_count = */ ggml_backend_zdnn_reg_device_count,
+    /* .get_device       = */ ggml_backend_zdnn_reg_device_get,
+    /* .get_proc_address = */ ggml_backend_zdnn_get_proc_address,
+};
+
+static void ggml_zdnn_cleanup(void) {
+    ggml_backend_zdnn_device_rel(&g_ggml_ctx_dev_main);
+}
+
+// TODO: make thread-safe
+ggml_backend_reg_t ggml_backend_zdnn_reg(void) {
+    ggml_backend_zdnn_device_acq(&g_ggml_ctx_dev_main);
+
+    // register cleanup callback
+    atexit(ggml_zdnn_cleanup);
+
+    {
+        g_ggml_backend_zdnn_reg = (ggml_backend_reg) {
+            /* .api_version = */ GGML_ZDNN_VERSION,
+            /* .iface       = */ ggml_backend_zdnn_reg_i,
+            /* .context     = */ NULL,
+        };
+
+        g_ggml_backend_zdnn_device = (ggml_backend_device) {
+            /* .iface       = */ ggml_backend_zdnn_device_i,
+            /* .reg         = */ &g_ggml_backend_zdnn_reg,
+            /* .context     = */ &g_ggml_ctx_dev_main,
+        };
+
+        return &g_ggml_backend_zdnn_reg;
+    }
+}
+
+GGML_BACKEND_DL_IMPL(ggml_backend_zdnn_reg)

From 3a8f07ee4ea55e87c72bd1a59472536efdc01a05 Mon Sep 17 00:00:00 2001
From: Reese Levine <reeselevine1@gmail.com>
Date: Mon, 18 Aug 2025 19:22:37 +0300
Subject: [PATCH 84/86] ggml: Add initial WebGPU backend (llama/14521)

ggml-ci
---
 ggml/src/ggml-webgpu/CMakeLists.txt           |   54 +
 ggml/src/ggml-webgpu/ggml-webgpu.cpp          | 1190 +++++++++++++++++
 ggml/src/ggml-webgpu/wgsl-shaders/cpy.wgsl    |   60 +
 .../ggml-webgpu/wgsl-shaders/embed_wgsl.py    |   35 +
 ggml/src/ggml-webgpu/wgsl-shaders/memset.wgsl |   40 +
 .../src/ggml-webgpu/wgsl-shaders/mul_mat.wgsl |   56 +
 .../ggml-webgpu/wgsl-shaders/set_rows.wgsl    |   82 ++
 7 files changed, 1517 insertions(+)
 create mode 100644 ggml/src/ggml-webgpu/CMakeLists.txt
 create mode 100644 ggml/src/ggml-webgpu/ggml-webgpu.cpp
 create mode 100644 ggml/src/ggml-webgpu/wgsl-shaders/cpy.wgsl
 create mode 100755 ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py
 create mode 100644 ggml/src/ggml-webgpu/wgsl-shaders/memset.wgsl
 create mode 100644 ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.wgsl
 create mode 100644 ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl

diff --git a/ggml/src/ggml-webgpu/CMakeLists.txt b/ggml/src/ggml-webgpu/CMakeLists.txt
new file mode 100644
index 00000000000..79ef68b85a4
--- /dev/null
+++ b/ggml/src/ggml-webgpu/CMakeLists.txt
@@ -0,0 +1,54 @@
+cmake_minimum_required(VERSION 3.13)
+
+find_package(Python3 REQUIRED)
+
+# Shader locations
+set(SHADER_DIR "${CMAKE_CURRENT_SOURCE_DIR}/wgsl-shaders")
+set(SHADER_OUTPUT_DIR "${CMAKE_CURRENT_BINARY_DIR}/generated")
+set(SHADER_HEADER "${SHADER_OUTPUT_DIR}/ggml-wgsl-shaders.hpp")
+file(MAKE_DIRECTORY ${SHADER_OUTPUT_DIR})
+
+message(STATUS "Shader output dir: ${SHADER_OUTPUT_DIR}")
+
+# Find all WGSL files
+file(GLOB WGSL_SHADER_FILES "${SHADER_DIR}/*.wgsl")
+
+# Generate the header using a Python script
+add_custom_command(
+    OUTPUT ${SHADER_HEADER}
+    COMMAND ${CMAKE_COMMAND} -E echo "Embedding WGSL shaders to ggml-wgsl-shaders.hpp"
+    COMMAND ${CMAKE_COMMAND} -E make_directory ${SHADER_OUTPUT_DIR}
+    COMMAND ${CMAKE_COMMAND} -E env PYTHONIOENCODING=utf-8
+        ${Python3_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/wgsl-shaders/embed_wgsl.py
+            --input "${SHADER_DIR}"
+            --output "${SHADER_HEADER}"
+    DEPENDS ${WGSL_SHADER_FILES} ${CMAKE_CURRENT_SOURCE_DIR}/wgsl-shaders/embed_wgsl.py
+    VERBATIM
+)
+
+add_custom_target(generate_shaders DEPENDS ${SHADER_HEADER})
+
+ggml_add_backend_library(ggml-webgpu
+    ggml-webgpu.cpp
+    ${SHADER_HEADER}
+    ../../include/ggml-webgpu.h
+)
+
+add_dependencies(ggml-webgpu generate_shaders)
+
+if(EMSCRIPTEN)
+    set(EMDAWNWEBGPU_DIR "" CACHE PATH "Path to emdawnwebgpu_pkg")
+
+    target_compile_options(ggml-webgpu PRIVATE "--use-port=${EMDAWNWEBGPU_DIR}/emdawnwebgpu.port.py")
+    target_link_options(ggml-webgpu PRIVATE "--use-port=${EMDAWNWEBGPU_DIR}/emdawnwebgpu.port.py")
+else()
+    find_package(Dawn REQUIRED)
+    set(DawnWebGPU_TARGET dawn::webgpu_dawn)
+endif()
+
+if (GGML_WEBGPU_DEBUG)
+    target_compile_definitions(ggml-webgpu PRIVATE GGML_WEBGPU_DEBUG=1)
+endif()
+
+target_include_directories(ggml-webgpu PRIVATE ${SHADER_OUTPUT_DIR})
+target_link_libraries(ggml-webgpu PRIVATE ${DawnWebGPU_TARGET})
diff --git a/ggml/src/ggml-webgpu/ggml-webgpu.cpp b/ggml/src/ggml-webgpu/ggml-webgpu.cpp
new file mode 100644
index 00000000000..ba1addc8d9f
--- /dev/null
+++ b/ggml/src/ggml-webgpu/ggml-webgpu.cpp
@@ -0,0 +1,1190 @@
+/*
+    WebGPU backend implementation.
+    Note: Use ClangFormat to format this file.
+*/
+
+#include "ggml-webgpu.h"
+
+#include "ggml-backend-impl.h"
+#include "ggml-impl.h"
+#include "ggml-wgsl-shaders.hpp"
+
+#include <webgpu/webgpu_cpp.h>
+
+#include <condition_variable>
+#include <cstring>
+#include <iostream>
+#include <mutex>
+#include <string>
+#include <vector>
+
+#ifdef GGML_WEBGPU_DEBUG
+#    define WEBGPU_LOG_DEBUG(msg)  std::cout << msg << std::endl
+#    define WEBGPU_DEBUG_BUF_ELEMS 32
+#else
+#    define WEBGPU_LOG_DEBUG(msg) ((void) 0)
+#endif  // GGML_WEBGPU_DEBUG
+
+/* Constants */
+
+#define WEBGPU_COMMAND_SUBMIT_BATCH_SIZE     16
+#define WEBGPU_MUL_MAT_WG_SIZE               64
+#define WEBGPU_NUM_PARAM_BUFS                100
+#define WEBGPU_PARAMS_BUF_SIZE_BYTES         128  // enough for 32 parameters
+#define WEBGPU_NUM_SET_ROWS_ERROR_BUFS       32
+#define WEBGPU_SET_ROWS_ERROR_BUF_SIZE_BYTES 4
+#define WEBGPU_STORAGE_BUF_BINDING_MULT      4  // a storage buffer binding size must be a multiple of 4
+
+/* End Constants */
+
+// This is a "fake" base pointer, since WebGPU buffers do not have pointers to their locations.
+static void * const webgpu_ptr_base = (void *) (uintptr_t) 0x1000;  // NOLINT
+
+// Always returns the base offset of a tensor, regardless of views.
+static uint64_t webgpu_tensor_offset(const ggml_tensor * tensor) {
+    if (tensor->view_src) {
+        return (uint8_t *) tensor->view_src->data - (uint8_t *) webgpu_ptr_base;
+    }
+    return (uint8_t *) tensor->data - (uint8_t *) webgpu_ptr_base;
+}
+
+/* Struct definitions */
+
+// Forward reference
+static void ggml_webgpu_create_buffer(wgpu::Device &    device,
+                                      wgpu::Buffer &    buffer,
+                                      size_t            size,
+                                      wgpu::BufferUsage usage,
+                                      const char *      label);
+
+struct webgpu_pool_bufs {
+    wgpu::Buffer host_buf;
+    wgpu::Buffer dev_buf;
+};
+
+// Holds a pool of parameter buffers for WebGPU operations
+struct webgpu_buf_pool {
+    std::vector<webgpu_pool_bufs> free;
+
+    std::mutex mutex;
+
+    std::condition_variable cv;
+
+    void init(wgpu::Device      device,
+              int               num_bufs,
+              size_t            buf_size,
+              wgpu::BufferUsage dev_buf_usage,
+              wgpu::BufferUsage host_buf_usage) {
+        for (int i = 0; i < num_bufs; i++) {
+            wgpu::Buffer host_buf;
+            wgpu::Buffer dev_buf;
+            ggml_webgpu_create_buffer(device, host_buf, buf_size, host_buf_usage, "ggml_webgpu_host_pool_buf");
+            ggml_webgpu_create_buffer(device, dev_buf, buf_size, dev_buf_usage, "ggml_webgpu_dev_pool_buf");
+            free.push_back({ host_buf, dev_buf });
+        }
+    }
+
+    webgpu_pool_bufs alloc_bufs() {
+        std::unique_lock<std::mutex> lock(mutex);
+        cv.wait(lock, [this] { return !free.empty(); });
+        webgpu_pool_bufs bufs = free.back();
+        free.pop_back();
+        return bufs;
+    }
+
+    void free_bufs(std::vector<webgpu_pool_bufs> bufs) {
+        std::lock_guard<std::mutex> lock(mutex);
+        free.insert(free.end(), bufs.begin(), bufs.end());
+        cv.notify_all();
+    }
+
+    void cleanup() {
+        std::lock_guard<std::mutex> lock(mutex);
+        for (auto & bufs : free) {
+            bufs.host_buf.Destroy();
+            bufs.dev_buf.Destroy();
+        }
+        free.clear();
+    }
+};
+
+// All the base objects needed to run operations on a WebGPU device
+struct webgpu_context_struct {
+    wgpu::Instance instance;
+    wgpu::Adapter  adapter;
+    wgpu::Device   device;
+    wgpu::Queue    queue;
+    wgpu::Limits   limits;
+
+    std::recursive_mutex mutex;
+
+    bool device_init = false;
+
+    webgpu_buf_pool param_buf_pool;
+    webgpu_buf_pool set_rows_error_buf_pool;
+
+    wgpu::ComputePipeline memset_pipeline;
+    wgpu::ComputePipeline mul_mat_pipeline;
+    wgpu::ComputePipeline set_rows_pipeline;
+    wgpu::ComputePipeline cpy_pipeline;
+
+    size_t memset_bytes_per_thread;
+
+    // Staging buffer for reading data from the GPU
+    wgpu::Buffer get_tensor_staging_buf;
+
+    // Command buffers which need to be submitted
+    std::vector<wgpu::CommandBuffer> staged_command_bufs;
+
+    // Parameter buffers associated with the staged command buffers
+    std::vector<webgpu_pool_bufs> staged_param_bufs;
+    // Buffers associated with set_rows operations, used to store potential errors
+    std::vector<webgpu_pool_bufs> staged_set_row_error_bufs;
+
+    std::vector<wgpu::FutureWaitInfo> callback_futures;
+
+#ifdef GGML_WEBGPU_DEBUG
+    wgpu::Buffer debug_host_buf;
+    wgpu::Buffer debug_dev_buf;
+#endif
+};
+
+typedef std::shared_ptr<webgpu_context_struct> webgpu_context;
+
+struct ggml_backend_webgpu_reg_context {
+    webgpu_context webgpu_ctx;
+    size_t         device_count;
+    const char *   name;
+};
+
+struct ggml_backend_webgpu_device_context {
+    webgpu_context webgpu_ctx;
+    std::string    device_name;
+    std::string    device_desc;
+};
+
+struct ggml_backend_webgpu_context {
+    webgpu_context webgpu_ctx;
+    std::string    name;
+};
+
+struct ggml_backend_webgpu_buffer_context {
+    webgpu_context webgpu_ctx;
+    wgpu::Buffer   buffer;
+
+    ggml_backend_webgpu_buffer_context(webgpu_context ctx, wgpu::Buffer buf) :
+        webgpu_ctx(std::move(ctx)),
+        buffer(std::move(buf)) {}
+};
+
+/* End struct definitions */
+
+/* WebGPU object initializations */
+
+static void ggml_webgpu_create_pipeline(wgpu::Device &                           device,
+                                        wgpu::ComputePipeline &                  pipeline,
+                                        const char *                             shader_code,
+                                        const char *                             label,
+                                        const std::vector<wgpu::ConstantEntry> & constants = {}) {
+    WEBGPU_LOG_DEBUG("ggml_webgpu_create_pipeline()");
+
+    wgpu::ShaderSourceWGSL shader_source;
+    shader_source.code = shader_code;
+
+    wgpu::ShaderModuleDescriptor shader_desc;
+    shader_desc.nextInChain = &shader_source;
+
+    wgpu::ShaderModule shader_module = device.CreateShaderModule(&shader_desc);
+
+    wgpu::ComputePipelineDescriptor pipeline_desc;
+    pipeline_desc.label              = label;
+    pipeline_desc.compute.module     = shader_module;
+    pipeline_desc.compute.entryPoint = "main";   // Entry point in the WGSL code
+    pipeline_desc.layout             = nullptr;  // nullptr means auto layout
+    if (constants.size() > 0) {
+        pipeline_desc.compute.constants     = constants.data();
+        pipeline_desc.compute.constantCount = constants.size();
+    }
+    pipeline = device.CreateComputePipeline(&pipeline_desc);
+}
+
+static void ggml_webgpu_create_buffer(wgpu::Device &    device,
+                                      wgpu::Buffer &    buffer,
+                                      size_t            size,
+                                      wgpu::BufferUsage usage,
+                                      const char *      label) {
+    WEBGPU_LOG_DEBUG("ggml_webgpu_create_buffer()");
+
+    wgpu::BufferDescriptor buffer_desc;
+    buffer_desc.size             = size;
+    buffer_desc.usage            = usage;
+    buffer_desc.label            = label;
+    buffer_desc.mappedAtCreation = false;
+
+    // TODO: error handling
+    buffer = device.CreateBuffer(&buffer_desc);
+}
+
+/** End WebGPU object initializations */
+
+/** WebGPU Actions */
+
+// Wait for the queue to finish processing all submitted work
+static void ggml_backend_webgpu_wait_on_submission(webgpu_context & ctx) {
+    std::lock_guard<std::recursive_mutex> lock(ctx->mutex);
+    if (ctx->callback_futures.empty()) {
+        // no existing callbacks, wait on queue submission
+        ctx->instance.WaitAny(ctx->queue.OnSubmittedWorkDone(
+                                  wgpu::CallbackMode::AllowSpontaneous,
+                                  [](wgpu::QueueWorkDoneStatus status, wgpu::StringView message) {
+                                      if (status != wgpu::QueueWorkDoneStatus::Success) {
+                                          GGML_LOG_ERROR("ggml_webgpu: Failed to submit commands: %s\n", message.data);
+                                      }
+                                  }),
+                              UINT64_MAX);
+    } else {
+        // existing callbacks, wait on them
+        ctx->instance.WaitAny(ctx->callback_futures.size(), ctx->callback_futures.data(), UINT64_MAX);
+        ctx->callback_futures.clear();
+    }
+}
+
+static void ggml_backend_webgpu_submit_queue(webgpu_context & ctx) {
+    std::lock_guard<std::recursive_mutex> lock(ctx->mutex);
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_submit_queue()");
+    if (ctx->staged_command_bufs.empty()) {
+        // Nothing to submit
+        return;
+    }
+    ctx->queue.Submit(ctx->staged_command_bufs.size(), ctx->staged_command_bufs.data());
+
+    // If there are SET_ROWS operations in this submission, copy their error buffers to the host.
+    if (ctx->staged_set_row_error_bufs.size() > 0) {
+        wgpu::CommandEncoder encoder = ctx->device.CreateCommandEncoder();
+        for (auto & error_bufs : ctx->staged_set_row_error_bufs) {
+            // Copy the error buffer to the host buffer
+            encoder.CopyBufferToBuffer(error_bufs.dev_buf, 0, error_bufs.host_buf, 0, error_bufs.host_buf.GetSize());
+        }
+        wgpu::CommandBuffer commands = encoder.Finish();
+        ctx->queue.Submit(1, &commands);
+    }
+
+    ctx->staged_command_bufs.clear();
+    std::vector<webgpu_pool_bufs> staged_param_bufs         = std::move(ctx->staged_param_bufs);
+    std::vector<webgpu_pool_bufs> staged_set_row_error_bufs = std::move(ctx->staged_set_row_error_bufs);
+
+    // Free the staged parameter buffers once the submission completes
+    wgpu::Future p_f = ctx->queue.OnSubmittedWorkDone(
+        wgpu::CallbackMode::AllowSpontaneous,
+        [ctx, staged_param_bufs](wgpu::QueueWorkDoneStatus status, wgpu::StringView message) {
+            if (status != wgpu::QueueWorkDoneStatus::Success) {
+                GGML_LOG_ERROR("ggml_webgpu: Failed to submit commands: %s\n", message.data);
+            }
+            // Free the staged buffers
+            ctx->param_buf_pool.free_bufs(staged_param_bufs);
+        });
+    ctx->callback_futures.push_back({ p_f });
+
+    // Check for errrors in SET_ROWS operations
+    for (auto & error_bufs : staged_set_row_error_bufs) {
+        wgpu::Future f = error_bufs.host_buf.MapAsync(
+            wgpu::MapMode::Read,
+            0,
+            error_bufs.host_buf.GetSize(),
+            wgpu::CallbackMode::AllowSpontaneous,
+            [ctx, error_bufs](wgpu::MapAsyncStatus status, wgpu::StringView message) {
+                if (status != wgpu::MapAsyncStatus::Success) {
+                    GGML_LOG_ERROR("ggml_webgpu: Failed to map error buffer: %s\n", message.data);
+                } else {
+                    const uint32_t * error_data = (const uint32_t *) error_bufs.host_buf.GetConstMappedRange();
+                    if (*error_data) {
+                        GGML_ABORT("ggml_webgpu: SET_ROWS index > 2^32, unsupported.");
+                    }
+                    // We can't unmap in here due to WebGPU reentrancy limitations.
+                    ctx->set_rows_error_buf_pool.free_bufs({ error_bufs });
+                }
+            });
+        ctx->callback_futures.push_back({ f });
+    }
+}
+
+static void ggml_backend_webgpu_map_buffer(webgpu_context & ctx,
+                                           wgpu::Buffer &   buffer,
+                                           wgpu::MapMode    mode,
+                                           size_t           offset,
+                                           size_t           size) {
+    ctx->instance.WaitAny(buffer.MapAsync(mode,
+                                          offset,
+                                          size,
+                                          wgpu::CallbackMode::AllowSpontaneous,
+                                          [](wgpu::MapAsyncStatus status, wgpu::StringView message) {
+                                              if (status != wgpu::MapAsyncStatus::Success) {
+                                                  GGML_LOG_ERROR("ggml_webgpu: Failed to map buffer: %s\n",
+                                                                 message.data);
+                                              }
+                                          }),
+                          UINT64_MAX);
+}
+
+#ifdef GGML_WEBGPU_DEBUG
+// This function adds debugging information to shaders, as WebGPU does not support printing directly.
+// To use, add a bind group entry to the setup for the shader you are debugging, add the buffer and
+// debug statements in the shader, and then call this function after encoding the commands and submitting them.
+static void ggml_backend_webgpu_debug(webgpu_context & ctx) {
+    wgpu::CommandEncoder encoder = ctx->device.CreateCommandEncoder();
+    encoder.CopyBufferToBuffer(ctx->debug_dev_buf, 0, ctx->debug_host_buf, 0, ctx->debug_host_buf.GetSize());
+    wgpu::CommandBuffer commands = encoder.Finish();
+    ctx->queue.Submit(1, &commands);
+
+    ggml_backend_webgpu_map_buffer(ctx, ctx->debug_host_buf, wgpu::MapMode::Read, 0, ctx->debug_host_buf.GetSize());
+    const uint32_t * debug_data = (const uint32_t *) ctx->debug_host_buf.GetConstMappedRange();
+    std::cout << "debug data:";
+    for (size_t i = 0; i < WEBGPU_DEBUG_BUF_ELEMS; i++) {
+        std::cout << "  " << i << ": " << debug_data[i];
+    }
+    std::cout << "\n";
+    ctx->debug_host_buf.Unmap();
+}
+#endif
+
+static void ggml_backend_webgpu_build_and_enqueue(webgpu_context &                  ctx,
+                                                  wgpu::ComputePipeline &           pipeline,
+                                                  std::vector<uint32_t>             params,
+                                                  std::vector<wgpu::BindGroupEntry> bind_group_entries,
+                                                  uint32_t                          wg_x,
+                                                  bool                              submit_and_wait = false) {
+    webgpu_pool_bufs params_bufs = ctx->param_buf_pool.alloc_bufs();
+
+    ggml_backend_webgpu_map_buffer(ctx, params_bufs.host_buf, wgpu::MapMode::Write, 0, params_bufs.host_buf.GetSize());
+    uint32_t * _params = (uint32_t *) params_bufs.host_buf.GetMappedRange();
+    for (size_t i = 0; i < params.size(); i++) {
+        _params[i] = params[i];
+    };
+
+    params_bufs.host_buf.Unmap();
+
+    uint32_t params_bufs_binding_num = bind_group_entries.size();
+    bind_group_entries.push_back({ .binding = params_bufs_binding_num,
+                                   .buffer  = params_bufs.dev_buf,
+                                   .offset  = 0,
+                                   .size    = params_bufs.dev_buf.GetSize() });
+
+    wgpu::BindGroupDescriptor bind_group_desc;
+    bind_group_desc.layout     = pipeline.GetBindGroupLayout(0);
+    bind_group_desc.entryCount = bind_group_entries.size();
+    bind_group_desc.entries    = bind_group_entries.data();
+    wgpu::BindGroup bind_group = ctx->device.CreateBindGroup(&bind_group_desc);
+
+    wgpu::CommandEncoder encoder = ctx->device.CreateCommandEncoder();
+    encoder.CopyBufferToBuffer(params_bufs.host_buf, 0, params_bufs.dev_buf, 0, params_bufs.dev_buf.GetSize());
+    wgpu::ComputePassEncoder pass = encoder.BeginComputePass();
+    pass.SetPipeline(pipeline);
+    pass.SetBindGroup(0, bind_group);
+    pass.DispatchWorkgroups(wg_x, 1, 1);
+    pass.End();
+    wgpu::CommandBuffer commands = encoder.Finish();
+    if (submit_and_wait) {
+        // Submit and wait immediately
+        ctx->queue.Submit(1, &commands);
+        ctx->instance.WaitAny(ctx->queue.OnSubmittedWorkDone(
+                                  wgpu::CallbackMode::AllowSpontaneous,
+                                  [ctx, params_bufs](wgpu::QueueWorkDoneStatus status, wgpu::StringView message) {
+                                      if (status != wgpu::QueueWorkDoneStatus::Success) {
+                                          GGML_LOG_ERROR("ggml_webgpu: Failed to submit commands: %s\n", message.data);
+                                      }
+                                      ctx->param_buf_pool.free_bufs({ params_bufs });
+                                  }),
+                              UINT64_MAX);
+    } else {
+        // Lock the context mutex when pushing to the staging vectors.
+        std::lock_guard<std::recursive_mutex> lock(ctx->mutex);
+        // Enqueue commands and only submit if we have enough staged commands
+        ctx->staged_command_bufs.push_back(commands);
+        ctx->staged_param_bufs.push_back(params_bufs);
+        if (ctx->staged_command_bufs.size() == WEBGPU_COMMAND_SUBMIT_BATCH_SIZE) {
+            ggml_backend_webgpu_submit_queue(ctx);
+        }
+    }
+}
+
+static void ggml_backend_webgpu_buffer_memset(webgpu_context & ctx,
+                                              wgpu::Buffer &   buf,
+                                              uint32_t         value,
+                                              size_t           offset,
+                                              size_t           size) {
+    std::vector<uint32_t>             params  = { (uint32_t) offset, (uint32_t) size, value };
+    std::vector<wgpu::BindGroupEntry> entries = {
+        { .binding = 0, .buffer = buf, .offset = 0, .size = buf.GetSize() }
+    };
+    size_t   bytes_per_wg = ctx->limits.maxComputeWorkgroupSizeX * ctx->memset_bytes_per_thread;
+    uint32_t wg_x         = ((size + 3) + bytes_per_wg - 1) / bytes_per_wg;
+    ggml_backend_webgpu_build_and_enqueue(ctx, ctx->memset_pipeline, params, entries, wg_x, true);
+}
+
+static size_t ggml_backend_webgpu_tensor_offset(const ggml_tensor * tensor) {
+    return webgpu_tensor_offset(tensor) + tensor->view_offs;
+}
+
+static wgpu::Buffer ggml_backend_webgpu_tensor_buf(const ggml_tensor * tensor) {
+    ggml_backend_webgpu_buffer_context * ctx = (ggml_backend_webgpu_buffer_context *) tensor->buffer->context;
+    return ctx->buffer;
+}
+
+/** End WebGPU Actions */
+
+/** GGML Backend Interface */
+
+static const char * ggml_backend_webgpu_name(ggml_backend_t backend) {
+    ggml_backend_webgpu_context * ctx = (ggml_backend_webgpu_context *) backend->context;
+    return ctx->name.c_str();
+}
+
+static void ggml_backend_webgpu_free(ggml_backend_t backend) {
+    ggml_backend_webgpu_context * ctx = (ggml_backend_webgpu_context *) backend->context;
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_free(" << ctx->name << ")");
+
+    // TODO: cleanup
+    GGML_UNUSED(ctx);
+}
+
+static void ggml_webgpu_cpy(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) {
+    size_t src_offset       = ggml_backend_webgpu_tensor_offset(src);
+    // assumes power of 2 offset alignment
+    size_t src_misalignment = src_offset & (ctx->limits.minStorageBufferOffsetAlignment - 1);
+    // align to minimum offset alignment
+    src_offset &= ~(ctx->limits.minStorageBufferOffsetAlignment - 1);
+    size_t dst_offset       = ggml_backend_webgpu_tensor_offset(dst);
+    size_t dst_misalignment = dst_offset & (ctx->limits.minStorageBufferOffsetAlignment - 1);
+    dst_offset &= ~(ctx->limits.minStorageBufferOffsetAlignment - 1);
+    uint32_t              ne     = (uint32_t) ggml_nelements(dst);
+    std::vector<uint32_t> params = { ne,
+                                     (uint32_t) (src_misalignment / ggml_type_size(src->type)),
+                                     (uint32_t) (dst_misalignment / ggml_type_size(dst->type)),
+                                     // Convert byte-strides to element-strides
+                                     (uint32_t) (src->nb[0] / ggml_type_size(src->type)),
+                                     (uint32_t) (src->nb[1] / ggml_type_size(src->type)),
+                                     (uint32_t) (src->nb[2] / ggml_type_size(src->type)),
+                                     (uint32_t) (src->nb[3] / ggml_type_size(src->type)),
+                                     (uint32_t) (dst->nb[0] / ggml_type_size(dst->type)),
+                                     (uint32_t) (dst->nb[1] / ggml_type_size(dst->type)),
+                                     (uint32_t) (dst->nb[2] / ggml_type_size(dst->type)),
+                                     (uint32_t) (dst->nb[3] / ggml_type_size(dst->type)),
+                                     // Logical shape — same for both tensors even if permuted
+                                     (uint32_t) src->ne[0],
+                                     (uint32_t) src->ne[1],
+                                     (uint32_t) src->ne[2],
+                                     (uint32_t) src->ne[3] };
+
+    std::vector<wgpu::BindGroupEntry> entries = {
+        { .binding = 0,
+         .buffer  = ggml_backend_webgpu_tensor_buf(src),
+         .offset  = src_offset,
+         .size    = (ggml_nbytes(src) + src_misalignment + WEBGPU_STORAGE_BUF_BINDING_MULT - 1) &
+                  ~(WEBGPU_STORAGE_BUF_BINDING_MULT - 1) },
+        { .binding = 1,
+         .buffer  = ggml_backend_webgpu_tensor_buf(dst),
+         .offset  = dst_offset,
+         .size    = (ggml_nbytes(dst) + dst_misalignment + WEBGPU_STORAGE_BUF_BINDING_MULT - 1) &
+                  ~(WEBGPU_STORAGE_BUF_BINDING_MULT - 1) }
+    };
+
+    size_t   max_wg_size = ctx->limits.maxComputeWorkgroupSizeX;
+    uint32_t wg_x        = (ne + max_wg_size - 1) / max_wg_size;
+    ggml_backend_webgpu_build_and_enqueue(ctx, ctx->cpy_pipeline, params, entries, wg_x);
+}
+
+static void ggml_webgpu_set_rows(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * idx, ggml_tensor * dst) {
+    // For set rows specifically, we need to check if src and idx are empty tensors.
+    if (ggml_is_empty(src) || ggml_is_empty(idx)) {
+        return;
+    }
+
+    webgpu_pool_bufs error_bufs = ctx->set_rows_error_buf_pool.alloc_bufs();
+    if (error_bufs.host_buf.GetMapState() == wgpu::BufferMapState::Mapped) {
+        error_bufs.host_buf.Unmap();
+    }
+
+    size_t src_offset       = ggml_backend_webgpu_tensor_offset(src);
+    // assumes power of 2 offset alignment
+    size_t src_misalignment = src_offset & (ctx->limits.minStorageBufferOffsetAlignment - 1);
+    // align to minimum offset alignment
+    src_offset &= ~(ctx->limits.minStorageBufferOffsetAlignment - 1);
+    size_t idx_offset       = ggml_backend_webgpu_tensor_offset(idx);
+    size_t idx_misalignment = idx_offset & (ctx->limits.minStorageBufferOffsetAlignment - 1);
+    idx_offset &= ~(ctx->limits.minStorageBufferOffsetAlignment - 1);
+    size_t dst_offset       = ggml_backend_webgpu_tensor_offset(dst);
+    size_t dst_misalignment = dst_offset & (ctx->limits.minStorageBufferOffsetAlignment - 1);
+    dst_offset &= ~(ctx->limits.minStorageBufferOffsetAlignment - 1);
+
+    std::vector<uint32_t> params = { (uint32_t) (src_misalignment / ggml_type_size(src->type)),
+                                     (uint32_t) (idx_misalignment / ggml_type_size(idx->type)),
+                                     (uint32_t) (dst_misalignment / ggml_type_size(dst->type)),
+                                     // Convert byte-strides to element-strides
+                                     (uint32_t) (src->nb[1] / ggml_type_size(src->type)),
+                                     (uint32_t) (src->nb[2] / ggml_type_size(src->type)),
+                                     (uint32_t) (src->nb[3] / ggml_type_size(src->type)),
+                                     (uint32_t) (idx->nb[0] / ggml_type_size(idx->type)),
+                                     (uint32_t) (idx->nb[1] / ggml_type_size(idx->type)),
+                                     (uint32_t) (idx->nb[2] / ggml_type_size(idx->type)),
+                                     (uint32_t) (dst->nb[1] / ggml_type_size(dst->type)),
+                                     (uint32_t) (dst->nb[2] / ggml_type_size(dst->type)),
+                                     (uint32_t) (dst->nb[3] / ggml_type_size(dst->type)),
+                                     // Shape of src
+                                     (uint32_t) src->ne[0],
+                                     (uint32_t) src->ne[1],
+                                     (uint32_t) src->ne[2],
+                                     (uint32_t) src->ne[3],
+                                     // Shape of idx
+                                     (uint32_t) (idx->ne[1]),
+                                     (uint32_t) (idx->ne[2]) };
+
+    std::vector<wgpu::BindGroupEntry> entries = {
+        { .binding = 0,
+         .buffer  = ggml_backend_webgpu_tensor_buf(src),
+         .offset  = ggml_backend_webgpu_tensor_offset(src),
+         .size    = ggml_nbytes(src)                                                                       },
+        { .binding = 1,
+         .buffer  = ggml_backend_webgpu_tensor_buf(idx),
+         .offset  = ggml_backend_webgpu_tensor_offset(idx),
+         .size    = ggml_nbytes(idx)                                                                       },
+        { .binding = 2,
+         .buffer  = ggml_backend_webgpu_tensor_buf(dst),
+         .offset  = ggml_backend_webgpu_tensor_offset(dst),
+         .size    = ggml_nbytes(dst)                                                                       },
+        { .binding = 3, .buffer = error_bufs.dev_buf,    .offset = 0, .size = error_bufs.dev_buf.GetSize() }
+    };
+
+    size_t   max_wg_size = ctx->limits.maxComputeWorkgroupSizeX;
+    uint32_t wg_x        = (src->ne[1] * src->ne[2] * src->ne[3] + max_wg_size - 1) / max_wg_size;
+
+    std::lock_guard<std::recursive_mutex> lock(ctx->mutex);
+    ctx->staged_set_row_error_bufs.push_back(error_bufs);
+
+    ggml_backend_webgpu_build_and_enqueue(ctx, ctx->set_rows_pipeline, params, entries, wg_x);
+}
+
+static void ggml_webgpu_mul_mat(webgpu_context & ctx, ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst) {
+    std::vector<uint32_t> params = {
+        (uint32_t) dst->ne[1],                                  // number of rows in result (M)
+        (uint32_t) dst->ne[0],                                  // number of columns in result (N)
+        (uint32_t) src0->ne[0],                                 // number of columns in src0/src1 (K)
+        (uint32_t) (src0->nb[1] / ggml_type_size(src0->type)),  // stride (elements) of src0 in dimension 1
+        (uint32_t) (src1->nb[1] / ggml_type_size(src1->type)),  // stride (elements) of src1 in dimension 1
+        (uint32_t) (src0->nb[2] / ggml_type_size(src0->type)),  // stride (elements) of src0 in dimension 2
+        (uint32_t) (src1->nb[2] / ggml_type_size(src1->type)),  // stride (elements) of src1 in dimension 2
+        (uint32_t) (src0->nb[3] / ggml_type_size(src0->type)),  // stride (elements) of src0 in dimension 3
+        (uint32_t) (src1->nb[3] / ggml_type_size(src1->type)),  // stride (elements) of src1 in dimension 3
+        (uint32_t) src0->ne[2],                                 // batch size in dimension 2
+        (uint32_t) src0->ne[3],                                 // batch size in dimension 3
+        (uint32_t) (src1->ne[2] / src0->ne[2]),                 // broadcast in dimension 2
+        (uint32_t) (src1->ne[3] / src0->ne[3])                  // broadcast in dimension 3
+    };
+
+    std::vector<wgpu::BindGroupEntry> entries = {
+        { .binding = 0,
+         .buffer  = ggml_backend_webgpu_tensor_buf(src0),
+         .offset  = ggml_backend_webgpu_tensor_offset(src0),
+         .size    = ggml_nbytes(src0) },
+        { .binding = 1,
+         .buffer  = ggml_backend_webgpu_tensor_buf(src1),
+         .offset  = ggml_backend_webgpu_tensor_offset(src1),
+         .size    = ggml_nbytes(src1) },
+        { .binding = 2,
+         .buffer  = ggml_backend_webgpu_tensor_buf(dst),
+         .offset  = ggml_backend_webgpu_tensor_offset(dst),
+         .size    = ggml_nbytes(dst)  }
+    };
+
+    uint32_t wg_x =
+        (dst->ne[0] * dst->ne[1] * dst->ne[2] * dst->ne[3] + WEBGPU_MUL_MAT_WG_SIZE - 1) / WEBGPU_MUL_MAT_WG_SIZE;
+    ggml_backend_webgpu_build_and_enqueue(ctx, ctx->mul_mat_pipeline, params, entries, wg_x);
+}
+
+// Returns true if node has enqueued work into the queue, false otherwise
+static bool ggml_webgpu_encode_node(webgpu_context ctx, ggml_tensor * node) {
+    if (ggml_is_empty(node)) {
+        return false;
+    }
+    WEBGPU_LOG_DEBUG("ggml_webgpu_encode_node(" << node << ", " << ggml_op_name(node->op) << ")");
+
+    ggml_tensor * src0 = node->src[0];
+    ggml_tensor * src1 = node->src[1];
+
+    switch (node->op) {
+            // no-ops
+        case GGML_OP_NONE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+            return false;
+        case GGML_OP_CPY:
+            {
+                ggml_webgpu_cpy(ctx, src0, node);
+                break;
+            }
+        case GGML_OP_SET_ROWS:
+            {
+                ggml_webgpu_set_rows(ctx, src0, src1, node);
+                break;
+            }
+        case GGML_OP_MUL_MAT:
+            {
+                ggml_webgpu_mul_mat(ctx, src0, src1, node);
+                break;
+            }
+        default:
+            return false;
+    }
+    return true;
+}
+
+static ggml_status ggml_backend_webgpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_graph_compute(" << cgraph->n_nodes << " nodes)");
+
+    ggml_backend_webgpu_context * backend_ctx = static_cast<ggml_backend_webgpu_context *>(backend->context);
+    webgpu_context                ctx         = backend_ctx->webgpu_ctx;
+
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        ggml_webgpu_encode_node(ctx, cgraph->nodes[i]);
+    }
+
+    ggml_backend_webgpu_submit_queue(ctx);
+    ggml_backend_webgpu_wait_on_submission(ctx);
+
+    return GGML_STATUS_SUCCESS;
+}
+
+static ggml_backend_i ggml_backend_webgpu_i = {
+    /* .get_name                = */ ggml_backend_webgpu_name,
+    /* .free                    = */ ggml_backend_webgpu_free,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_async        = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_webgpu_graph_compute,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+};
+
+/* End GGML Backend Interface */
+
+/* GGML Backend Buffer Interface */
+
+static void ggml_backend_webgpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_free_buffer()");
+    ggml_backend_webgpu_buffer_context * ctx = static_cast<ggml_backend_webgpu_buffer_context *>(buffer->context);
+    ctx->buffer.Destroy();
+}
+
+// Returns the "fake" base pointer.
+static void * ggml_backend_webgpu_buffer_get_base(ggml_backend_buffer_t buffer) {
+    GGML_UNUSED(buffer);
+    return webgpu_ptr_base;
+}
+
+static void ggml_backend_webgpu_buffer_memset_tensor(ggml_backend_buffer_t buffer,
+                                                     ggml_tensor *         tensor,
+                                                     uint8_t               value,
+                                                     size_t                offset,
+                                                     size_t                size) {
+    if (size == 0) {
+        WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_memset_tensor: size is zero, nothing to do.");
+        return;
+    }
+
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_memset_tensor(" << buffer << ", " << tensor << ", " << value << ", "
+                                                                 << offset << ", " << size << ")");
+
+    ggml_backend_webgpu_buffer_context * buf_ctx = (ggml_backend_webgpu_buffer_context *) buffer->context;
+
+    size_t total_offset = webgpu_tensor_offset(tensor) + tensor->view_offs + offset;
+
+    // This is a trick to set all bytes of a u32 to the same 1 byte value.
+    uint32_t val32 = (uint32_t) value * 0x01010101;
+    ggml_backend_webgpu_buffer_memset(buf_ctx->webgpu_ctx, buf_ctx->buffer, val32, total_offset, size);
+}
+
+static void ggml_backend_webgpu_buffer_set_tensor(ggml_backend_buffer_t buffer,
+                                                  ggml_tensor *         tensor,
+                                                  const void *          data,
+                                                  size_t                offset,
+                                                  size_t                size) {
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_set_tensor(" << buffer << ", " << tensor << ", " << data << ", "
+                                                              << offset << ", " << size << ")");
+    ggml_backend_webgpu_buffer_context * buf_ctx    = (ggml_backend_webgpu_buffer_context *) buffer->context;
+    webgpu_context                       webgpu_ctx = buf_ctx->webgpu_ctx;
+
+    size_t total_offset = webgpu_tensor_offset(tensor) + tensor->view_offs + offset;
+
+    webgpu_ctx->queue.WriteBuffer(buf_ctx->buffer, total_offset, data, (size / 4) * 4);
+
+    if (size % 4 != 0) {
+        // If size is not a multiple of 4, we need to memset the remaining bytes
+        size_t remaining_size = size % 4;
+
+        // pack the remaining bytes into a uint32_t
+        uint32_t val32 = 0;
+
+        for (size_t i = 0; i < remaining_size; i++) {
+            ((uint8_t *) &val32)[i] = ((const uint8_t *) data)[size - remaining_size + i];
+        }
+        // memset the remaining bytes
+        ggml_backend_webgpu_buffer_memset(
+            webgpu_ctx, buf_ctx->buffer, val32, total_offset + (size - remaining_size), remaining_size);
+    } else {
+        // wait for WriteBuffer to complete
+        ggml_backend_webgpu_wait_on_submission(webgpu_ctx);
+    }
+}
+
+static void ggml_backend_webgpu_buffer_get_tensor(ggml_backend_buffer_t buffer,
+                                                  const ggml_tensor *   tensor,
+                                                  void *                data,
+                                                  size_t                offset,
+                                                  size_t                size) {
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_get_tensor(" << buffer << ", " << tensor << ", " << data << ", "
+                                                              << offset << ", " << size << ")");
+
+    ggml_backend_webgpu_buffer_context * buf_ctx    = (ggml_backend_webgpu_buffer_context *) buffer->context;
+    webgpu_context                       webgpu_ctx = buf_ctx->webgpu_ctx;
+    wgpu::Device                         device     = webgpu_ctx->device;
+
+    size_t total_offset = webgpu_tensor_offset(tensor) + tensor->view_offs + offset;
+
+    size_t final_size = size;
+    if (size % 4 != 0) {
+        // If size is not a multiple of 4, we need to round it up to the next multiple of 4
+        final_size = size + (4 - (size % 4));
+    }
+
+    std::lock_guard<std::recursive_mutex> lock(webgpu_ctx->mutex);
+
+    if (webgpu_ctx->get_tensor_staging_buf == nullptr || webgpu_ctx->get_tensor_staging_buf.GetSize() < final_size) {
+        // Create a new staging buffer if it doesn't exist or is too small
+        if (webgpu_ctx->get_tensor_staging_buf) {
+            webgpu_ctx->get_tensor_staging_buf.Destroy();
+        }
+        ggml_webgpu_create_buffer(device,
+                                  webgpu_ctx->get_tensor_staging_buf,
+                                  final_size,
+                                  wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead,
+                                  "get_tensor_staging_buf");
+    }
+
+    // Copy the data from the buffer to the staging buffer
+    wgpu::CommandEncoder encoder = device.CreateCommandEncoder();
+    encoder.CopyBufferToBuffer(buf_ctx->buffer, total_offset, webgpu_ctx->get_tensor_staging_buf, 0, final_size);
+    wgpu::CommandBuffer commands = encoder.Finish();
+
+    // Submit the command buffer to the queue
+    webgpu_ctx->queue.Submit(1, &commands);
+
+    // Map the staging buffer to read the data
+    ggml_backend_webgpu_map_buffer(webgpu_ctx, webgpu_ctx->get_tensor_staging_buf, wgpu::MapMode::Read, 0, final_size);
+    // Must specify size here since the staging buffer might be larger than the tensor size
+    const void * mapped_range = webgpu_ctx->get_tensor_staging_buf.GetConstMappedRange(0, final_size);
+
+    // Copy the data from the mapped range to the output buffer
+    std::memcpy(data, mapped_range, size);
+    webgpu_ctx->get_tensor_staging_buf.Unmap();
+}
+
+static void ggml_backend_webgpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_clear(" << buffer << ", " << (uint32_t) value << ")");
+    ggml_backend_webgpu_buffer_context * buf_ctx = (ggml_backend_webgpu_buffer_context *) buffer->context;
+    ggml_backend_webgpu_buffer_memset(buf_ctx->webgpu_ctx, buf_ctx->buffer, value, 0, buffer->size);
+}
+
+static ggml_backend_buffer_i ggml_backend_webgpu_buffer_interface = {
+    /* .free_buffer     = */ ggml_backend_webgpu_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_webgpu_buffer_get_base,
+    /* .init_tensor     = */ NULL,  // TODO: optional, needed?
+    /* .memset_tensor   = */ ggml_backend_webgpu_buffer_memset_tensor,
+    /* .set_tensor      = */ ggml_backend_webgpu_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_webgpu_buffer_get_tensor,
+    /* .cpy_tensor      = */ NULL,  // TODO: optional, implement this
+    /* .clear           = */ ggml_backend_webgpu_buffer_clear,
+    /* .reset           = */ NULL,  // TODO: optional, think it coordinates with .init_tensor
+};
+
+/* End GGML Backend Buffer Interface */
+
+/* GGML Backend Buffer Type Interface */
+
+static const char * ggml_backend_webgpu_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(buft->device->context);
+    return ctx->device_name.c_str();
+}
+
+static ggml_backend_buffer_t ggml_backend_webgpu_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
+                                                                          size_t                     size) {
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_type_alloc_buffer(" << size << ")");
+    ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(buft->device->context);
+
+    wgpu::Buffer buf;
+    ggml_webgpu_create_buffer(ctx->webgpu_ctx->device,
+                              buf,
+                              size,
+                              wgpu::BufferUsage::Storage | wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::CopyDst,
+                              "allocated_buffer");
+
+    ggml_backend_webgpu_buffer_context * buf_ctx = new ggml_backend_webgpu_buffer_context(ctx->webgpu_ctx, buf);
+
+    return ggml_backend_buffer_init(buft, ggml_backend_webgpu_buffer_interface, buf_ctx, size);
+}
+
+static size_t ggml_backend_webgpu_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(buft->device->context);
+    return ctx->webgpu_ctx->limits.minStorageBufferOffsetAlignment;
+}
+
+// maxBufferSize might be larger, but you can't bind more than maxStorageBufferBindingSize to a single binding.
+static size_t ggml_backend_webgpu_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
+    ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(buft->device->context);
+    return ctx->webgpu_ctx->limits.maxStorageBufferBindingSize;
+}
+
+/* End GGML Backend Buffer Type Interface */
+
+/* GGML Backend Device Interface */
+
+static const char * ggml_backend_webgpu_device_get_name(ggml_backend_dev_t dev) {
+    ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(dev->context);
+    return ctx->device_name.c_str();
+}
+
+static const char * ggml_backend_webgpu_device_get_description(ggml_backend_dev_t dev) {
+    ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(dev->context);
+    return ctx->device_desc.c_str();
+}
+
+static void ggml_backend_webgpu_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(dev->context);
+    // TODO: what do we actually want to return here? maxBufferSize might not be the full available memory.
+    *free                                    = ctx->webgpu_ctx->limits.maxBufferSize;
+    *total                                   = ctx->webgpu_ctx->limits.maxBufferSize;
+}
+
+static enum ggml_backend_dev_type ggml_backend_webgpu_device_get_type(ggml_backend_dev_t dev) {
+    GGML_UNUSED(dev);
+    return GGML_BACKEND_DEVICE_TYPE_GPU;
+}
+
+static void ggml_backend_webgpu_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_webgpu_device_get_name(dev);
+    props->description = ggml_backend_webgpu_device_get_description(dev);
+    props->type        = ggml_backend_webgpu_device_get_type(dev);
+    ggml_backend_webgpu_device_get_memory(dev, &props->memory_free, &props->memory_total);
+    props->caps = {
+        /* .async                 = */ false,
+        /* .host_buffer           = */ false,
+        /* .buffer_from_host_ptr  = */ false,
+        /* .events                = */ false,
+    };
+}
+
+static ggml_guid_t ggml_backend_webgpu_guid(void) {
+    static const char * guid_str = "__ggml_webgpu :)";
+    return reinterpret_cast<ggml_guid_t>((void *) guid_str);
+}
+
+static void ggml_webgpu_init_memset_pipeline(webgpu_context & webgpu_ctx) {
+    // we use the maximum workgroup size for the memset pipeline
+    size_t max_wg_size = webgpu_ctx->limits.maxComputeWorkgroupSizeX;
+    size_t max_threads = max_wg_size * webgpu_ctx->limits.maxComputeWorkgroupsPerDimension;
+    // Size the bytes_per_thread so that the largest buffer size can be handled
+    webgpu_ctx->memset_bytes_per_thread =
+        (webgpu_ctx->limits.maxStorageBufferBindingSize + max_threads - 1) / max_threads;
+    std::vector<wgpu::ConstantEntry> constants(2);
+    constants[0].key   = "wg_size";
+    constants[0].value = max_wg_size;
+    constants[1].key   = "bytes_per_thread";
+    constants[1].value = webgpu_ctx->memset_bytes_per_thread;
+    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->memset_pipeline, wgsl_memset, "memset", constants);
+}
+
+static void ggml_webgpu_init_mul_mat_pipeline(webgpu_context & webgpu_ctx) {
+    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline, wgsl_mul_mat, "mul_mat");
+}
+
+static void ggml_webgpu_init_set_rows_pipeline(webgpu_context & webgpu_ctx) {
+    std::vector<wgpu::ConstantEntry> constants(1);
+    constants[0].key   = "wg_size";
+    constants[0].value = webgpu_ctx->limits.maxComputeWorkgroupSizeX;
+    ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, webgpu_ctx->set_rows_pipeline, wgsl_set_rows, "set_rows", constants);
+}
+
+static void ggml_webgpu_init_cpy_pipeline(webgpu_context & webgpu_ctx) {
+    std::vector<wgpu::ConstantEntry> constants(1);
+    constants[0].key   = "wg_size";
+    constants[0].value = webgpu_ctx->limits.maxComputeWorkgroupSizeX;
+    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->cpy_pipeline, wgsl_cpy, "cpy", constants);
+}
+
+static ggml_backend_t ggml_backend_webgpu_device_init(ggml_backend_dev_t dev, const char * params) {
+    GGML_UNUSED(params);
+
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_device_init()");
+
+    ggml_backend_webgpu_device_context * dev_ctx    = static_cast<ggml_backend_webgpu_device_context *>(dev->context);
+    webgpu_context                       webgpu_ctx = dev_ctx->webgpu_ctx;
+
+    // Multiple threads may try to initialize the device
+    std::lock_guard<std::recursive_mutex> lock(webgpu_ctx->mutex);
+    if (!webgpu_ctx->device_init) {
+        // Initialize device
+        std::vector<wgpu::FeatureName> required_features = { wgpu::FeatureName::ShaderF16,
+                                                             wgpu::FeatureName::ImplicitDeviceSynchronization };
+        wgpu::DeviceDescriptor         dev_desc;
+        dev_desc.requiredLimits       = &webgpu_ctx->limits;
+        dev_desc.requiredFeatures     = required_features.data();
+        dev_desc.requiredFeatureCount = required_features.size();
+        dev_desc.SetDeviceLostCallback(
+            wgpu::CallbackMode::AllowSpontaneous,
+            [](const wgpu::Device & device, wgpu::DeviceLostReason reason, wgpu::StringView message) {
+                GGML_UNUSED(device);
+                GGML_LOG_ERROR(
+                    "ggml_webgpu: Device lost! Reason: %d, Message: %s\n", static_cast<int>(reason), message.data);
+            });
+        dev_desc.SetUncapturedErrorCallback(
+            [](const wgpu::Device & device, wgpu::ErrorType reason, wgpu::StringView message) {
+                GGML_UNUSED(device);
+                GGML_LOG_ERROR(
+                    "ggml_webgpu: Device error! Reason: %d, Message: %s\n", static_cast<int>(reason), message.data);
+            });
+        webgpu_ctx->instance.WaitAny(
+            webgpu_ctx->adapter.RequestDevice(
+                &dev_desc,
+                wgpu::CallbackMode::AllowSpontaneous,
+                [webgpu_ctx](wgpu::RequestDeviceStatus status, wgpu::Device device, wgpu::StringView message) {
+                    if (status != wgpu::RequestDeviceStatus::Success) {
+                        GGML_LOG_ERROR("ggml_webgpu: Failed to get a device: %s\n", message.data);
+                        return;
+                    }
+                    webgpu_ctx->device = std::move(device);
+                }),
+            UINT64_MAX);
+        GGML_ASSERT(webgpu_ctx->device != nullptr);
+
+        // Initialize (compute) queue
+        webgpu_ctx->queue = webgpu_ctx->device.GetQueue();
+
+        // Create buffer pool for shader parameters
+        webgpu_ctx->param_buf_pool.init(webgpu_ctx->device,
+                                        WEBGPU_NUM_PARAM_BUFS,
+                                        WEBGPU_PARAMS_BUF_SIZE_BYTES,
+                                        wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::Uniform,
+                                        wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::MapWrite);
+        webgpu_ctx->set_rows_error_buf_pool.init(webgpu_ctx->device,
+                                                 WEBGPU_NUM_SET_ROWS_ERROR_BUFS,
+                                                 WEBGPU_SET_ROWS_ERROR_BUF_SIZE_BYTES,
+                                                 wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::Storage,
+                                                 wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead);
+
+        ggml_webgpu_init_memset_pipeline(webgpu_ctx);
+        ggml_webgpu_init_mul_mat_pipeline(webgpu_ctx);
+        ggml_webgpu_init_set_rows_pipeline(webgpu_ctx);
+        ggml_webgpu_init_cpy_pipeline(webgpu_ctx);
+
+#ifdef GGML_WEBGPU_DEBUG
+        // Initialize debug buffers
+        ggml_webgpu_create_buffer(webgpu_ctx->device,
+                                  webgpu_ctx->debug_host_buf,
+                                  WEBGPU_DEBUG_BUF_ELEMS * sizeof(uint32_t),
+                                  wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead,
+                                  "debug_host_buf");
+        ggml_webgpu_create_buffer(webgpu_ctx->device,
+                                  webgpu_ctx->debug_dev_buf,
+                                  WEBGPU_DEBUG_BUF_ELEMS * sizeof(uint32_t),
+                                  wgpu::BufferUsage::Storage | wgpu::BufferUsage::CopySrc,
+                                  "debug_dev_buf");
+#endif
+        webgpu_ctx->device_init = true;
+    }
+
+    static ggml_backend_webgpu_context backend_ctx;
+    backend_ctx.name       = GGML_WEBGPU_NAME + std::string(": ") + dev_ctx->device_name;
+    backend_ctx.webgpu_ctx = webgpu_ctx;
+
+    // See GGML Backend Interface section
+    static ggml_backend backend = {
+        /* .guid      = */ ggml_backend_webgpu_guid(),
+        /* .interface = */ ggml_backend_webgpu_i,
+        /* .device    = */ dev,
+        /* .context   = */ &backend_ctx,
+    };
+
+    return &backend;
+}
+
+static ggml_backend_buffer_type_t ggml_backend_webgpu_device_get_buffer_type(ggml_backend_dev_t dev) {
+    // See GGML Backend Buffer Type Interface section
+    static struct ggml_backend_buffer_type ggml_backend_webgpu_buffer_type = {
+        /* .iface = */ {
+                        /* .get_name         = */ ggml_backend_webgpu_buffer_type_get_name,
+                        /* .alloc_buffer     = */ ggml_backend_webgpu_buffer_type_alloc_buffer,
+                        /* .get_alignment    = */ ggml_backend_webgpu_buffer_type_get_alignment,
+                        /* .get_max_size     = */ ggml_backend_webgpu_buffer_type_get_max_size,
+                        /* .get_alloc_size   = */ NULL,  // defaults to ggml_nbytes
+            /* .is_host          = */ NULL,  // defaults to false
+        },
+        /* .device  = */
+        dev,
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_webgpu_buffer_type;
+}
+
+static bool ggml_backend_webgpu_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    GGML_UNUSED(dev);
+    return buft->iface.get_name == ggml_backend_webgpu_buffer_type_get_name;
+}
+
+static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
+    GGML_UNUSED(dev);
+
+    switch (op->op) {
+        case GGML_OP_NONE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+            return true;
+        case GGML_OP_CPY | GGML_OP_SET_ROWS:
+            return op->type == GGML_TYPE_F16 && op->src[0]->type == GGML_TYPE_F32;
+        case GGML_OP_MUL_MAT:
+            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
+        default:
+            return false;
+    }
+}
+
+static struct ggml_backend_device_i ggml_backend_webgpu_device_i = {
+    /* .get_name             = */ ggml_backend_webgpu_device_get_name,
+    /* .get_description      = */ ggml_backend_webgpu_device_get_description,
+    /* .get_memory           = */ ggml_backend_webgpu_device_get_memory,
+    /* .get_type             = */ ggml_backend_webgpu_device_get_type,
+    /* .get_props            = */ ggml_backend_webgpu_device_get_props,
+    /* .init_backend         = */ ggml_backend_webgpu_device_init,
+    /* .get_buffer_type      = */ ggml_backend_webgpu_device_get_buffer_type,
+    /* .get_host_buffer_type = */ NULL,
+    /* .buffer_from_host_ptr = */ NULL,
+    /* .supports_op          = */ ggml_backend_webgpu_device_supports_op,
+    /* .supports_buft        = */ ggml_backend_webgpu_device_supports_buft,
+    /* .offload_op           = */ NULL,
+    /* .event_new            = */ NULL,
+    /* .event_free           = */ NULL,
+    /* .event_synchronize    = */ NULL,
+};
+
+/* End GGML Backend Device Interface */
+
+/* GGML Backend Registration Interface */
+
+static const char * ggml_backend_webgpu_reg_get_name(ggml_backend_reg_t reg) {
+    ggml_backend_webgpu_reg_context * ctx = static_cast<ggml_backend_webgpu_reg_context *>(reg->context);
+    return ctx->name;
+}
+
+static size_t ggml_backend_webgpu_reg_get_device_count(ggml_backend_reg_t reg) {
+    ggml_backend_webgpu_reg_context * ctx = static_cast<ggml_backend_webgpu_reg_context *>(reg->context);
+    return ctx->device_count;
+}
+
+// TODO: Does this need to be thread safe? Is it only called once?
+// Only one device is supported for now
+static ggml_backend_dev_t ggml_backend_webgpu_reg_get_device(ggml_backend_reg_t reg, size_t index) {
+    GGML_ASSERT(index == 0);
+    WEBGPU_LOG_DEBUG("ggml_backend_reg_get_device()");
+
+    ggml_backend_webgpu_reg_context * reg_ctx = static_cast<ggml_backend_webgpu_reg_context *>(reg->context);
+
+    webgpu_context ctx = reg_ctx->webgpu_ctx;
+
+    wgpu::RequestAdapterOptions options = {};
+    auto                        callback =
+        [](wgpu::RequestAdapterStatus status, wgpu::Adapter adapter, const char * message, void * userdata) {
+            if (status != wgpu::RequestAdapterStatus::Success) {
+                GGML_LOG_ERROR("ggml_webgpu: Failed to get an adapter: %s\n", message);
+                return;
+            }
+            *static_cast<wgpu::Adapter *>(userdata) = std::move(adapter);
+        };
+    void * userdata = &ctx->adapter;
+    ctx->instance.WaitAny(
+        ctx->instance.RequestAdapter(&options, wgpu::CallbackMode::AllowSpontaneous, callback, userdata), UINT64_MAX);
+    GGML_ASSERT(ctx->adapter != nullptr);
+
+    ctx->adapter.GetLimits(&ctx->limits);
+
+    wgpu::AdapterInfo info{};
+    ctx->adapter.GetInfo(&info);
+
+    static ggml_backend_webgpu_device_context device_ctx;
+    device_ctx.webgpu_ctx  = ctx;
+    device_ctx.device_name = GGML_WEBGPU_NAME;
+    device_ctx.device_desc = std::string(info.description.data);
+
+    GGML_LOG_INFO(
+        "ggml_webgpu: adapter_info: vendor_id: %u | vendor: %s | architecture: %s | device_id: %u | name: %s | "
+        "device_desc: %s\n",
+        info.vendorID,
+        info.vendor.data,
+        info.architecture.data,
+        info.deviceID,
+        info.device.data,
+        info.description.data);
+
+    // See GGML Backend Device Interface section
+    static ggml_backend_device device = {
+        /* .iface   = */ ggml_backend_webgpu_device_i,
+        /* .reg     = */ reg,
+        /* .context = */ &device_ctx,
+    };
+    return &device;
+}
+
+static const struct ggml_backend_reg_i ggml_backend_webgpu_reg_i = {
+    /* .get_name         = */ ggml_backend_webgpu_reg_get_name,
+    /* .get_device_count = */ ggml_backend_webgpu_reg_get_device_count,
+    /* .get_device       = */ ggml_backend_webgpu_reg_get_device,
+    /* .get_proc_address = */ NULL,
+};
+
+/* End GGML Backend Registration Interface */
+
+ggml_backend_reg_t ggml_backend_webgpu_reg() {
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_reg()");
+
+    webgpu_context webgpu_ctx = std::make_shared<webgpu_context_struct>();
+
+    static ggml_backend_webgpu_reg_context ctx;
+    ctx.webgpu_ctx   = webgpu_ctx;
+    ctx.name         = GGML_WEBGPU_NAME;
+    ctx.device_count = 1;
+
+    wgpu::InstanceDescriptor               instance_descriptor{};
+    std::vector<wgpu::InstanceFeatureName> instance_features = { wgpu::InstanceFeatureName::TimedWaitAny };
+    instance_descriptor.requiredFeatures                     = instance_features.data();
+    instance_descriptor.requiredFeatureCount                 = instance_features.size();
+    webgpu_ctx->instance                                     = wgpu::CreateInstance(&instance_descriptor);
+    GGML_ASSERT(webgpu_ctx->instance != nullptr);
+
+    static ggml_backend_reg reg = {
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_webgpu_reg_i,
+        /* .context     = */ &ctx,
+    };
+    return &reg;
+}
+
+ggml_backend_t ggml_backend_webgpu_init(void) {
+    ggml_backend_dev_t dev = ggml_backend_reg_dev_get(ggml_backend_webgpu_reg(), 0);
+
+    return ggml_backend_webgpu_device_init(dev, nullptr);
+}
+
+GGML_BACKEND_DL_IMPL(ggml_backend_webgpu_reg)
diff --git a/ggml/src/ggml-webgpu/wgsl-shaders/cpy.wgsl b/ggml/src/ggml-webgpu/wgsl-shaders/cpy.wgsl
new file mode 100644
index 00000000000..6fe924c554c
--- /dev/null
+++ b/ggml/src/ggml-webgpu/wgsl-shaders/cpy.wgsl
@@ -0,0 +1,60 @@
+enable f16;
+
+@group(0) @binding(0)
+var<storage, read_write> src: array<f32>;
+
+@group(0) @binding(1)
+var<storage, read_write> dst: array<f16>;
+
+struct Params {
+    ne: u32,            // total number of elements
+    offset_src: u32,    // in elements
+    offset_dst: u32,    // in elements
+
+    // Strides (in elements) — may be permuted
+    stride_src0: u32,
+    stride_src1: u32,
+    stride_src2: u32,
+    stride_src3: u32,
+
+    stride_dst0: u32,
+    stride_dst1: u32,
+    stride_dst2: u32,
+    stride_dst3: u32,
+
+    // Logical shape (same for both tensors)
+    ne0: u32,
+    ne1: u32,
+    ne2: u32,
+    ne3: u32,
+};
+
+@group(0) @binding(2)
+var<uniform> params: Params;
+
+override wg_size: u32;
+@compute @workgroup_size(wg_size)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    if (gid.x >= params.ne) {
+        return;
+    }
+
+    var i = gid.x;
+
+    let i3 = i / (params.ne2 * params.ne1 * params.ne0);
+    i = i % (params.ne2 * params.ne1 * params.ne0);
+
+    let i2 = i / (params.ne1 * params.ne0);
+    i = i % (params.ne1 * params.ne0);
+
+    let i1 = i / params.ne0;
+    let i0 = i % params.ne0;
+
+    let src_idx = i0 * params.stride_src0 + i1 * params.stride_src1 +
+                  i2 * params.stride_src2 + i3 * params.stride_src3;
+
+    let dst_idx = i0 * params.stride_dst0 + i1 * params.stride_dst1 +
+                  i2 * params.stride_dst2 + i3 * params.stride_dst3;
+
+    dst[params.offset_dst + dst_idx] = f16(src[params.offset_src + src_idx]);
+}
diff --git a/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py b/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py
new file mode 100755
index 00000000000..962dcd6b170
--- /dev/null
+++ b/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py
@@ -0,0 +1,35 @@
+import os
+import argparse
+
+
+def escape_triple_quotes(wgsl):
+    # Simple defense in case of embedded """
+    return wgsl.replace('"""', '\\"""')
+
+
+def to_cpp_string_literal(varname, content):
+    return f'const char* wgsl_{varname} = R"({content})";\n'
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--input', required=True)
+    parser.add_argument('--output', required=True)
+    args = parser.parse_args()
+
+    with open(args.output, 'w', encoding='utf-8') as out:
+        out.write("// Auto-generated shader embedding \n\n")
+        for fname in sorted(os.listdir(args.input)):
+            if not fname.endswith('.wgsl'):
+                continue
+            shader_path = os.path.join(args.input, fname)
+            varname = os.path.splitext(fname)[0]
+            with open(shader_path, 'r', encoding='utf-8') as f:
+                content = f.read()
+            content = escape_triple_quotes(content)
+            out.write(to_cpp_string_literal(varname, content))
+            out.write('\n')
+
+
+if __name__ == '__main__':
+    main()
diff --git a/ggml/src/ggml-webgpu/wgsl-shaders/memset.wgsl b/ggml/src/ggml-webgpu/wgsl-shaders/memset.wgsl
new file mode 100644
index 00000000000..cb7c8c3e09e
--- /dev/null
+++ b/ggml/src/ggml-webgpu/wgsl-shaders/memset.wgsl
@@ -0,0 +1,40 @@
+@group(0) @binding(0)
+var<storage, read_write> output_buffer: array<u32>;
+
+struct Params {
+    offset: u32, // in bytes
+    size: u32,   // in bytes
+    value: u32,  // 4 8-bit values, which are either repeating (memset_tensor) or may be separate (cleaning up unaligned set_tensor operations)
+};
+
+@group(0) @binding(1)
+var<uniform> params: Params;
+
+override wg_size: u32;
+override bytes_per_thread: u32;
+
+@compute @workgroup_size(wg_size)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let i = gid.x * bytes_per_thread;
+    let start = params.offset;
+    let end = params.offset + params.size;
+
+    for (var j: u32 = 0u; j < bytes_per_thread; j = j + 1u) {
+        let byte_index = start + i + j;
+        if (byte_index + 4u <= end) {
+            output_buffer[(byte_index >> 2u)] = params.value;
+        } else {
+            // Handle tail (unaligned)
+            for (var k: u32 = 0u; k < 4u; k = k + 1u) {
+                let idx = byte_index + k;
+                if (idx < end) {
+                    let word_idx = idx >> 2u;
+                    let byte_offset = (idx & 3u) * 8u;
+                    let mask = ~(0xffu << byte_offset);
+                    let existing = output_buffer[word_idx];
+                    output_buffer[word_idx] = (existing & mask) | ((params.value & 0xffu) << byte_offset);
+                }
+            }
+        }
+    }
+}
diff --git a/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.wgsl b/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.wgsl
new file mode 100644
index 00000000000..054aab566f9
--- /dev/null
+++ b/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.wgsl
@@ -0,0 +1,56 @@
+struct MulMatParams {
+    m: u32,
+    n: u32,
+    k: u32,
+    // all strides are in elements
+    stride_01: u32,
+    stride_11: u32,
+    stride_02: u32,
+    stride_12: u32,
+    stride_03: u32,
+    stride_13: u32,
+
+    bs02: u32,
+    bs03: u32,
+    broadcast2: u32,
+    broadcast3: u32
+};
+
+@group(0) @binding(0) var<storage, read_write> src0: array<f32>; // N rows, K columns
+@group(0) @binding(1) var<storage, read_write> src1: array<f32>; // M rows, K columns (transposed)
+@group(0) @binding(2) var<storage, read_write> dst: array<f32>; // M rows, N columns
+
+@group(0) @binding(3) var<uniform> params: MulMatParams;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
+    let total = params.m * params.n * params.bs02 * params.broadcast2 * params.bs03 * params.broadcast3;
+    if (global_id.x >= total) {
+        return;
+    }
+
+    let dst2_stride = params.m * params.n;
+    let dst3_stride = dst2_stride * params.bs02 * params.broadcast2;
+
+    let dst3_idx = global_id.x / dst3_stride;
+    let src03_idx = dst3_idx / params.broadcast3; // src0 may be broadcast along the third dimension
+    let src13_idx = dst3_idx; // src1 is not broadcast
+    let dst3_rem = global_id.x % dst3_stride;
+
+    let dst2_idx = dst3_rem / dst2_stride;
+    let src02_idx = dst2_idx / params.broadcast2; // src0 may also be broadcast along the second dimension
+    let src12_idx = dst2_idx; // src1 is not broadcast
+
+    let dst2_rem = dst3_rem % dst2_stride;
+
+    let row = dst2_rem / params.n; // output row
+    let col = dst2_rem % params.n; // output column
+
+    var sum = 0.0;
+    for (var i: u32 = 0u; i < params.k; i = i + 1u) {
+        let src0_idx = src03_idx * params.stride_03 + src02_idx * params.stride_02 + col * params.stride_01 + i;
+        let src1_idx = src13_idx * params.stride_13 + src12_idx * params.stride_12 + row * params.stride_11 + i;
+        sum = sum + src0[src0_idx] * src1[src1_idx];
+    }
+    dst[dst3_idx * dst3_stride + dst2_idx * dst2_stride + row * params.n + col] = sum;
+}
diff --git a/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl b/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl
new file mode 100644
index 00000000000..4bd6f94a235
--- /dev/null
+++ b/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl
@@ -0,0 +1,82 @@
+enable f16;
+
+@group(0) @binding(0)
+var<storage, read_write> src: array<f32>;
+
+@group(0) @binding(1)
+var<storage, read_write> idx: array<u32>;
+
+@group(0) @binding(2)
+var<storage, read_write> dst: array<f16>;
+
+@group(0) @binding(3)
+var<storage, read_write> error: atomic<u32>;
+
+struct Params {
+    offset_src: u32, // in elements
+    offset_idx: u32, // in elements
+    offset_dst: u32, // in elements
+
+    // Strides (in elements)
+    stride_src1: u32,
+    stride_src2: u32,
+    stride_src3: u32,
+
+    stride_idx0: u32,
+    stride_idx1: u32,
+    stride_idx2: u32,
+
+    stride_dst1: u32,
+    stride_dst2: u32,
+    stride_dst3: u32,
+
+    // Shape of src
+    ne0: u32,
+    n_rows: u32,
+    ne2: u32,
+    ne3: u32,
+
+    // Shape of idx
+    idx1: u32,
+    idx2: u32,
+};
+
+@group(0) @binding(4)
+var<uniform> params: Params;
+
+override wg_size: u32;
+@compute @workgroup_size(wg_size)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    if (gid.x >= params.n_rows * params.ne2 * params.ne3) {
+        return;
+    }
+    var i = gid.x;
+    let i_src3 = i / (params.ne2 * params.n_rows);
+    let i_dst3 = i / (params.ne2 * 3);
+
+    i = i % (params.ne2 * params.n_rows);
+    let i_src2 = i / params.n_rows;
+    let i_src1 = i % params.n_rows;
+
+    let i_idx2 = i_src3 % params.idx2;
+    let i_idx1 = i_src2 % params.idx1;
+    let i_idx0 = i_src1;
+
+    let idx_high = (params.offset_idx + i_idx0 * params.stride_idx0 + i_idx1 * params.stride_idx1 + i_idx2 * params.stride_idx2) * 2;
+
+    let idx_high_val = idx[idx_high];
+    let idx_low_val = idx[idx_high + 1];
+
+    if (idx_low_val != 0) {
+        // Upper bits of index are not zero, output will be incorrect
+        atomicStore(&error, 1);
+        return;
+    }
+
+    let i_dst_row = params.offset_dst + idx_high_val * params.stride_dst1 + i_src2 * params.stride_dst2 + i_src3 * params.stride_dst3;
+    let i_src_row = params.offset_src + i_src1 * params.stride_src1 + i_src2 * params.stride_src2 + i_src3 * params.stride_src3;
+
+    for (var i: u32 = 0; i < params.ne0; i++) {
+      dst[i_dst_row + i] = f16(src[i_src_row + i]);
+    }
+}

From 293c1780070be7ef157fd04ba01b4a1bd2c7aad4 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Mon, 18 Aug 2025 19:31:43 +0300
Subject: [PATCH 85/86] sync : ggml

---
 scripts/sync-ggml.last | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last
index 5be1dce60aa..043c5a767a1 100644
--- a/scripts/sync-ggml.last
+++ b/scripts/sync-ggml.last
@@ -1 +1 @@
-b96890f3ab5ffbdbe56bc126df5366c34bd08d39
+a933eb1ccbfe3b290384a67a9db8eb2b596ccfff

From b74528f2844448971b78ec7d85df3de7557ef3c5 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Mon, 18 Aug 2025 19:32:04 +0300
Subject: [PATCH 86/86] talk-llama : sync llama.cpp

ggml-ci
---
 examples/talk-llama/llama-arch.cpp            |  127 ++
 examples/talk-llama/llama-arch.h              |   13 +
 examples/talk-llama/llama-batch.cpp           |    4 +-
 examples/talk-llama/llama-chat.cpp            |   38 +-
 examples/talk-llama/llama-chat.h              |    2 +
 examples/talk-llama/llama-context.cpp         |   75 +-
 examples/talk-llama/llama-context.h           |   16 +-
 examples/talk-llama/llama-graph.cpp           |  229 +++-
 examples/talk-llama/llama-graph.h             |   89 +-
 examples/talk-llama/llama-hparams.cpp         |   12 +-
 examples/talk-llama/llama-hparams.h           |   21 +-
 .../llama-kv-cache-unified-iswa.cpp           |   18 +-
 .../talk-llama/llama-kv-cache-unified-iswa.h  |    4 +-
 .../talk-llama/llama-kv-cache-unified.cpp     |  166 ++-
 examples/talk-llama/llama-kv-cache-unified.h  |    6 +-
 examples/talk-llama/llama-memory-hybrid.cpp   |   11 +-
 examples/talk-llama/llama-memory-hybrid.h     |    5 +-
 .../talk-llama/llama-memory-recurrent.cpp     |    8 +-
 examples/talk-llama/llama-memory-recurrent.h  |    4 +-
 examples/talk-llama/llama-memory.h            |    4 +-
 examples/talk-llama/llama-model-loader.cpp    |    1 +
 examples/talk-llama/llama-model-loader.h      |    5 +-
 examples/talk-llama/llama-model.cpp           | 1167 ++++++++++++++++-
 examples/talk-llama/llama-model.h             |   29 +-
 examples/talk-llama/llama-quant.cpp           |   44 +-
 examples/talk-llama/llama-vocab.cpp           |   51 +-
 examples/talk-llama/llama-vocab.h             |    1 +
 examples/talk-llama/llama.h                   |   38 +-
 28 files changed, 1905 insertions(+), 283 deletions(-)

diff --git a/examples/talk-llama/llama-arch.cpp b/examples/talk-llama/llama-arch.cpp
index 062a9977678..18dcc6ddfe5 100644
--- a/examples/talk-llama/llama-arch.cpp
+++ b/examples/talk-llama/llama-arch.cpp
@@ -62,6 +62,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_DEEPSEEK2,        "deepseek2"        },
     { LLM_ARCH_CHATGLM,          "chatglm"          },
     { LLM_ARCH_GLM4,             "glm4"             },
+    { LLM_ARCH_GLM4_MOE,         "glm4moe"          },
     { LLM_ARCH_BITNET,           "bitnet"           },
     { LLM_ARCH_T5,               "t5"               },
     { LLM_ARCH_T5ENCODER,        "t5encoder"        },
@@ -85,9 +86,13 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_ERNIE4_5,         "ernie4_5"         },
     { LLM_ARCH_ERNIE4_5_MOE,     "ernie4_5-moe"     },
     { LLM_ARCH_HUNYUAN_MOE,      "hunyuan-moe"      },
+    { LLM_ARCH_HUNYUAN_DENSE,    "hunyuan-dense"    },
     { LLM_ARCH_SMOLLM3,          "smollm3"          },
+    { LLM_ARCH_OPENAI_MOE,       "gpt-oss"          },
     { LLM_ARCH_LFM2,             "lfm2"             },
     { LLM_ARCH_DREAM,            "dream"            },
+    { LLM_ARCH_SMALLTHINKER,     "smallthinker"     },
+    { LLM_ARCH_LLADA,            "llada"            },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -124,6 +129,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_EXPERT_WEIGHTS_NORM,               "%s.expert_weights_norm"               },
     { LLM_KV_EXPERT_GATING_FUNC,                "%s.expert_gating_func"                },
     { LLM_KV_MOE_EVERY_N_LAYERS,                "%s.moe_every_n_layers"                },
+    { LLM_KV_NEXTN_PREDICT_LAYERS,              "%s.nextn_predict_layers"              },
     { LLM_KV_POOLING_TYPE,                      "%s.pooling_type"                      },
     { LLM_KV_LOGIT_SCALE,                       "%s.logit_scale"                       },
     { LLM_KV_DECODER_START_TOKEN_ID,            "%s.decoder_start_token_id"            },
@@ -1388,6 +1394,40 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
         },
     },
+    {
+        LLM_ARCH_GLM4_MOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_POST_NORM,     "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_FFN_GATE,           "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,           "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,             "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_GATE_SHEXP,     "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,     "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
+            { LLM_TENSOR_FFN_EXP_PROBS_B,    "blk.%d.exp_probs_b" },
+            // NextN/MTP tensors - preserved but unused (in final layer, dynamic layer number)
+            { LLM_TENSOR_NEXTN_EH_PROJ,      "blk.%d.nextn.eh_proj" },
+            { LLM_TENSOR_NEXTN_EMBED_TOKENS, "blk.%d.nextn.embed_tokens" },
+            { LLM_TENSOR_NEXTN_ENORM,        "blk.%d.nextn.enorm" },
+            { LLM_TENSOR_NEXTN_HNORM,        "blk.%d.nextn.hnorm" },
+            { LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "blk.%d.nextn.shared_head_head" },
+            { LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "blk.%d.nextn.shared_head_norm" },
+        },
+    },
     {
         LLM_ARCH_BITNET,
         {
@@ -1895,6 +1935,26 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
         },
     },
+    {
+        LLM_ARCH_HUNYUAN_DENSE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+
+        },
+    },
     {
         LLM_ARCH_SMOLLM3,
         {
@@ -1912,6 +1972,25 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,         "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_OPENAI_MOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_POST_NORM,     "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_SINKS,         "blk.%d.attn_sinks" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+        },
+    },
     {
         LLM_ARCH_LFM2,
         {
@@ -1933,6 +2012,27 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_TOKEN_EMBD_NORM,   "token_embd_norm" },
         }
     },
+    {
+        LLM_ARCH_SMALLTHINKER,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,           "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,           "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,             "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" }
+        },
+    },
     {
         LLM_ARCH_DREAM,
         {
@@ -1950,6 +2050,23 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_LLADA,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -1989,6 +2106,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_ATTN_KV_B,                  {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_K_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_V_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ATTN_SINKS,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SCALE}},
     {LLM_TENSOR_DEC_ATTN_Q,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_DEC_ATTN_K,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_DEC_ATTN_V,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
@@ -2120,6 +2238,14 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_SHORTCONV_CONV,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
     {LLM_TENSOR_SHORTCONV_INPROJ,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_SHORTCONV_OUTPROJ,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    // NextN/MTP tensors are currently ignored (reserved for future MTP support)
+    // These tensors only exist in the last layer(s) and are treated as output tensors
+    {LLM_TENSOR_NEXTN_EH_PROJ,              {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_NEXTN_EMBED_TOKENS,         {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_GET_ROWS}},
+    {LLM_TENSOR_NEXTN_ENORM,                {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_GET_ROWS}},
+    {LLM_TENSOR_NEXTN_HNORM,                {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
+    {LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,     {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,     {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
 };
 
 LLM_KV::LLM_KV(llm_arch arch, const char * suffix) : arch(arch), suffix(suffix) {}
@@ -2202,6 +2328,7 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
 bool llm_arch_is_diffusion(const llm_arch & arch) {
     switch (arch) {
         case LLM_ARCH_DREAM:
+        case LLM_ARCH_LLADA:
             return true;
         default:
             return false;
diff --git a/examples/talk-llama/llama-arch.h b/examples/talk-llama/llama-arch.h
index d09b7d7810b..7af587e7951 100644
--- a/examples/talk-llama/llama-arch.h
+++ b/examples/talk-llama/llama-arch.h
@@ -66,6 +66,7 @@ enum llm_arch {
     LLM_ARCH_DEEPSEEK2,
     LLM_ARCH_CHATGLM,
     LLM_ARCH_GLM4,
+    LLM_ARCH_GLM4_MOE,
     LLM_ARCH_BITNET,
     LLM_ARCH_T5,
     LLM_ARCH_T5ENCODER,
@@ -89,9 +90,13 @@ enum llm_arch {
     LLM_ARCH_ERNIE4_5,
     LLM_ARCH_ERNIE4_5_MOE,
     LLM_ARCH_HUNYUAN_MOE,
+    LLM_ARCH_HUNYUAN_DENSE,
     LLM_ARCH_SMOLLM3,
+    LLM_ARCH_OPENAI_MOE,
     LLM_ARCH_LFM2,
     LLM_ARCH_DREAM,
+    LLM_ARCH_SMALLTHINKER,
+    LLM_ARCH_LLADA,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -128,6 +133,7 @@ enum llm_kv {
     LLM_KV_EXPERT_WEIGHTS_NORM,
     LLM_KV_EXPERT_GATING_FUNC,
     LLM_KV_MOE_EVERY_N_LAYERS,
+    LLM_KV_NEXTN_PREDICT_LAYERS,
     LLM_KV_POOLING_TYPE,
     LLM_KV_LOGIT_SCALE,
     LLM_KV_DECODER_START_TOKEN_ID,
@@ -260,6 +266,7 @@ enum llm_tensor {
     LLM_TENSOR_ATTN_OUT_NORM,
     LLM_TENSOR_ATTN_POST_NORM,
     LLM_TENSOR_ATTN_ROT_EMBD,
+    LLM_TENSOR_ATTN_SINKS,
     LLM_TENSOR_FFN_GATE_INP,
     LLM_TENSOR_FFN_GATE_INP_SHEXP,
     LLM_TENSOR_FFN_NORM,
@@ -406,6 +413,12 @@ enum llm_tensor {
     LLM_TENSOR_SHORTCONV_CONV,
     LLM_TENSOR_SHORTCONV_INPROJ,
     LLM_TENSOR_SHORTCONV_OUTPROJ,
+    LLM_TENSOR_NEXTN_EH_PROJ,
+    LLM_TENSOR_NEXTN_EMBED_TOKENS,
+    LLM_TENSOR_NEXTN_ENORM,
+    LLM_TENSOR_NEXTN_HNORM,
+    LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,
+    LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,
 };
 
 enum llm_tensor_layer {
diff --git a/examples/talk-llama/llama-batch.cpp b/examples/talk-llama/llama-batch.cpp
index a546063c0a7..55d89eca0ad 100644
--- a/examples/talk-llama/llama-batch.cpp
+++ b/examples/talk-llama/llama-batch.cpp
@@ -59,7 +59,7 @@ bool llama_batch_allocr::init(
         for (int32_t i = 0; i < batch.n_tokens; ++i) {
             for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) {
                 if (batch.seq_id && (batch.seq_id[i][s] < 0 || batch.seq_id[i][s] >= (llama_seq_id) n_seq_max)) {
-                    LLAMA_LOG_ERROR("%s: invalid seq_id[%d][%d] = %d > %d\n", __func__, i, s, batch.seq_id[i][s], (llama_seq_id) n_seq_max);
+                    LLAMA_LOG_ERROR("%s: invalid seq_id[%d][%d] = %d >= %d\n", __func__, i, s, batch.seq_id[i][s], (llama_seq_id) n_seq_max);
                     return false;
                 }
             }
@@ -477,7 +477,7 @@ llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) {
 
 llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential) {
     if (sequential && has_cpl) {
-        LLAMA_LOG_ERROR("%s: sequential split is not supported when there are coupled sequences in the input batch\n", __func__);
+        LLAMA_LOG_ERROR("%s: sequential split is not supported when there are coupled sequences in the input batch (you may need to use the -kvu flag)\n", __func__);
 
         return {};
     }
diff --git a/examples/talk-llama/llama-chat.cpp b/examples/talk-llama/llama-chat.cpp
index d34bb26878c..0a96a9a579e 100644
--- a/examples/talk-llama/llama-chat.cpp
+++ b/examples/talk-llama/llama-chat.cpp
@@ -66,6 +66,8 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "llama4",            LLM_CHAT_TEMPLATE_LLAMA4            },
     { "smolvlm",           LLM_CHAT_TEMPLATE_SMOLVLM           },
     { "hunyuan-moe",       LLM_CHAT_TEMPLATE_HUNYUAN_MOE       },
+    { "gpt-oss",           LLM_CHAT_TEMPLATE_OPENAI_MOE        },
+    { "hunyuan-dense",     LLM_CHAT_TEMPLATE_HUNYUAN_DENSE     },
     { "kimi-k2",           LLM_CHAT_TEMPLATE_KIMI_K2           },
 };
 
@@ -191,8 +193,12 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_LLAMA4;
     } else if (tmpl_contains("<|endofuserprompt|>")) {
         return LLM_CHAT_TEMPLATE_DOTS1;
-    } else if (tmpl_contains("<|startoftext|>") && tmpl_contains("<|extra_4|>")) {
+    } else if (tmpl_contains("<|extra_0|>") && tmpl_contains("<|extra_4|>")) {
         return LLM_CHAT_TEMPLATE_HUNYUAN_MOE;
+    } else if (tmpl_contains("<|start|>") && tmpl_contains("<|channel|>")) {
+        return LLM_CHAT_TEMPLATE_OPENAI_MOE;
+    } else if (tmpl_contains("<｜hy_Assistant｜>") && tmpl_contains("<｜hy_place▁holder▁no▁3｜>")) {
+        return LLM_CHAT_TEMPLATE_HUNYUAN_DENSE;
     } else if (tmpl_contains("<|im_assistant|>assistant<|im_middle|>")) {
         return LLM_CHAT_TEMPLATE_KIMI_K2;
     }
@@ -619,8 +625,6 @@ int32_t llm_chat_apply_template(
     } else if (tmpl == LLM_CHAT_TEMPLATE_YANDEX) {
         // Yandex template ("\n\n" is defined as EOT token)
 
-        ss << "<s>";
-
         for (size_t i = 0; i < chat.size(); i++) {
             std::string role(chat[i]->role);
             if (role == "user") {
@@ -698,11 +702,37 @@ int32_t llm_chat_apply_template(
             if (role == "system") {
                 ss << "<|startoftext|>" << message->content << "<|extra_4|>";
             } else if (role == "assistant") {
-                ss << "<|startoftext|>" << message->content << "<|eos|>";
+                ss << message->content << "<|eos|>";
             } else {
                 ss << "<|startoftext|>" << message->content << "<|extra_0|>";
             }
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_OPENAI_MOE) {
+        // OpenAI MoE (based on Harmony chat template)
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|start|>" << role << "<|message|>" << message->content;
+            ss << (role == "assistant" ? "<|return|>" : "<|end|>");
+        }
+        if (add_ass) {
+            ss << "<|start|>assistant";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_HUNYUAN_DENSE) {
+        // tencent/Hunyuan-4B-Instruct
+        for (size_t i = 0; i < chat.size(); i++) {
+            std::string role(chat[i]->role);
+            if (i == 0) {
+                if (role == "system") {
+                    ss << chat[i]->content << "<｜hy_place▁holder▁no▁3｜>";
+                }
+            }
+
+            if (role == "assistant") {
+                ss << "<｜hy_Assistant｜>" << chat[i]->content << "<｜hy_place▁holder▁no▁2｜>";
+            } else if (role == "user") {
+                ss << "<｜hy_User｜>" << chat[i]->content << "<｜hy_Assistant｜>";
+            }
+        }
     } else if (tmpl == LLM_CHAT_TEMPLATE_KIMI_K2) {
         // moonshotai/Kimi-K2-Instruct
         for (auto message : chat) {
diff --git a/examples/talk-llama/llama-chat.h b/examples/talk-llama/llama-chat.h
index 6968a19fbe1..35a943856fa 100644
--- a/examples/talk-llama/llama-chat.h
+++ b/examples/talk-llama/llama-chat.h
@@ -46,6 +46,8 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_SMOLVLM,
     LLM_CHAT_TEMPLATE_DOTS1,
     LLM_CHAT_TEMPLATE_HUNYUAN_MOE,
+    LLM_CHAT_TEMPLATE_OPENAI_MOE,
+    LLM_CHAT_TEMPLATE_HUNYUAN_DENSE,
     LLM_CHAT_TEMPLATE_KIMI_K2,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
diff --git a/examples/talk-llama/llama-context.cpp b/examples/talk-llama/llama-context.cpp
index 9e77fe6d869..7d7abad5d4a 100644
--- a/examples/talk-llama/llama-context.cpp
+++ b/examples/talk-llama/llama-context.cpp
@@ -105,7 +105,7 @@ llama_context::llama_context(
 
     {
         const char * LLAMA_SET_ROWS = getenv("LLAMA_SET_ROWS");
-        supports_set_rows = LLAMA_SET_ROWS ? (atoi(LLAMA_SET_ROWS) != 0) : false;
+        supports_set_rows = LLAMA_SET_ROWS ? (atoi(LLAMA_SET_ROWS) != 0) : supports_set_rows;
 
         if (!supports_set_rows && !cparams.kv_unified) {
             LLAMA_LOG_WARN("%s: non-unified KV cache requires ggml_set_rows() - forcing unified KV cache\n", __func__);
@@ -113,6 +113,15 @@ llama_context::llama_context(
         }
     }
 
+    {
+        const char * LLAMA_GRAPH_REUSE_DISABLE = getenv("LLAMA_GRAPH_REUSE_DISABLE");
+        graph_reuse_disable = LLAMA_GRAPH_REUSE_DISABLE ? (atoi(LLAMA_GRAPH_REUSE_DISABLE) != 0) : graph_reuse_disable;
+
+        if (graph_reuse_disable) {
+            LLAMA_LOG_WARN("%s: graph reuse disabled\n", __func__);
+        }
+    }
+
     const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;
 
     LLAMA_LOG_INFO("%s: n_seq_max     = %u\n",   __func__, cparams.n_seq_max);
@@ -716,7 +725,7 @@ llm_graph_result * llama_context::process_ubatch(const llama_ubatch & ubatch, ll
     // in order to correctly reuse a graph, it's full topology has to be uniquely determined by these parameters
     const auto gparams = graph_params(res, ubatch, mctx, gtype);
 
-    if (res->can_reuse(gparams)) {
+    if (!graph_reuse_disable && res->can_reuse(gparams)) {
         //LLAMA_LOG_DEBUG("%s: reusing previous graph\n", __func__);
 
         n_reused++;
@@ -777,7 +786,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
     const auto & hparams = model.hparams;
 
     const int64_t n_embd  = hparams.n_embd;
-    const int32_t n_vocab = model.vocab.n_tokens();
+    const int64_t n_vocab = model.vocab.n_tokens();
 
     // note: during encode, we always pass the full sequence starting from pos = 0
     if (!balloc->init(batch_inp, model.vocab, nullptr, n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
@@ -950,7 +959,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
     const auto & vocab   = model.vocab;
     const auto & hparams = model.hparams;
 
-    const int32_t n_vocab = vocab.n_tokens();
+    const int64_t n_vocab = vocab.n_tokens();
     const int64_t n_embd  = hparams.n_embd;
 
     // when computing embeddings, all tokens are output
@@ -1319,21 +1328,21 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
 }
 
 void llama_context::output_reorder() {
-    const uint32_t n_vocab = model.vocab.n_tokens();
+    const uint64_t n_vocab = model.vocab.n_tokens();
     const uint64_t n_embd  = model.hparams.n_embd;
 
-    for (uint32_t s = 0; s < output_swaps.size(); ++s) {
-        const uint32_t i0 = output_swaps[s].i0;
-        const uint32_t i1 = output_swaps[s].i1;
+    for (size_t s = 0; s < output_swaps.size(); ++s) {
+        const uint64_t i0 = output_swaps[s].i0;
+        const uint64_t i1 = output_swaps[s].i1;
 
         if (logits_size > 0) {
-            for (uint32_t k = 0; k < n_vocab; k++) {
+            for (uint64_t k = 0; k < n_vocab; k++) {
                 std::swap(logits[i0*n_vocab + k], logits[i1*n_vocab + k]);
             }
         }
 
         if (embd_size > 0) {
-            for (uint32_t k = 0; k < n_embd; k++) {
+            for (uint64_t k = 0; k < n_embd; k++) {
                 std::swap(embd[i0*n_embd + k], embd[i1*n_embd + k]);
             }
         }
@@ -1648,30 +1657,30 @@ size_t llama_context::state_set_data(const uint8_t * src, size_t size) {
     }
 }
 
-size_t llama_context::state_seq_get_size(llama_seq_id seq_id) {
+size_t llama_context::state_seq_get_size(llama_seq_id seq_id, llama_state_seq_flags flags) {
     llama_io_write_dummy io;
     try {
-        return state_seq_write_data(io, seq_id);
+        return state_seq_write_data(io, seq_id, flags);
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: error getting state size: %s\n", __func__, err.what());
         return 0;
     }
 }
 
-size_t llama_context::state_seq_get_data(llama_seq_id seq_id, uint8_t * dst, size_t size) {
+size_t llama_context::state_seq_get_data(llama_seq_id seq_id, uint8_t * dst, size_t size, llama_state_seq_flags flags) {
     llama_io_write_buffer io(dst, size);
     try {
-        return state_seq_write_data(io, seq_id);
+        return state_seq_write_data(io, seq_id, flags);
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: error saving state: %s\n", __func__, err.what());
         return 0;
     }
 }
 
-size_t llama_context::state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size) {
+size_t llama_context::state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size, llama_state_seq_flags flags) {
     llama_io_read_buffer io(src, size);
     try {
-        return state_seq_read_data(io, seq_id);
+        return state_seq_read_data(io, seq_id, flags);
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: error loading state: %s\n", __func__, err.what());
         return 0;
@@ -1769,7 +1778,7 @@ size_t llama_context::state_seq_load_file(llama_seq_id seq_id, const char * file
     {
         const size_t state_size = file.size() - file.tell();
         llama_io_read_file io(&file);
-        const size_t nread = state_seq_read_data(io, seq_id);
+        const size_t nread = state_seq_read_data(io, seq_id, 0);
         if (!nread) {
             LLAMA_LOG_ERROR("%s: failed to restore sequence state\n", __func__);
             return 0;
@@ -1793,7 +1802,7 @@ size_t llama_context::state_seq_save_file(llama_seq_id seq_id, const char * file
 
     // save the context state using stream saving
     llama_io_write_file io(&file);
-    state_seq_write_data(io, seq_id);
+    state_seq_write_data(io, seq_id, 0);
 
     const size_t res = file.tell();
     GGML_ASSERT(res == sizeof(uint32_t) * 3 + sizeof(llama_token) * n_token_count + io.n_bytes());
@@ -1962,21 +1971,21 @@ size_t llama_context::state_read_data(llama_io_read_i & io) {
     return io.n_bytes();
 }
 
-size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id) {
+size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
     GGML_UNUSED(seq_id);
 
     if (memory) {
-        memory->state_write(io, seq_id);
+        memory->state_write(io, seq_id, flags);
     }
 
     return io.n_bytes();
 }
 
-size_t llama_context::state_seq_read_data(llama_io_read_i & io, llama_seq_id seq_id) {
+size_t llama_context::state_seq_read_data(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
     GGML_UNUSED(seq_id);
 
     if (memory) {
-        memory->state_read(io, seq_id);
+        memory->state_read(io, seq_id, flags);
     }
 
     return io.n_bytes();
@@ -2039,7 +2048,7 @@ void llama_context::opt_init(struct llama_model * model, struct llama_opt_params
     opt_params.opt_period      = n_batch / n_ubatch;
     opt_params.get_opt_pars    = lopt_params.get_opt_pars;
     opt_params.get_opt_pars_ud = lopt_params.get_opt_pars_ud;
-
+    opt_params.optimizer       = lopt_params.optimizer_type;
     opt_ctx = ggml_opt_init(opt_params);
 
     llama_opt_param_filter param_filter = lopt_params.param_filter;
@@ -2792,19 +2801,31 @@ bool llama_state_save_file(llama_context * ctx, const char * path_session, const
 }
 
 size_t llama_state_seq_get_size(llama_context * ctx, llama_seq_id seq_id) {
-    return ctx->state_seq_get_size(seq_id);
+    return llama_state_seq_get_size_ext(ctx, seq_id, 0);
 }
 
 size_t llama_state_seq_get_data(llama_context * ctx, uint8_t * dst, size_t size, llama_seq_id seq_id) {
+    return llama_state_seq_get_data_ext(ctx, dst, size, seq_id, 0);
+}
+
+size_t llama_state_seq_set_data(llama_context * ctx, const uint8_t * src, size_t size, llama_seq_id seq_id) {
+    return llama_state_seq_set_data_ext(ctx, src, size, seq_id, 0);
+}
+
+size_t llama_state_seq_get_size_ext(llama_context * ctx, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    return ctx->state_seq_get_size(seq_id, flags);
+}
+
+size_t llama_state_seq_get_data_ext(llama_context * ctx, uint8_t * dst, size_t size, llama_seq_id seq_id, llama_state_seq_flags flags) {
     ctx->synchronize();
 
-    return ctx->state_seq_get_data(seq_id, dst, size);
+    return ctx->state_seq_get_data(seq_id, dst, size, flags);
 }
 
-size_t llama_state_seq_set_data(llama_context * ctx, const uint8_t * src, size_t size, llama_seq_id seq_id) {
+size_t llama_state_seq_set_data_ext(llama_context * ctx, const uint8_t * src, size_t size, llama_seq_id seq_id, llama_state_seq_flags flags) {
     ctx->synchronize();
 
-    return ctx->state_seq_set_data(seq_id, src, size);
+    return ctx->state_seq_set_data(seq_id, src, size, flags);
 }
 
 size_t llama_state_seq_save_file(llama_context * ctx, const char * filepath, llama_seq_id seq_id, const llama_token * tokens, size_t n_token_count) {
diff --git a/examples/talk-llama/llama-context.h b/examples/talk-llama/llama-context.h
index 5c3a1c09886..230ef8962b8 100644
--- a/examples/talk-llama/llama-context.h
+++ b/examples/talk-llama/llama-context.h
@@ -111,9 +111,9 @@ struct llama_context {
     size_t state_get_data(      uint8_t * dst, size_t size);
     size_t state_set_data(const uint8_t * src, size_t size);
 
-    size_t state_seq_get_size(llama_seq_id seq_id);
-    size_t state_seq_get_data(llama_seq_id seq_id,       uint8_t * dst, size_t size);
-    size_t state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size);
+    size_t state_seq_get_size(llama_seq_id seq_id, llama_state_seq_flags flags);
+    size_t state_seq_get_data(llama_seq_id seq_id,       uint8_t * dst, size_t size, llama_state_seq_flags flags);
+    size_t state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size, llama_state_seq_flags flags);
 
     bool state_load_file(
             const char * filepath,
@@ -152,6 +152,7 @@ struct llama_context {
 
     void opt_init(struct llama_model * model, struct llama_opt_params lopt_params);
 
+    // TODO: more flexible combinations of logical/physical batch size and context size
     void opt_epoch(
             ggml_opt_dataset_t      dataset,
             ggml_opt_result_t       result_train,
@@ -212,8 +213,8 @@ struct llama_context {
     size_t state_write_data(llama_io_write_i & io);
     size_t state_read_data (llama_io_read_i  & io);
 
-    size_t state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id);
-    size_t state_seq_read_data (llama_io_read_i  & io, llama_seq_id seq_id);
+    size_t state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags);
+    size_t state_seq_read_data (llama_io_read_i  & io, llama_seq_id seq_id, llama_state_seq_flags flags);
 
     //
     // members
@@ -289,7 +290,10 @@ struct llama_context {
 
     // env: LLAMA_SET_ROWS (temporary)
     // ref: https://github.com/ggml-org/llama.cpp/pull/14285
-    bool supports_set_rows = false;
+    bool supports_set_rows = true;
+
+    // env: LLAMA_GRAPH_REUSE_DISABLE
+    bool graph_reuse_disable = false;
 
     // perf
     mutable int64_t t_start_us  = 0;
diff --git a/examples/talk-llama/llama-graph.cpp b/examples/talk-llama/llama-graph.cpp
index b63a41053b4..053c72d6dc8 100644
--- a/examples/talk-llama/llama-graph.cpp
+++ b/examples/talk-llama/llama-graph.cpp
@@ -188,38 +188,23 @@ void llm_graph_input_mean::set_input(const llama_ubatch * ubatch) {
 
 void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) {
     const int64_t n_tokens     = ubatch->n_tokens;
-    const int64_t n_seq_tokens = ubatch->n_seq_tokens;
     const int64_t n_seqs_unq   = ubatch->n_seqs_unq;
 
     if (cparams.embeddings && (
-            cparams.pooling_type == LLAMA_POOLING_TYPE_CLS ||
-            cparams.pooling_type == LLAMA_POOLING_TYPE_RANK
-        )) {
+        cparams.pooling_type == LLAMA_POOLING_TYPE_CLS  ||
+        cparams.pooling_type == LLAMA_POOLING_TYPE_RANK ||
+        cparams.pooling_type == LLAMA_POOLING_TYPE_LAST
+    )) {
         GGML_ASSERT(cls);
         GGML_ASSERT(ggml_backend_buffer_is_host(cls->buffer));
 
         uint32_t * data = (uint32_t *) cls->data;
         memset(cls->data, 0, n_seqs_unq*ggml_element_size(cls));
 
-        for (int i = 0; i < n_tokens; i += n_seq_tokens) {
-            for (int s = 0; s < ubatch->n_seq_id[i]; ++s) {
-                const llama_seq_id seq_id  = ubatch->seq_id[i][s];
-                const int32_t      seq_idx = ubatch->seq_idx[seq_id];
-
-                data[seq_idx] = i;
-            }
-        }
-    }
-
-    if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_LAST) {
-        GGML_ASSERT(cls);
-        GGML_ASSERT(ggml_backend_buffer_is_host(cls->buffer));
-
-        uint32_t * data = (uint32_t *) cls->data;
-        memset(cls->data, 0, n_seqs_unq*ggml_element_size(cls));
+        std::vector<int> target_pos(n_seqs_unq, -1);
+        std::vector<int> target_row(n_seqs_unq, -1);
 
-        std::vector<int> last_pos(n_seqs_unq, -1);
-        std::vector<int> last_row(n_seqs_unq, -1);
+        bool last = cparams.pooling_type == LLAMA_POOLING_TYPE_LAST;
 
         for (int i = 0; i < n_tokens; ++i) {
             const llama_pos pos = ubatch->pos[i];
@@ -228,16 +213,20 @@ void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) {
                 const llama_seq_id seq_id  = ubatch->seq_id[i][s];
                 const int32_t      seq_idx = ubatch->seq_idx[seq_id];
 
-                if (pos >= last_pos[seq_idx]) {
-                    last_pos[seq_idx] = pos;
-                    last_row[seq_idx] = i;
+                if (
+                    (target_pos[seq_idx] == -1) ||
+                    ( last && pos >= target_pos[seq_idx]) ||
+                    (!last && pos <  target_pos[seq_idx])
+                ) {
+                    target_pos[seq_idx] = pos;
+                    target_row[seq_idx] = i;
                 }
             }
         }
 
         for (int s = 0; s < n_seqs_unq; ++s) {
-            if (last_row[s] >= 0) {
-                data[s] = last_row[s];
+            if (target_row[s] >= 0) {
+                data[s] = target_row[s];
             }
         }
     }
@@ -751,6 +740,8 @@ ggml_tensor * llm_graph_context::build_ffn(
                 cur = ggml_reglu(ctx0, cur);
                 cb(cur, "ffn_reglu", il);
             } break;
+        default:
+            GGML_ABORT("fatal error");
     }
 
     if (gate && type_gate == LLM_FFN_PAR) {
@@ -760,8 +751,8 @@ ggml_tensor * llm_graph_context::build_ffn(
 
     if (down) {
         cur = build_lora_mm(down, cur);
-        if (arch == LLM_ARCH_GLM4) {
-            // GLM4 seems to have numerical issues with half-precision accumulators
+        if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) {
+            // GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators
             ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
         }
     }
@@ -796,13 +787,64 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
                 bool   scale_w,
                float   w_scale,
          llama_expert_gating_func_type gating_op,
-                 int   il) const {
+                 int   il,
+         ggml_tensor * probs_in) const {
+    return build_moe_ffn(
+        cur,
+        gate_inp,  /* gate_inp_b  */ nullptr,
+        up_exps,   /* up_exps_b   */ nullptr,
+        gate_exps, /* gate_exps_b */ nullptr,
+        down_exps, /* down_exps_b */ nullptr,
+        exp_probs_b,
+        n_expert,
+        n_expert_used,
+        type_op,
+        norm_w,
+        scale_w,
+        w_scale,
+        gating_op,
+        il,
+        probs_in
+    );
+}
+
+ggml_tensor * llm_graph_context::build_moe_ffn(
+         ggml_tensor * cur,
+         ggml_tensor * gate_inp,
+         ggml_tensor * gate_inp_b,
+         ggml_tensor * up_exps,
+         ggml_tensor * up_exps_b,
+         ggml_tensor * gate_exps,
+         ggml_tensor * gate_exps_b,
+         ggml_tensor * down_exps,
+         ggml_tensor * down_exps_b,
+         ggml_tensor * exp_probs_b,
+             int64_t   n_expert,
+             int64_t   n_expert_used,
+     llm_ffn_op_type   type_op,
+                bool   norm_w,
+                bool   scale_w,
+               float   w_scale,
+        llama_expert_gating_func_type gating_op,
+                 int   il,
+         ggml_tensor * probs_in) const {
     const int64_t n_embd   = cur->ne[0];
     const int64_t n_tokens = cur->ne[1];
     const bool weight_before_ffn = arch == LLM_ARCH_LLAMA4; // for llama4, we apply the sigmoid-ed weights before the FFN
 
-    ggml_tensor * logits = build_lora_mm(gate_inp, cur); // [n_expert, n_tokens]
-    cb(logits, "ffn_moe_logits", il);
+    ggml_tensor * logits = nullptr;
+
+    if (probs_in == nullptr) {
+        logits = build_lora_mm(gate_inp, cur); // [n_expert, n_tokens]
+        cb(logits, "ffn_moe_logits", il);
+    } else {
+        logits = probs_in;
+    }
+
+    if (gate_inp_b) {
+        logits = ggml_add(ctx0, logits, gate_inp_b);
+        cb(logits, "ffn_moe_logits_biased", il);
+    }
 
     ggml_tensor * probs = nullptr;
     switch (gating_op) {
@@ -814,6 +856,10 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
             {
                 probs = ggml_sigmoid(ctx0, logits); // [n_expert, n_tokens]
             } break;
+        case LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX_WEIGHT:
+            {
+                probs = logits; // [n_expert, n_tokens]
+            } break;
         default:
             GGML_ABORT("fatal error");
     }
@@ -842,6 +888,13 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
             ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens), selected_experts); // [1, n_expert_used, n_tokens]
     cb(weights, "ffn_moe_weights", il);
 
+    if (gating_op == LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX_WEIGHT) {
+        weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens);
+        weights = ggml_soft_max(ctx0, weights); // [n_expert_used, n_tokens]
+        weights = ggml_reshape_3d(ctx0, weights, 1, n_expert_used, n_tokens);
+        cb(weights, "ffn_moe_weights_softmax", il);
+    }
+
     if (norm_w) {
         weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens);
 
@@ -870,6 +923,11 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
     ggml_tensor * up = build_lora_mm_id(up_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
     cb(up, "ffn_moe_up", il);
 
+    if (up_exps_b) {
+        up = ggml_add_id(ctx0, up, up_exps_b, selected_experts);
+        cb(up, "ffn_moe_up_biased", il);
+    }
+
     ggml_tensor * experts = nullptr;
     if (gate_exps) {
         cur = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
@@ -878,6 +936,11 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         cur = up;
     }
 
+    if (gate_exps_b) {
+        cur = ggml_add_id(ctx0, cur, gate_exps_b, selected_experts);
+        cb(cur, "ffn_moe_gate_biased", il);
+    }
+
     switch (type_op) {
         case LLM_FFN_SILU:
             if (gate_exps) {
@@ -895,6 +958,22 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
                 cur = ggml_gelu(ctx0, cur);
                 cb(cur, "ffn_moe_gelu", il);
             } break;
+        case LLM_FFN_SWIGLU_OAI_MOE:
+            {
+                // TODO: move to hparams?
+                constexpr float alpha = 1.702f;
+                constexpr float limit = 7.0f;
+                cur = ggml_swiglu_oai(ctx0, cur, up, alpha, limit);
+                cb(cur, "ffn_moe_swiglu_oai", il);
+            } break;
+        case LLM_FFN_RELU:
+            if (gate_exps) {
+                cur = ggml_reglu_split(ctx0, cur, up);
+                cb(cur, "ffn_moe_reglu", il);
+            } else {
+                cur = ggml_relu(ctx0, cur);
+                cb(cur, "ffn_moe_relu", il);
+            } break;
         default:
             GGML_ABORT("fatal error");
     }
@@ -902,6 +981,11 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
     experts = build_lora_mm_id(down_exps, cur, selected_experts); // [n_embd, n_expert_used, n_tokens]
     cb(experts, "ffn_moe_down", il);
 
+    if (down_exps_b) {
+        experts = ggml_add_id(ctx0, experts, down_exps_b, selected_experts);
+        cb(experts, "ffn_moe_down_biased", il);
+    }
+
     if (!weight_before_ffn) {
         experts = ggml_mul(ctx0, experts, weights);
         cb(cur, "ffn_moe_weighted", il);
@@ -1140,6 +1224,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(
          ggml_tensor * kq_b,
          ggml_tensor * kq_mask,
          ggml_tensor * v_mla,
+         ggml_tensor * sinks,
              float     kq_scale) const {
     const bool v_trans = v->nb[1] > v->nb[2];
 
@@ -1176,7 +1261,8 @@ ggml_tensor * llm_graph_context::build_attn_mha(
         cur = ggml_flash_attn_ext(ctx0, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias,
                                   hparams.attn_soft_cap ? hparams.f_attn_logit_softcapping : 0.0f);
 
-        ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
+        ggml_flash_attn_ext_add_sinks(cur, sinks);
+        ggml_flash_attn_ext_set_prec (cur, GGML_PREC_F32);
 
         if (v_mla) {
 #if 0
@@ -1224,6 +1310,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(
         }
 
         kq = ggml_soft_max_ext(ctx0, kq, kq_mask, kq_scale, hparams.f_max_alibi_bias);
+        ggml_soft_max_add_sinks(kq, sinks);
 
         if (!v_trans) {
             // note: avoid this branch
@@ -1294,7 +1381,7 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_tensor * k = k_cur;
     ggml_tensor * v = v_cur;
 
-    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, nullptr, kq_scale);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1382,13 +1469,13 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_tensor * k = mctx_cur->get_k(ctx0, il);
     ggml_tensor * v = mctx_cur->get_v(ctx0, il);
 
-    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, nullptr, kq_scale);
     cb(cur, "kqv_out", il);
 
     if (wo) {
         cur = build_lora_mm(wo, cur);
-        if (arch == LLM_ARCH_GLM4) {
-            // GLM4 seems to have numerical issues with half-precision accumulators
+        if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) {
+            // GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators
             ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
         }
     }
@@ -1411,6 +1498,32 @@ ggml_tensor * llm_graph_context::build_attn(
         ggml_tensor * v_mla,
             float     kq_scale,
             int       il) const {
+    return build_attn_with_sinks(
+            inp,
+            wo,
+            wo_b,
+            q_cur,
+            k_cur,
+            v_cur,
+            kq_b,
+            v_mla,
+            nullptr,
+            kq_scale,
+            il);
+}
+
+ggml_tensor * llm_graph_context::build_attn_with_sinks(
+        llm_graph_input_attn_kv_unified_iswa * inp,
+        ggml_tensor * wo,
+        ggml_tensor * wo_b,
+        ggml_tensor * q_cur,
+        ggml_tensor * k_cur,
+        ggml_tensor * v_cur,
+        ggml_tensor * kq_b,
+        ggml_tensor * v_mla,
+        ggml_tensor * sinks,
+            float     kq_scale,
+            int       il) const {
     // these nodes are added to the graph together so that they are not reordered
     // by doing so, the number of splits in the graph is reduced
     ggml_build_forward_expand(gf, q_cur);
@@ -1448,7 +1561,7 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_tensor * k = mctx_cur->get_k(ctx0, il);
     ggml_tensor * v = mctx_cur->get_v(ctx0, il);
 
-    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, sinks, kq_scale);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1502,7 +1615,7 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_tensor * k = k_cur;
     ggml_tensor * v = v_cur;
 
-    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, nullptr, kq_scale);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1561,16 +1674,17 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif
 
 ggml_tensor * llm_graph_context::build_rs(
         ggml_tensor * s,
-        ggml_tensor * state_copy,
+        ggml_tensor * state_copy_main,
+        ggml_tensor * state_copy_extra,
             int32_t   state_size,
             int32_t   n_seqs,
-           uint32_t   n_kv,
-           uint32_t   kv_head,
-           uint32_t   kv_size,
+           uint32_t   n_rs,
+           uint32_t   rs_head,
+           uint32_t   rs_size,
             int32_t   rs_zero,
         const llm_graph_get_rows_fn & get_state_rows) const {
 
-    ggml_tensor * states = ggml_reshape_2d(ctx0, s, state_size, kv_size);
+    ggml_tensor * states = ggml_reshape_2d(ctx0, s, state_size, rs_size);
 
     // Clear a single state which will then be copied to the other cleared states.
     // Note that this is a no-op when the view is zero-sized.
@@ -1578,39 +1692,44 @@ ggml_tensor * llm_graph_context::build_rs(
     ggml_build_forward_expand(gf, ggml_scale_inplace(ctx0, state_zero, 0));
 
     // copy states
-    // NOTE: assuming the copy destinations are ALL contained between kv_head and kv_head + n_kv
-    // {state_size, kv_size} -> {state_size, n_seqs}
-    ggml_tensor * output_states = get_state_rows(ctx0, states, ggml_view_1d(ctx0, state_copy, n_seqs, 0));
+    // NOTE: assuming the copy destinations are ALL contained between rs_head and rs_head + n_rs
+    // {state_size, rs_size} -> {state_size, n_seqs}
+    ggml_tensor * output_states = get_state_rows(ctx0, states, state_copy_main);
     ggml_build_forward_expand(gf, output_states);
 
-    // copy extra states which won't be changed further (between n_seqs and n_kv)
-    ggml_tensor * states_extra = ggml_get_rows(ctx0, states, ggml_view_1d(ctx0, state_copy, n_kv - n_seqs, n_seqs*state_copy->nb[0]));
+    // copy extra states which won't be changed further (between n_seqs and n_rs)
+    ggml_tensor * states_extra = ggml_get_rows(ctx0, states, state_copy_extra);
     ggml_build_forward_expand(gf,
         ggml_cpy(ctx0,
             states_extra,
-            ggml_view_1d(ctx0, s, state_size*(n_kv - n_seqs), (kv_head + n_seqs)*state_size*ggml_element_size(s))));
+            ggml_view_1d(ctx0, s, state_size*(n_rs - n_seqs), (rs_head + n_seqs)*state_size*ggml_element_size(s))));
 
     return output_states;
 }
 
 static std::unique_ptr<llm_graph_input_rs> build_rs_inp_impl(
            ggml_context * ctx0,
+     const llama_ubatch & ubatch,
     const llama_memory_recurrent_context * mctx_cur) {
 
     auto inp = std::make_unique<llm_graph_input_rs>(mctx_cur);
 
-    const auto n_rs = mctx_cur->get_n_rs();
+    const int64_t n_rs   = mctx_cur->get_n_rs();
+    const int64_t n_seqs = ubatch.n_seqs;
 
     inp->s_copy = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_rs);
     ggml_set_input(inp->s_copy);
 
+    inp->s_copy_main  = ggml_view_1d(ctx0, inp->s_copy, n_seqs, 0);
+    inp->s_copy_extra = ggml_view_1d(ctx0, inp->s_copy, n_rs - n_seqs, n_seqs * inp->s_copy->nb[0]);
+
     return inp;
 }
 
 llm_graph_input_rs * llm_graph_context::build_rs_inp() const {
     const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);
 
-    auto inp = build_rs_inp_impl(ctx0, mctx_cur);
+    auto inp = build_rs_inp_impl(ctx0, ubatch, mctx_cur);
 
     return (llm_graph_input_rs *) res->add_input(std::move(inp));
 }
@@ -1623,7 +1742,9 @@ ggml_tensor * llm_graph_context::build_rs(
         const llm_graph_get_rows_fn & get_state_rows) const {
     const auto * kv_state = inp->mctx;
 
-    return build_rs(s, inp->s_copy, state_size, n_seqs, kv_state->get_n_rs(), kv_state->get_head(), kv_state->get_size(), kv_state->get_rs_z(), get_state_rows);
+    return build_rs(s, inp->s_copy_main, inp->s_copy_extra, state_size, n_seqs,
+                    kv_state->get_n_rs(), kv_state->get_head(), kv_state->get_size(), kv_state->get_rs_z(),
+                    get_state_rows);
 }
 
 ggml_tensor * llm_graph_context::build_rwkv_token_shift_load(
@@ -1670,7 +1791,7 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_store(
 llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const {
     const auto * mctx_cur = static_cast<const llama_memory_hybrid_context *>(mctx);
 
-    auto inp_rs   = build_rs_inp_impl(ctx0, mctx_cur->get_recr());
+    auto inp_rs   = build_rs_inp_impl(ctx0, ubatch, mctx_cur->get_recr());
     auto inp_attn = build_attn_inp_kv_unified_impl(ctx0, ubatch, hparams, cparams, mctx_cur->get_attn());
 
     auto inp = std::make_unique<llm_graph_input_mem_hybrid>(std::move(inp_attn), std::move(inp_rs), mctx_cur);
diff --git a/examples/talk-llama/llama-graph.h b/examples/talk-llama/llama-graph.h
index a28a8c4bdda..6ff49de3a1c 100644
--- a/examples/talk-llama/llama-graph.h
+++ b/examples/talk-llama/llama-graph.h
@@ -39,6 +39,7 @@ enum llm_ffn_op_type {
     LLM_FFN_SWIGLU,
     LLM_FFN_GEGLU,
     LLM_FFN_REGLU,
+    LLM_FFN_SWIGLU_OAI_MOE,
 };
 
 enum llm_ffn_gate_type {
@@ -144,7 +145,7 @@ class llm_graph_input_pos_bucket : public llm_graph_input_i {
 
     ggml_tensor * pos_bucket = nullptr; // I32 [n_batch, n_batch]
 
-    const llama_hparams & hparams;
+    const llama_hparams hparams;
 };
 
 class llm_graph_input_pos_bucket_kv : public llm_graph_input_i {
@@ -158,7 +159,7 @@ class llm_graph_input_pos_bucket_kv : public llm_graph_input_i {
 
     ggml_tensor * pos_bucket = nullptr; // I32 [n_kv, n_batch]
 
-    const llama_hparams & hparams;
+    const llama_hparams hparams;
 
     const llama_kv_cache_unified_context * mctx;
 };
@@ -177,8 +178,8 @@ class llm_graph_input_out_ids : public llm_graph_input_i {
 
     ggml_tensor * out_ids; // I32 [n_outputs]
 
-    const llama_hparams & hparams;
-    const llama_cparams & cparams;
+    const llama_hparams hparams;
+    const llama_cparams cparams;
 
     const uint32_t n_outputs;
 };
@@ -192,7 +193,7 @@ class llm_graph_input_mean : public llm_graph_input_i {
 
     ggml_tensor * mean; // F32 [n_batch, n_batch]
 
-    const llama_cparams & cparams;
+    const llama_cparams cparams;
 };
 
 class llm_graph_input_cls : public llm_graph_input_i {
@@ -204,7 +205,7 @@ class llm_graph_input_cls : public llm_graph_input_i {
 
     ggml_tensor * cls; // I32 [n_batch]
 
-    const llama_cparams & cparams;
+    const llama_cparams cparams;
 };
 
 class llm_graph_input_rs : public llm_graph_input_i {
@@ -214,7 +215,12 @@ class llm_graph_input_rs : public llm_graph_input_i {
 
     void set_input(const llama_ubatch * ubatch) override;
 
-    ggml_tensor * s_copy; // I32 [kv_size]
+    ggml_tensor * s_copy;  // I32 [n_rs]
+
+    // views of s_copy, computed once per graph
+    // and shared across layers which use build_rs
+    ggml_tensor * s_copy_main;   // I32 [n_seqs]
+    ggml_tensor * s_copy_extra;  // I32 [n_rs - n_seqs]
 
     const llama_memory_recurrent_context * mctx;
 };
@@ -247,8 +253,8 @@ class llm_graph_input_attn_no_cache : public llm_graph_input_i {
     ggml_tensor * kq_mask     = nullptr; // F32 [n_tokens, n_batch, 1, 1]
     ggml_tensor * kq_mask_cnv = nullptr; //     [n_tokens, n_batch, 1, 1]
 
-    const llama_hparams & hparams;
-    const llama_cparams & cparams;
+    const llama_hparams hparams;
+    const llama_cparams cparams;
 };
 
 class llm_graph_input_attn_kv_unified : public llm_graph_input_i {
@@ -278,8 +284,11 @@ class llm_graph_input_attn_kv_unified : public llm_graph_input_i {
     ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
     ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch/n_stream, 1, n_stream]
 
-    const llama_hparams & hparams;
-    const llama_cparams & cparams;
+    // note: these have to be copies because in order to be able to reuse a graph, its inputs
+    //       need to carry these parameters with them. otherwise, they can point to freed
+    //       llm_graph_params from a previous batch, causing stack-use-after-return
+    const llama_hparams hparams;
+    const llama_cparams cparams;
 
     const llama_kv_cache_unified_context * mctx;
 };
@@ -318,8 +327,8 @@ class llm_graph_input_attn_kv_unified_iswa : public llm_graph_input_i {
     ggml_tensor * self_kq_mask_swa     = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
     ggml_tensor * self_kq_mask_swa_cnv = nullptr; //     [n_kv, n_batch/n_stream, 1, n_stream]
 
-    const llama_hparams & hparams;
-    const llama_cparams & cparams;
+    const llama_hparams hparams;
+    const llama_cparams cparams;
 
     const llama_kv_cache_unified_iswa_context * mctx;
 };
@@ -415,7 +424,9 @@ struct llm_graph_params {
                 (!ubatch.embd  && !other.ubatch.embd)
             );
 
-        if (can_reuse_ubatch && !ubatch.equal_seqs()) {
+        // when we split the batch using "equal_seqs" we have to verify that the participating sequences are the same
+        //   the reason is because the set of attention streams would be different for different sequences
+        if (can_reuse_ubatch && ubatch.equal_seqs()) {
             if (!ubatch.data) {
                 // if the old ubatch does not own it's data, then we cannot guarantee that it is still alive, and
                 //   therefore we cannot perform the sequence id check. normally should never happen
@@ -609,6 +620,7 @@ struct llm_graph_context {
        llm_ffn_gate_type   type_gate,
                      int   il) const;
 
+    // build MoE FFN without bias tensors
     ggml_tensor * build_moe_ffn(
              ggml_tensor * cur,
              ggml_tensor * gate_inp,
@@ -623,7 +635,29 @@ struct llm_graph_context {
                     bool   scale_w,
                    float   w_scale,
             llama_expert_gating_func_type gating_op,
-                     int   il) const;
+                     int   il,
+             ggml_tensor * probs_in = nullptr) const;
+
+    ggml_tensor * build_moe_ffn(
+             ggml_tensor * cur,
+             ggml_tensor * gate_inp,
+             ggml_tensor * gate_inp_b,
+             ggml_tensor * up_exps,
+             ggml_tensor * up_exps_b,
+             ggml_tensor * gate_exps,
+             ggml_tensor * gate_exps_b,
+             ggml_tensor * down_exps,
+             ggml_tensor * down_exps_b,
+             ggml_tensor * exp_probs_b,
+                 int64_t   n_expert,
+                 int64_t   n_expert_used,
+         llm_ffn_op_type   type_op,
+                    bool   norm_w,
+                    bool   scale_w,
+                   float   w_scale,
+            llama_expert_gating_func_type gating_op,
+                     int   il,
+             ggml_tensor * probs_in = nullptr) const;
 
     //
     // inputs
@@ -651,6 +685,7 @@ struct llm_graph_context {
              ggml_tensor * v,       // [n_embd_head_v, n_head_v, n_tokens] (v_trans == false)
              ggml_tensor * kq_b,
              ggml_tensor * kq_mask,
+             ggml_tensor * sinks,
              ggml_tensor * v_mla,   // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                    float   kq_scale) const;
 
@@ -697,6 +732,20 @@ struct llm_graph_context {
                   float   kq_scale,
                     int   il) const;
 
+    // TODO: temporary to keep the diff small. after the code is public will refactor to simplify this
+    ggml_tensor * build_attn_with_sinks(
+            llm_graph_input_attn_kv_unified_iswa * inp,
+            ggml_tensor * wo,
+            ggml_tensor * wo_b,
+            ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
+            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens] optional
+            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] optional
+            ggml_tensor * kq_b,
+            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
+            ggml_tensor * sinks, // [n_head_q]
+                  float   kq_scale,
+                    int   il) const;
+
     llm_graph_input_attn_cross * build_attn_inp_cross() const;
 
     ggml_tensor * build_attn(
@@ -715,7 +764,6 @@ struct llm_graph_context {
     // recurrent
     //
 
-    // TODO: avoid notion of "kv"
     // TODO: move this implementation to llama_memory_recurrent.
     //       this is analogous to llama_kv_cache_unified::cpy_k / cpy_v
     //       when moving, avoid passing `ggml_cgraph` - only pass `ggml_context`. would likely need to split the
@@ -723,12 +771,13 @@ struct llm_graph_context {
     //         `llama_memory_recurrent`
     ggml_tensor * build_rs(
             ggml_tensor * s,
-            ggml_tensor * state_copy,
+            ggml_tensor * state_copy_main,
+            ggml_tensor * state_copy_extra,
                 int32_t   state_size,
                 int32_t   n_seqs,
-               uint32_t   n_kv,
-               uint32_t   kv_head,
-               uint32_t   kv_size,
+               uint32_t   n_rs,
+               uint32_t   rs_head,
+               uint32_t   rs_size,
                 int32_t   rs_zero,
             const llm_graph_get_rows_fn & get_state_rows = ggml_get_rows) const;
 
diff --git a/examples/talk-llama/llama-hparams.cpp b/examples/talk-llama/llama-hparams.cpp
index c6c67d26f93..7a06368dcda 100644
--- a/examples/talk-llama/llama-hparams.cpp
+++ b/examples/talk-llama/llama-hparams.cpp
@@ -2,9 +2,15 @@
 
 #include "ggml.h"
 
-void llama_hparams::set_swa_pattern(uint32_t n_pattern) {
-    for (uint32_t il = 0; il < n_layer; ++il) {
-        swa_layers[il] = n_pattern == 0 || (il % n_pattern < (n_pattern - 1));
+void llama_hparams::set_swa_pattern(uint32_t n_pattern, bool dense_first) {
+    if (dense_first) {
+        for (uint32_t il = 0; il < n_layer; ++il) {
+            swa_layers[il] = n_pattern == 0 || (il % n_pattern != 0);
+        }
+    } else {
+        for (uint32_t il = 0; il < n_layer; ++il) {
+            swa_layers[il] = n_pattern == 0 || (il % n_pattern < (n_pattern - 1));
+        }
     }
 }
 
diff --git a/examples/talk-llama/llama-hparams.h b/examples/talk-llama/llama-hparams.h
index ec7fd6a42bf..bd231224432 100644
--- a/examples/talk-llama/llama-hparams.h
+++ b/examples/talk-llama/llama-hparams.h
@@ -9,9 +9,10 @@
 #define LLAMA_MAX_EXPERTS 384  // Kimi-K2
 
 enum llama_expert_gating_func_type {
-    LLAMA_EXPERT_GATING_FUNC_TYPE_NONE    = 0,
-    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX = 1,
-    LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID = 2,
+    LLAMA_EXPERT_GATING_FUNC_TYPE_NONE           = 0,
+    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX        = 1,
+    LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID        = 2,
+    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX_WEIGHT = 3, // applied to the router weights instead of the logits
 };
 
 enum llama_swa_type {
@@ -73,6 +74,7 @@ struct llama_hparams {
     bool     expert_weights_norm  = false;
     uint32_t expert_gating_func   = LLAMA_EXPERT_GATING_FUNC_TYPE_NONE;
     uint32_t moe_every_n_layers   = 0;
+    uint32_t nextn_predict_layers = 0;
 
     float f_norm_eps;
     float f_norm_rms_eps;
@@ -140,7 +142,7 @@ struct llama_hparams {
     // for Classifiers
     uint32_t n_cls_out = 1;
 
-    // llama4
+    // llama4 smallthinker
     uint32_t n_moe_layer_step        = 0;
     uint32_t n_no_rope_layer_step    = 4;
     uint32_t n_attn_temp_floor_scale = 8192;
@@ -161,9 +163,10 @@ struct llama_hparams {
     enum llama_rope_scaling_type rope_scaling_type_train = LLAMA_ROPE_SCALING_TYPE_NONE;
 
     // this value n_pattern means that every nth layer is dense (i.e. non-SWA)
+    // dense_first means whether the pattern is start with a dense layer
     // note that if n_pattern == 0, all layers are SWA
     //           if n_pattern == 1, all layers are dense
-    // example: n_pattern = 3
+    // example 1: n_pattern = 3, dense_first = false
     //   il == 0: swa
     //   il == 1: swa
     //   il == 2: dense
@@ -172,7 +175,13 @@ struct llama_hparams {
     //   il == 5: dense
     //   il == 6: swa
     //   etc ...
-    void set_swa_pattern(uint32_t n_pattern);
+    // example 2: n_pattern = 2, dense_first = true
+    //   il == 0: dense
+    //   il == 1: swa
+    //   il == 2: dense
+    //   il == 3: swa
+    //   etc ...
+    void set_swa_pattern(uint32_t n_pattern, bool dense_first = false);
 
     // return true if one of the layers is SWA
     bool is_swa_any() const;
diff --git a/examples/talk-llama/llama-kv-cache-unified-iswa.cpp b/examples/talk-llama/llama-kv-cache-unified-iswa.cpp
index 01d27fb4db9..1e363fff2a5 100644
--- a/examples/talk-llama/llama-kv-cache-unified-iswa.cpp
+++ b/examples/talk-llama/llama-kv-cache-unified-iswa.cpp
@@ -194,14 +194,20 @@ bool llama_kv_cache_unified_iswa::get_can_shift() const {
     return kv_base->get_size() == kv_swa->get_size();
 }
 
-void llama_kv_cache_unified_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {
-    kv_base->state_write(io, seq_id);
-    kv_swa ->state_write(io, seq_id);
+void llama_kv_cache_unified_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
+    if ((flags & LLAMA_STATE_SEQ_FLAGS_SWA_ONLY) == 0) {
+        kv_base->state_write(io, seq_id, flags);
+    }
+
+    kv_swa->state_write(io, seq_id, flags);
 }
 
-void llama_kv_cache_unified_iswa::state_read(llama_io_read_i & io, llama_seq_id seq_id) {
-    kv_base->state_read(io, seq_id);
-    kv_swa ->state_read(io, seq_id);
+void llama_kv_cache_unified_iswa::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    if ((flags & LLAMA_STATE_SEQ_FLAGS_SWA_ONLY) == 0) {
+        kv_base->state_read(io, seq_id, flags);
+    }
+
+    kv_swa->state_read(io, seq_id, flags);
 }
 
 llama_kv_cache_unified * llama_kv_cache_unified_iswa::get_base() const {
diff --git a/examples/talk-llama/llama-kv-cache-unified-iswa.h b/examples/talk-llama/llama-kv-cache-unified-iswa.h
index d2650dadd35..7bc4df718d3 100644
--- a/examples/talk-llama/llama-kv-cache-unified-iswa.h
+++ b/examples/talk-llama/llama-kv-cache-unified-iswa.h
@@ -56,8 +56,8 @@ class llama_kv_cache_unified_iswa : public llama_memory_i {
 
     // state write/load
 
-    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
-    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1)       override;
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override;
 
     //
     // llama_kv_cache_unified_iswa specific API
diff --git a/examples/talk-llama/llama-kv-cache-unified.cpp b/examples/talk-llama/llama-kv-cache-unified.cpp
index 321dc79fc36..478ebffac0f 100644
--- a/examples/talk-llama/llama-kv-cache-unified.cpp
+++ b/examples/talk-llama/llama-kv-cache-unified.cpp
@@ -39,6 +39,10 @@ llama_kv_cache_unified::llama_kv_cache_unified(
     if (model.arch == LLM_ARCH_GEMMA3N) {
         n_layer_cache = 20;
     }
+    if (model.arch == LLM_ARCH_GLM4_MOE) {
+        // GLM-4.5: Only process up to last layer, skip final NextN layer
+        n_layer_cache = hparams.n_layer - hparams.nextn_predict_layers;
+    }
 
     // create a context for each buffer type
     std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
@@ -183,7 +187,7 @@ llama_kv_cache_unified::llama_kv_cache_unified(
         const size_t memory_size_k = size_k_bytes();
         const size_t memory_size_v = size_v_bytes();
 
-        LLAMA_LOG_INFO("%s: size = %7.2f MiB (%6u cells, %3d layers, %2u/%2u seqs), K (%s): %7.2f MiB, V (%s): %7.2f MiB\n", __func__,
+        LLAMA_LOG_INFO("%s: size = %7.2f MiB (%6u cells, %3d layers, %2u/%u seqs), K (%s): %7.2f MiB, V (%s): %7.2f MiB\n", __func__,
                 (float)(memory_size_k + memory_size_v) / (1024.0f * 1024.0f), kv_size, (int) layers.size(), n_seq_max, n_stream,
                 ggml_type_name(type_k), (float)memory_size_k / (1024.0f * 1024.0f),
                 ggml_type_name(type_v), (float)memory_size_v / (1024.0f * 1024.0f));
@@ -193,7 +197,7 @@ llama_kv_cache_unified::llama_kv_cache_unified(
     debug = LLAMA_KV_CACHE_DEBUG ? atoi(LLAMA_KV_CACHE_DEBUG) : 0;
 
     const char * LLAMA_SET_ROWS = getenv("LLAMA_SET_ROWS");
-    supports_set_rows = LLAMA_SET_ROWS ? atoi(LLAMA_SET_ROWS) != 0 : 0;
+    supports_set_rows = LLAMA_SET_ROWS ? atoi(LLAMA_SET_ROWS) != 0 : supports_set_rows;
 
     if (!supports_set_rows) {
         // ref: https://github.com/ggml-org/llama.cpp/pull/14363
@@ -219,12 +223,7 @@ void llama_kv_cache_unified::clear(bool data) {
 }
 
 bool llama_kv_cache_unified::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
-    GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size());
-
-    auto & cells = v_cells[seq_to_stream[seq_id]];
-    auto & head  = v_heads[seq_to_stream[seq_id]];
-
-    uint32_t new_head = cells.size();
+    GGML_ASSERT(seq_id == -1 || (seq_id >= 0 && (size_t) seq_id < seq_to_stream.size()));
 
     if (p0 < 0) {
         p0 = 0;
@@ -235,6 +234,11 @@ bool llama_kv_cache_unified::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
     }
 
     if (seq_id >= 0) {
+        auto & cells = v_cells[seq_to_stream[seq_id]];
+        auto & head  = v_heads[seq_to_stream[seq_id]];
+
+        uint32_t new_head = cells.size();
+
         for (uint32_t i = 0; i < cells.size(); ++i) {
             if (!cells.pos_in(i, p0, p1)) {
                 continue;
@@ -246,24 +250,36 @@ bool llama_kv_cache_unified::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
                 }
             }
         }
+
+        // If we freed up a slot, set head to it so searching can start there.
+        if (new_head != cells.size() && new_head < head) {
+            head = new_head;
+        }
     } else {
         // match any sequence
-        for (uint32_t i = 0; i < cells.size(); ++i) {
-            if (!cells.pos_in(i, p0, p1)) {
-                continue;
-            }
+        for (uint32_t s = 0; s < n_stream; ++s) {
+            auto & cells = v_cells[s];
+            auto & head  = v_heads[s];
 
-            cells.rm(i);
+            uint32_t new_head = cells.size();
 
-            if (new_head == cells.size()) {
-                new_head = i;
+            for (uint32_t i = 0; i < cells.size(); ++i) {
+                if (!cells.pos_in(i, p0, p1)) {
+                    continue;
+                }
+
+                cells.rm(i);
+
+                if (new_head == cells.size()) {
+                    new_head = i;
+                }
             }
-        }
-    }
 
-    // If we freed up a slot, set head to it so searching can start there.
-    if (new_head != cells.size() && new_head < head) {
-        head = new_head;
+            // If we freed up a slot, set head to it so searching can start there.
+            if (new_head != cells.size() && new_head < head) {
+                head = new_head;
+            }
+        }
     }
 
     return true;
@@ -734,66 +750,70 @@ bool llama_kv_cache_unified::update(llama_context * lctx, bool do_shift, const d
 }
 
 llama_kv_cache_unified::slot_info llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch, bool cont) const {
-    if (debug > 0) {
-        const auto & cells = v_cells[seq_to_stream[1]];
 
-        const uint32_t head_cur = v_heads[1];
-
-        LLAMA_LOG_DEBUG("%s: n = %5d, used = %5d, head = %5d, size = %5d, n_swa = %5d\n",
-                __func__, cells.used_max_p1(), cells.get_used(), head_cur, get_size(), n_swa);
+    if (debug > 0) {
+        for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
+            const auto seq_id = ubatch.seq_id_unq[s];
+            const auto stream_id = seq_to_stream[seq_id];
+            const auto & cells = v_cells[stream_id];
+            const uint32_t head_cur = v_heads[stream_id];
+
+            LLAMA_LOG_DEBUG("%s: stream[%d], n = %5d, used = %5d, head = %5d, size = %5d, n_swa = %5d\n",
+                    __func__, stream_id, cells.used_max_p1(), cells.get_used(), head_cur, get_size(), n_swa);
+
+            if ((debug == 2 && n_swa > 0) || debug > 2) {
+                std::string ss;
+                for (uint32_t i = 0; i < cells.size(); ++i) {
+                    if (cells.is_empty(i)) {
+                        ss += '.';
+                    } else {
+                        assert(cells.seq_count(i) >= 1);
 
-        if ((debug == 2 && n_swa > 0) || debug > 2) {
-            std::string ss;
-            for (uint32_t i = 0; i < cells.size(); ++i) {
-                if (cells.is_empty(i)) {
-                    ss += '.';
-                } else {
-                    assert(cells.seq_count(i) >= 1);
+                        if (cells.seq_count(i) == 1) {
+                            ss += std::to_string(cells.seq_get(i));
+                        } else {
+                            ss += 'M';
+                        }
+                    }
+                    if (i%256 == 255) {
+                        ss += " *";
+                        ss += '\n';
+                    }
+                }
+                LLAMA_LOG_DEBUG("\n%s\n", ss.c_str());
+            }
 
-                    if (cells.seq_count(i) == 1) {
-                        ss += std::to_string(cells.seq_get(i));
+            if ((debug == 2 && n_swa > 0) || debug > 2) {
+                std::string ss;
+                for (uint32_t i = 0; i < cells.size(); ++i) {
+                    std::string cur;
+                    if (cells.is_empty(i)) {
+                        cur = '.';
                     } else {
-                        ss += 'M';
+                        cur = std::to_string(cells.pos_get(i));
+                    }
+                    const int n = cur.size();
+                    for (int j = 0; j < 5 - n; ++j) {
+                        cur += ' ';
+                    }
+                    ss += cur;
+                    if (i%256 == 255) {
+                        ss += " *";
+                    }
+                    if (i%64 == 63) {
+                        ss += '\n';
                     }
                 }
-                if (i%256 == 255) {
-                    ss += " *";
-                    ss += '\n';
-                }
+                LLAMA_LOG_DEBUG("\n%s\n", ss.c_str());
             }
-            LLAMA_LOG_DEBUG("\n%s\n", ss.c_str());
-        }
 
-        if ((debug == 2 && n_swa > 0) || debug > 2) {
-            std::string ss;
-            for (uint32_t i = 0; i < cells.size(); ++i) {
-                std::string cur;
-                if (cells.is_empty(i)) {
-                    cur = '.';
-                } else {
-                    cur = std::to_string(cells.pos_get(i));
-                }
-                const int n = cur.size();
-                for (int j = 0; j < 5 - n; ++j) {
-                    cur += ' ';
-                }
-                ss += cur;
-                if (i%256 == 255) {
-                    ss += " *";
-                }
-                if (i%64 == 63) {
-                    ss += '\n';
+            for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
+                if (cells.seq_pos_min(s) < 0) {
+                    continue;
                 }
-            }
-            LLAMA_LOG_DEBUG("\n%s\n", ss.c_str());
-        }
 
-        for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
-            if (cells.seq_pos_min(s) < 0) {
-                continue;
+                LLAMA_LOG_DEBUG("%s: stream[%d] min[%d] = %5d, max[%d] = %5d\n", __func__, stream_id, s, cells.seq_pos_min(s), s, cells.seq_pos_max(s));
             }
-
-            LLAMA_LOG_DEBUG("%s: min[%d] = %5d, max[%d] = %5d\n", __func__, s, cells.seq_pos_min(s), s, cells.seq_pos_max(s));
         }
     }
 
@@ -1808,7 +1828,9 @@ bool llama_kv_cache_unified::is_masked_swa(llama_pos p0, llama_pos p1) const {
     return false;
 }
 
-void llama_kv_cache_unified::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {
+void llama_kv_cache_unified::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
+    GGML_UNUSED(flags);
+
     io.write(&n_stream, sizeof(n_stream));
 
     for (uint32_t s = 0; s < n_stream; ++s) {
@@ -1859,7 +1881,9 @@ void llama_kv_cache_unified::state_write(llama_io_write_i & io, llama_seq_id seq
     }
 }
 
-void llama_kv_cache_unified::state_read(llama_io_read_i & io, llama_seq_id seq_id) {
+void llama_kv_cache_unified::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    GGML_UNUSED(flags);
+
     GGML_ASSERT(seq_id == -1 || (seq_id >= 0 && (size_t) seq_id < seq_to_stream.size()));
 
     uint32_t n_stream_cur;
diff --git a/examples/talk-llama/llama-kv-cache-unified.h b/examples/talk-llama/llama-kv-cache-unified.h
index 3e28e346c3f..07a7c9e4e46 100644
--- a/examples/talk-llama/llama-kv-cache-unified.h
+++ b/examples/talk-llama/llama-kv-cache-unified.h
@@ -136,8 +136,8 @@ class llama_kv_cache_unified : public llama_memory_i {
 
     // state write/load
 
-    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
-    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1)       override;
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override;
 
     //
     // llama_kv_cache_unified specific API
@@ -230,7 +230,7 @@ class llama_kv_cache_unified : public llama_memory_i {
 
     // env: LLAMA_SET_ROWS (temporary)
     // ref: https://github.com/ggml-org/llama.cpp/pull/14285
-    bool supports_set_rows = false;
+    bool supports_set_rows = true;
 
     const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
 
diff --git a/examples/talk-llama/llama-memory-hybrid.cpp b/examples/talk-llama/llama-memory-hybrid.cpp
index d8e2086c875..cbeeb21344e 100644
--- a/examples/talk-llama/llama-memory-hybrid.cpp
+++ b/examples/talk-llama/llama-memory-hybrid.cpp
@@ -25,6 +25,7 @@ llama_memory_hybrid::llama_memory_hybrid(
                          /* common */
              uint32_t    n_seq_max,
                  bool    offload,
+                 bool    unified,
                          /* layer filters */
       layer_filter_cb && filter_attn,
       layer_filter_cb && filter_recr) :
@@ -38,7 +39,7 @@ llama_memory_hybrid::llama_memory_hybrid(
         type_v,
         v_trans,
         offload,
-        1,
+        unified,
         kv_size,
         n_seq_max,
         n_pad,
@@ -164,12 +165,16 @@ llama_pos llama_memory_hybrid::seq_pos_max(llama_seq_id seq_id) const {
     return std::min(mem_attn->seq_pos_max(seq_id), mem_recr->seq_pos_max(seq_id));
 }
 
-void llama_memory_hybrid::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {
+void llama_memory_hybrid::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
+    GGML_UNUSED(flags);
+
     mem_attn->state_write(io, seq_id);
     mem_recr->state_write(io, seq_id);
 }
 
-void llama_memory_hybrid::state_read(llama_io_read_i & io, llama_seq_id seq_id) {
+void llama_memory_hybrid::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    GGML_UNUSED(flags);
+
     mem_attn->state_read(io, seq_id);
     mem_recr->state_read(io, seq_id);
 }
diff --git a/examples/talk-llama/llama-memory-hybrid.h b/examples/talk-llama/llama-memory-hybrid.h
index 4ac31817578..acdbc26bfb6 100644
--- a/examples/talk-llama/llama-memory-hybrid.h
+++ b/examples/talk-llama/llama-memory-hybrid.h
@@ -39,6 +39,7 @@ class llama_memory_hybrid : public llama_memory_i {
                              /* common */
                  uint32_t    n_seq_max,
                      bool    offload,
+                     bool    unified,
                              /* layer filters */
           layer_filter_cb && filter_attn = nullptr,
           layer_filter_cb && filter_recr = nullptr);
@@ -73,8 +74,8 @@ class llama_memory_hybrid : public llama_memory_i {
 
     // state write/load
 
-    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
-    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1)       override;
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0)       override;
 
     //
     // llama_memory_hybrid specific API
diff --git a/examples/talk-llama/llama-memory-recurrent.cpp b/examples/talk-llama/llama-memory-recurrent.cpp
index c0c2ec084dc..849675c4188 100644
--- a/examples/talk-llama/llama-memory-recurrent.cpp
+++ b/examples/talk-llama/llama-memory-recurrent.cpp
@@ -680,7 +680,9 @@ size_t llama_memory_recurrent::size_s_bytes() const {
     return size_s_bytes;
 }
 
-void llama_memory_recurrent::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {
+void llama_memory_recurrent::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
+    GGML_UNUSED(flags);
+
     std::vector<std::pair<uint32_t, uint32_t>> cell_ranges; // ranges, from inclusive, to exclusive
     uint32_t cell_count = 0;
 
@@ -718,7 +720,9 @@ void llama_memory_recurrent::state_write(llama_io_write_i & io, llama_seq_id seq
     state_write_data(io, cell_ranges);
 }
 
-void llama_memory_recurrent::state_read(llama_io_read_i & io, llama_seq_id seq_id) {
+void llama_memory_recurrent::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    GGML_UNUSED(flags);
+
     uint32_t cell_count;
     io.read_to(&cell_count, sizeof(cell_count));
 
diff --git a/examples/talk-llama/llama-memory-recurrent.h b/examples/talk-llama/llama-memory-recurrent.h
index 4d094f9a057..95c617b2c94 100644
--- a/examples/talk-llama/llama-memory-recurrent.h
+++ b/examples/talk-llama/llama-memory-recurrent.h
@@ -63,8 +63,8 @@ class llama_memory_recurrent : public llama_memory_i {
 
     // state write/load
 
-    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
-    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) override;
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override;
 
     uint32_t head = 0; // the location where the batch will be placed in the cache (see find_slot())
     uint32_t size = 0; // total number of cells, shared across all sequences
diff --git a/examples/talk-llama/llama-memory.h b/examples/talk-llama/llama-memory.h
index e8ba336e852..171d312cc99 100644
--- a/examples/talk-llama/llama-memory.h
+++ b/examples/talk-llama/llama-memory.h
@@ -104,8 +104,8 @@ struct llama_memory_i {
     // state write/read
     //
 
-    virtual void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const = 0;
-    virtual void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) = 0;
+    virtual void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const = 0;
+    virtual void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) = 0;
 };
 
 using llama_memory_ptr = std::unique_ptr<llama_memory_i>;
diff --git a/examples/talk-llama/llama-model-loader.cpp b/examples/talk-llama/llama-model-loader.cpp
index bd9e6da8832..f71c40f8e3f 100644
--- a/examples/talk-llama/llama-model-loader.cpp
+++ b/examples/talk-llama/llama-model-loader.cpp
@@ -35,6 +35,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_Q5_0:     return "Q5_0";
         case LLAMA_FTYPE_MOSTLY_Q5_1:     return "Q5_1";
         case LLAMA_FTYPE_MOSTLY_Q8_0:     return "Q8_0";
+        case LLAMA_FTYPE_MOSTLY_MXFP4_MOE: return "MXFP4 MoE";
         case LLAMA_FTYPE_MOSTLY_Q2_K:     return "Q2_K - Medium";
         case LLAMA_FTYPE_MOSTLY_Q2_K_S:   return "Q2_K - Small";
         case LLAMA_FTYPE_MOSTLY_Q3_K_S:   return "Q3_K - Small";
diff --git a/examples/talk-llama/llama-model-loader.h b/examples/talk-llama/llama-model-loader.h
index 0f52b011b69..c9189f6cb44 100644
--- a/examples/talk-llama/llama-model-loader.h
+++ b/examples/talk-llama/llama-model-loader.h
@@ -58,8 +58,9 @@ struct llama_model_loader {
         }
     };
 
-    static const int TENSOR_NOT_REQUIRED = 1;
-    static const int TENSOR_DUPLICATED   = 2;
+    static const int TENSOR_NOT_REQUIRED = 1 << 0;
+    static const int TENSOR_DUPLICATED   = 1 << 1;
+    static const int TENSOR_SKIP         = 1 << 2;
 
     int n_kv      = 0;
     int n_tensors = 0;
diff --git a/examples/talk-llama/llama-model.cpp b/examples/talk-llama/llama-model.cpp
index 71f89e19072..23a26f0c64e 100644
--- a/examples/talk-llama/llama-model.cpp
+++ b/examples/talk-llama/llama-model.cpp
@@ -109,8 +109,10 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_A13B:          return "A13B";
         case LLM_TYPE_21B_A3B:       return "21B.A3B";
         case LLM_TYPE_30B_A3B:       return "30B.A3B";
+        case LLM_TYPE_106B_A12B:     return "106B.A12B";
         case LLM_TYPE_235B_A22B:     return "235B.A22B";
         case LLM_TYPE_300B_A47B:     return "300B.A47B";
+        case LLM_TYPE_355B_A32B:     return "355B.A32B";
         case LLM_TYPE_E2B:           return "E2B";
         case LLM_TYPE_E4B:           return "E4B";
         default:                     return "?B";
@@ -190,6 +192,13 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w
                 ggml_tensor * a = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, w->ne[0], w->ne[1], w->ne[2], w->ne[3]);
                 op_tensor = ggml_add(ctx, a, w);
             } break;
+        case GGML_OP_ADD_ID:
+            {
+                int n_expert_used = hparams.n_expert_used;
+                ggml_tensor * a = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, w->ne[0], n_expert_used, 512);
+                ggml_tensor * c = ggml_new_tensor_2d(ctx, GGML_TYPE_I32, n_expert_used, 512);
+                op_tensor = ggml_add_id(ctx, a, w, c);
+            } break;
         case GGML_OP_MUL:
             {
                 ggml_tensor * a = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, w->ne[0], w->ne[1], w->ne[2], w->ne[3]);
@@ -258,6 +267,10 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w
                 ggml_tensor * b = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, n_embd, w->ne[1], 1, 1);
                 op_tensor = ggml_im2col(ctx, w, b, 1, 0, 0, 0, 1, 0, false, GGML_TYPE_F16);
             } break;
+        case GGML_OP_SCALE:
+            {
+                op_tensor = ggml_scale(ctx, w, 1.0f);
+            } break;
         default:
             GGML_ABORT("%s: missing test for op %s for tensor %s", __func__, ggml_op_name(op), w->name);
     }
@@ -290,7 +303,7 @@ static ggml_backend_buffer_type_t select_weight_buft(const llama_hparams & hpara
 }
 
 // CPU: ACCEL -> GPU host -> CPU extra -> CPU
-static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & devices) {
+static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & devices, bool use_extra_bufts) {
     buft_list_t buft_list;
 
     // add ACCEL buffer types
@@ -319,21 +332,22 @@ static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & de
         }
     }
 
-    // add extra buffer types, only if no GPU device is present
-    // ref: https://github.com/ggml-org/llama.cpp/issues/12481#issuecomment-2743136094
-    auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
-    if (cpu_dev == nullptr) {
-        throw std::runtime_error(format("%s: no CPU backend found", __func__));
-    }
+    // add extra buffer types
+    if (use_extra_bufts) {
+        auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        if (cpu_dev == nullptr) {
+            throw std::runtime_error(format("%s: no CPU backend found", __func__));
+        }
 
-    auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
-    auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
-        ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
-    if (ggml_backend_dev_get_extra_bufts_fn) {
-        ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev);
-        while (extra_bufts && *extra_bufts) {
-            buft_list.emplace_back(cpu_dev, *extra_bufts);
-            ++extra_bufts;
+        auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
+        auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
+            ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
+        if (ggml_backend_dev_get_extra_bufts_fn) {
+            ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev);
+            while (extra_bufts && *extra_bufts) {
+                buft_list.emplace_back(cpu_dev, *extra_bufts);
+                ++extra_bufts;
+            }
         }
     }
 
@@ -869,6 +883,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 hparams.causal_attn = false;
             }
             break;
+        case LLM_ARCH_LLADA:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                // LLaDA-8B has 32 layers, similar to LLaMA but for diffusion
+                switch (hparams.n_layer) {
+                    case 32:
+                        type = LLM_TYPE_8B;
+                        break;
+                    default:
+                        type = LLM_TYPE_UNKNOWN;
+                }
+                // Set non-causal attention for diffusion models
+                hparams.causal_attn = false;
+            }
+            break;
         case LLM_ARCH_QWEN2MOE:
             {
                 ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp, false);
@@ -883,6 +912,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
             } break;
         case LLM_ARCH_QWEN3:
             {
+                ml.get_key(LLM_KV_POOLING_TYPE, hparams.pooling_type, false);
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 switch (hparams.n_layer) {
                     case 28: type = hparams.n_embd == 1024 ? LLM_TYPE_0_6B : LLM_TYPE_1_7B; break;
@@ -1065,6 +1095,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
                 switch (hparams.n_layer) {
+                    case 18: type = LLM_TYPE_537M; break;
                     case 26: type = LLM_TYPE_1B; break;
                     case 34: type = LLM_TYPE_4B; break;
                     case 48: type = LLM_TYPE_12B; break;
@@ -1417,6 +1448,34 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_GLM4_MOE:
+            {
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+                // MoE parameters
+                ml.get_key(LLM_KV_EXPERT_COUNT,                hparams.n_expert);
+                ml.get_key(LLM_KV_EXPERT_USED_COUNT,           hparams.n_expert_used);
+                ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
+                ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm, false);
+
+                // Expert gating function (GLM-4.5 uses sigmoid)
+                ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);
+                if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {
+                    hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;
+                }
+
+                // NextN/MTP parameters
+                ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS,        hparams.nextn_predict_layers, false);
+
+                switch (hparams.n_layer) {
+                    case 47: type = LLM_TYPE_106B_A12B; break; // GLM-4.5-Air (46 layers + 1 NextN layer)
+                    case 93: type = LLM_TYPE_355B_A32B; break; // GLM-4.5 (92 layers + 1 NextN layer)
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_BITNET:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -1744,6 +1803,18 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_HUNYUAN_DENSE:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+                switch (hparams.n_embd) {
+                    case 1024: type = LLM_TYPE_0_5B; break;
+                    case 2048: type = LLM_TYPE_1_8B; break;
+                    case 3072: type = LLM_TYPE_4B; break;
+                    case 4096: type = LLM_TYPE_7B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_SMOLLM3:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -1754,6 +1825,17 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_OPENAI_MOE:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+                ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa);
+
+                hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+                hparams.set_swa_pattern(2);
+
+                // TODO: switch (hparams.n_layer)
+            } break;
         case LLM_ARCH_LFM2:
             {
                 ml.get_key(LLM_KV_SHORTCONV_L_CACHE,           hparams.n_shortconv_l_cache);
@@ -1768,6 +1850,29 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default:   type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_SMALLTHINKER:
+            {
+                const bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
+
+                if (found_swa && hparams.n_swa > 0) {
+                    hparams.swa_type      = LLAMA_SWA_TYPE_STANDARD;
+                    hparams.n_swa         = 4096;
+                    hparams.set_swa_pattern(4, true);
+                } else {
+                    hparams.swa_type             = LLAMA_SWA_TYPE_NONE;
+                    hparams.n_no_rope_layer_step = hparams.n_layer;
+                }
+
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp, false);
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);
+
+                switch (hparams.n_layer) {
+                    case 32: type = LLM_TYPE_4B;  break;
+                    case 52: type = LLM_TYPE_20B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         default: throw std::runtime_error("unsupported model architecture");
     }
 
@@ -1801,7 +1906,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
     LLAMA_LOG_INFO("%s: loading model tensors, this can take a while... (mmap = %s)\n", __func__, ml.use_mmap ? "true" : "false");
 
     // build a list of buffer types for the CPU and GPU devices
-    pimpl->cpu_buft_list = make_cpu_buft_list(devices);
+    pimpl->cpu_buft_list = make_cpu_buft_list(devices, params.use_extra_bufts);
     for (auto * dev : devices) {
         buft_list_t buft_list = make_gpu_buft_list(dev, split_mode, tensor_split);
         // add CPU buffer types as a fallback
@@ -1897,6 +2002,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
     const auto TENSOR_DUPLICATED   = llama_model_loader::TENSOR_DUPLICATED;
     const auto TENSOR_NOT_REQUIRED = llama_model_loader::TENSOR_NOT_REQUIRED;
+    const auto TENSOR_SKIP         = llama_model_loader::TENSOR_SKIP;
 
     // create tensors for the weights
     {
@@ -1952,7 +2058,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
             }
 
             // skip unused tensors
-            if (info.op == GGML_OP_NONE) {
+            if (info.op == GGML_OP_NONE || flags & TENSOR_SKIP) {
                 const size_t nbytes = ggml_nbytes(t_meta);
                 LLAMA_LOG_WARN("model has unused tensor %s (size = %zu bytes) -- ignoring\n", tn.str().c_str(), nbytes);
 
@@ -1962,11 +2068,15 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 return nullptr;
             }
 
-            // tensors with "bias" suffix are always used with GGML_OP_ADD
+            // tensors with "bias" suffix are always used with GGML_OP_ADD or GGML_OP_ADD_ID
             ggml_op op;
             bool bias = tn.suffix != nullptr && strcmp(tn.suffix, "bias") == 0;
             if (bias) {
-                op = GGML_OP_ADD;
+                if (info.op == GGML_OP_MUL_MAT_ID) {
+                    op = GGML_OP_ADD_ID;
+                } else {
+                    op = GGML_OP_ADD;
+                }
             } else {
                 op = info.op;
             }
@@ -2006,7 +2116,13 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 for (const auto * overrides = ml.tensor_buft_overrides; overrides->pattern != nullptr; ++overrides) {
                     std::regex pattern(overrides->pattern);
                     if (std::regex_search(tensor_name, pattern)) {
-                        buft = overrides->buft;
+                        if (overrides->buft == ggml_backend_cpu_buffer_type()) {
+                            // when overriding to a CPU buffer, consider the extra buffer types
+                            buft = select_weight_buft(hparams, t_meta, op, pimpl->cpu_buft_list);
+                        } else {
+                            buft = overrides->buft;
+                        }
+
                         LLAMA_LOG_DEBUG("tensor %s (%zu MiB %s) buffer type overridden to %s\n",
                                 tensor_name.c_str(),
                                 ggml_nbytes(t_meta) / 1024 / 1024, ggml_type_name(t_meta->type),
@@ -2126,6 +2242,53 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         }
                     }
                 } break;
+            case LLM_ARCH_LLADA:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);
+                    output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED);
+
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output =
+                            create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED);
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);
+
+                        // Use separate Q, K, V projections without bias, matching LLaDALlamaBlock
+                        layer.wq =
+                            create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head }, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_k_gqa }, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, n_embd_v_gqa }, 0);
+                        // No bias for QKV projections as per config: include_bias=false, include_qkv_bias=false
+                        layer.wo =
+                            create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0);
+                        layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), { n_embd }, TENSOR_NOT_REQUIRED);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd }, 0);
+
+                        layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), { n_rot / 2 },
+                                                         TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+
+                        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), { n_embd, n_ff }, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd }, 0);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), { n_embd, n_ff }, 0);
+
+                        // optional MLP bias
+                        layer.ffn_gate_b =
+                            create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), { n_ff }, TENSOR_NOT_REQUIRED);
+                        layer.ffn_down_b =
+                            create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), { n_embd }, TENSOR_NOT_REQUIRED);
+                        layer.ffn_up_b = create_tensor(tn(LLM_TENSOR_FFN_UP, "bias", i), { n_ff }, TENSOR_NOT_REQUIRED);
+                    }
+                }
+                break;
             case LLM_ARCH_LLAMA4:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -4322,6 +4485,105 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ffn_post_norm  = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
                     }
                 } break;
+            case LLM_ARCH_GLM4_MOE:
+                {
+                    const int64_t n_expert        = hparams.n_expert;
+                    const int64_t n_expert_used   = hparams.n_expert_used;
+                    const int64_t n_expert_shared = hparams.n_expert_shared;
+
+                    GGML_ASSERT(hparams.n_expert > 0 && "n_expert must be > 0 for GLM4_MOE MoE layers");
+                    GGML_ASSERT(hparams.n_expert_used > 0 && "n_expert_used must be > 0 for GLM4_MOE MoE layers");
+
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);
+                    output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED);
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED);
+                    }
+
+                    // Load ALL tensors including NextN layer to satisfy total tensor count
+                    // but only PROCESS up to last layer (skipping final NextN layer) in forward pass
+                    for (int i = 0; i < n_layer; ++i) {
+                        int flags = 0;
+                        if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
+                            // skip all tensors in the NextN layers
+                            flags |= TENSOR_SKIP;
+                        }
+
+                        auto & layer = layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, flags);
+
+                        // GLM-style attention with bias terms
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head }, flags);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_k_gqa }, flags);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, n_embd_v_gqa }, flags);
+                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), { n_embd_head_k * n_head }, flags);
+                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), { n_embd_k_gqa }, flags);
+                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), { n_embd_v_gqa }, flags);
+
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, flags);
+
+                        // K/Q norm tensors (optional for GLM-4.5 355B variant)
+                        layer.attn_q_norm = create_tensor(
+                            tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), { n_embd_head_k }, TENSOR_NOT_REQUIRED | flags);
+                        layer.attn_k_norm = create_tensor(
+                            tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), { n_embd_head_k }, TENSOR_NOT_REQUIRED | flags);
+
+                        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), { n_embd }, flags);
+
+                        // Check if this layer uses MoE or dense FFN based on n_layer_dense_lead
+                        // GLM 4.5 uses hybrid architecture: layer 0 is dense, layers 1+ are MoE
+                        const bool use_moe = (static_cast<uint32_t>(i) >= hparams.n_layer_dense_lead);
+
+                        if (use_moe) {
+                            // MoE layers
+                            layer.ffn_gate_inp =
+                                create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), { n_embd, n_expert }, flags);
+                            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), { n_expert }, flags);
+
+                            // MoE branch
+                            const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+
+                            layer.ffn_gate_exps = create_tensor(
+                                tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, flags);
+                            layer.ffn_down_exps = create_tensor(
+                                tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff_exp, n_embd, n_expert }, flags);
+                            layer.ffn_up_exps = create_tensor(
+                                tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, flags);
+
+                            // Shared expert
+                            if (n_expert_shared > 0) {
+                                const int64_t n_ff_shexp = n_ff_exp * n_expert_shared;
+                                layer.ffn_gate_shexp = create_tensor(
+                                    tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), { n_embd, n_ff_shexp }, flags);
+                                layer.ffn_down_shexp = create_tensor(
+                                    tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_shexp, n_embd }, flags);
+                                layer.ffn_up_shexp = create_tensor(
+                                    tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), { n_embd, n_ff_shexp }, flags);
+                            }
+                        } else {
+                            // Dense layers (first k layers) - GLM uses separate gate/up projections
+                            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), { n_embd, n_ff }, flags);
+                            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd }, flags);
+                            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), { n_embd, n_ff }, flags);
+                        }
+
+                        // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers
+                        if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
+                            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags);
+                            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, flags);
+                            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, flags);
+                            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, flags);
+                            layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, flags);
+                            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags);
+                        }
+                    }
+                }
+                break;
             case LLM_ARCH_NEMOTRON:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -5103,6 +5365,39 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);
                     }
                 } break;
+            case LLM_ARCH_HUNYUAN_DENSE:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+                        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
+                        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+
+                    }
+                } break;
             case LLM_ARCH_SMOLLM3:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -5132,6 +5427,46 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
                     }
                 } break;
+            case LLM_ARCH_OPENAI_MOE:
+                {
+                    const int64_t n_ff_exp = hparams.n_ff_exp;
+
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm      = create_tensor(tn(LLM_TENSOR_ATTN_NORM,      "weight", i), {n_embd}, 0);
+                        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_head * n_rot}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_head_kv * n_rot}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_head_kv * n_rot}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_rot, n_embd}, 0);
+
+                        layer.attn_sinks = create_tensor(tn(LLM_TENSOR_ATTN_SINKS, "weight", i), {n_head}, 0);
+
+                        layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {  n_embd, n_expert}, 0);
+                        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+                        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
+                        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+
+                        // bias
+                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_head * n_rot}, 0);
+                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_head_kv * n_rot}, 0);
+                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_head_kv * n_rot}, 0);
+                        layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
+
+                        layer.ffn_gate_inp_b  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "bias", i), {n_expert}, 0);
+                        layer.ffn_gate_exps_b = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "bias", i), {n_ff_exp, n_expert}, 0);
+                        layer.ffn_down_exps_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "bias", i), {  n_embd, n_expert}, 0);
+                        layer.ffn_up_exps_b   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "bias", i), {n_ff_exp, n_expert}, 0);
+                    }
+                } break;
             case LLM_ARCH_LFM2:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -5165,6 +5500,42 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         }
                     }
                 } break;
+            case LLM_ARCH_SMALLTHINKER:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);
+
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head }, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_gqa }, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, n_embd_gqa }, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd }, 0);
+
+                        GGML_ASSERT(n_expert > 0 && "n_expert must be > 0 for SMALLTHINKER");
+                        GGML_ASSERT(n_expert_used > 0 && "n_expert_used must be > 0 for SMALLTHINKER");
+
+                        // MoE branch
+                        const int64_t n_ff_exp = hparams.n_ff_exp;
+                        layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), { n_embd, n_expert }, 0);
+                        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, 0);
+                        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff_exp, n_embd, n_expert }, 0);
+                        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, 0);
+                    }
+                } break;
             default:
                 throw std::runtime_error("unknown architecture");
         }
@@ -5468,7 +5839,7 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: n_ff_shexp       = %d\n",     __func__, hparams.n_ff_shexp);
     }
 
-    if (arch == LLM_ARCH_QWEN3MOE) {
+    if (arch == LLM_ARCH_QWEN3MOE || arch == LLM_ARCH_OPENAI_MOE) {
         LLAMA_LOG_INFO("%s: n_ff_exp         = %d\n",     __func__, hparams.n_ff_exp);
     }
 
@@ -5490,6 +5861,11 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: expert_weights_norm  = %d\n",     __func__, hparams.expert_weights_norm);
     }
 
+    if (arch == LLM_ARCH_SMALLTHINKER) {
+        LLAMA_LOG_INFO("%s: n_ff_exp             = %d\n",     __func__, hparams.n_ff_exp);
+        LLAMA_LOG_INFO("%s: expert_gating_func   = %s\n",     __func__, llama_expert_gating_func_name((llama_expert_gating_func_type) hparams.expert_gating_func));
+    }
+
     vocab.print_info();
 }
 
@@ -7978,8 +8354,10 @@ struct llm_build_dream : public llm_graph_context {
     }
 };
 
-struct llm_build_qwen2vl : public llm_graph_context {
-    llm_build_qwen2vl(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+struct llm_build_llada : public llm_graph_context {
+    llm_build_llada(const llama_model & model, const llm_graph_params & params) :
+        llm_graph_context(params) {
+        // LLaDA is similar to LLaMA but uses non-causal attention for diffusion
         const int64_t n_embd_head = hparams.n_embd_head_v;
 
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -7993,10 +8371,8 @@ struct llm_build_qwen2vl : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
-
-        int sections[4];
-        std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
+        // Non-causal attention for diffusion
+        auto * inp_attn = build_attn_inp_no_cache();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -8004,34 +8380,134 @@ struct llm_build_qwen2vl : public llm_graph_context {
             ggml_tensor * inpSA = inpL;
 
             // norm
-            cur = build_norm(inpL,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, il);
+            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
             cb(cur, "attn_norm", il);
 
             // self-attention
             {
-                // compute Q and K and RoPE them
+                // compute separate Q, K, V projections without bias, matching LLaDALlamaBlock
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                cb(Qcur, "Qcur", il);
-
                 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                cb(Kcur, "Kcur", il);
-
                 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
                 cb(Vcur, "Vcur", il);
 
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
                 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
                 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 
-                Qcur = ggml_rope_multi(
-                        ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
+                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                     ext_factor, attn_factor, beta_fast, beta_slow);
+
+                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                     ext_factor, attn_factor, beta_fast, beta_slow);
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                cur = build_attn(inp_attn, model.layers[il].wo, NULL, Qcur, Kcur, Vcur, nullptr, nullptr,
+                                 1.0f / sqrtf(float(n_embd_head)), il);
+            }
+
+            if (il == n_layer - 1 && inp_out_ids) {
+                cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur, model.layers[il].ffn_up, NULL, NULL, model.layers[il].ffn_gate, NULL, NULL,
+                            model.layers[il].ffn_down, NULL, NULL, NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+};
+
+struct llm_build_qwen2vl : public llm_graph_context {
+    llm_build_qwen2vl(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        auto * inp_attn = build_attn_inp_kv_unified();
+
+        int sections[4];
+        std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
+
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = ggml_rope_multi(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
                         );
 
                 Kcur = ggml_rope_multi(
@@ -13285,6 +13761,165 @@ struct llm_build_glm4 : public llm_graph_context {
     }
 };
 
+struct llm_build_glm4_moe : public llm_graph_context {
+    llm_build_glm4_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        auto * inp_attn = build_attn_inp_kv_unified();
+
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+        // Only process up to last layer (skip final NextN layer)
+        // Final layer tensors are loaded but not processed in forward pass
+        const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
+        for (int il = 0; il < n_transformer_layers; ++il) {
+            ggml_tensor * inpSA = inpL;
+
+            // Pre-attention norm
+            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                }
+                cb(Qcur, "Qcur", il);
+
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                }
+                cb(Kcur, "Kcur", il);
+
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                }
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                // Apply Q/K norm if available (GLM-4.5 355B variant)
+                if (model.layers[il].attn_q_norm) {
+                    Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+                    cb(Qcur, "Qcur_normed", il);
+                }
+                if (model.layers[il].attn_k_norm) {
+                    Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+                    cb(Kcur, "Kcur_normed", il);
+                }
+
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, NULL,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+            }
+
+            if (il == n_transformer_layers - 1 && inp_out_ids) {
+                cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // Post-attention norm
+            cur = build_norm(ffn_inp, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "post_attn_norm", il);
+
+            // Check if this is a dense layer (n_layer_dense_lead=1, so layer 0 is dense)
+            if (static_cast<uint32_t>(il) < hparams.n_layer_dense_lead) {
+                // Dense FFN layer
+                cur = build_ffn(cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(cur, "ffn_out", il);
+            } else {
+                // Process routed experts using existing MoE infrastructure
+                ggml_tensor * routed_out = build_moe_ffn(cur,
+                        model.layers[il].ffn_gate_inp,
+                        model.layers[il].ffn_up_exps,
+                        model.layers[il].ffn_gate_exps,
+                        model.layers[il].ffn_down_exps,
+                        model.layers[il].ffn_exp_probs_b,
+                        n_expert, n_expert_used,
+                        LLM_FFN_SILU, hparams.expert_weights_norm,
+                        true, hparams.expert_weights_scale,
+                        (llama_expert_gating_func_type) hparams.expert_gating_func,
+                        il);
+                cb(routed_out, "ffn_moe_out", il);
+
+                // Process shared expert on original input
+                ggml_tensor * shared_out = build_ffn(cur,
+                        model.layers[il].ffn_up_shexp,   NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(shared_out, "ffn_shexp_out", il);
+
+                // Final output: routed_output + shared_output
+                cur = ggml_add(ctx0, routed_out, shared_out);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+        cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+};
+
 struct llm_build_nemotron : public llm_graph_context {
     llm_build_nemotron(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
         const int64_t n_embd_head = hparams.n_embd_head_v;
@@ -16697,8 +17332,8 @@ struct llm_build_hunyuan_moe : public llm_graph_context {
     }
 };
 
-struct llm_build_smollm3 : public llm_graph_context {
-    llm_build_smollm3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+struct llm_build_hunyuan_dense : public llm_graph_context {
+    llm_build_hunyuan_dense(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
         const int64_t n_embd_head = hparams.n_embd_head_v;
 
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -16714,23 +17349,23 @@ struct llm_build_smollm3 : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
-        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+        const float kq_scale = 1.0f / sqrtf(float(n_embd_head));
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
-            const bool use_rope = (il + 1) % hparams.n_no_rope_layer_step != 0;
-
             // norm
             cur = build_norm(inpL,
                     model.layers[il].attn_norm, NULL,
                     LLM_NORM_RMS, il);
             cb(cur, "attn_norm", il);
-
             // self-attention
             {
+                // rope freq factors for llama3; may return nullptr for llama2 and other models
+                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
                 // compute Q and K and RoPE them
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
                 cb(Qcur, "Qcur", il);
@@ -16757,10 +17392,148 @@ struct llm_build_smollm3 : public llm_graph_context {
                 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
                 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 
-                if (use_rope) {
-                    Qcur = ggml_rope_ext(
-                            ctx0, Qcur, inp_pos, nullptr,
-                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                Qcur = ggml_rope_ext(
+                         ctx0, Qcur, inp_pos, rope_factors,
+                         n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                         ext_factor, attn_factor, beta_fast, beta_slow
+                         );
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                Kcur = ggml_rope_ext(
+                         ctx0, Kcur, inp_pos, rope_factors,
+                         n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                         ext_factor, attn_factor, beta_fast, beta_slow
+                         );
+
+                Kcur = build_norm(Kcur,
+                         model.layers[il].attn_k_norm, nullptr,
+                         LLM_NORM_RMS, il);
+                cb(Kcur, "Kcur_norm", il);
+
+                Qcur = build_norm(Qcur,
+                         model.layers[il].attn_q_norm, nullptr,
+                         LLM_NORM_RMS, il);
+                cb(Qcur, "Qcur_norm", il);
+
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                cb(cur, "attn_out", il);
+            }
+
+            if (il == n_layer - 1 && inp_out_ids) {
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+            // feed-forward network (non-MoE)
+            ggml_tensor * cur_mlp = build_ffn(cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur_mlp, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur_mlp, ffn_inp);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+};
+
+struct llm_build_smollm3 : public llm_graph_context {
+    llm_build_smollm3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        auto * inp_attn = build_attn_inp_kv_unified();
+
+        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * inpSA = inpL;
+
+            const bool use_rope = (il + 1) % hparams.n_no_rope_layer_step != 0;
+
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                if (use_rope) {
+                    Qcur = ggml_rope_ext(
+                            ctx0, Qcur, inp_pos, nullptr,
+                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                             ext_factor, attn_factor, beta_fast, beta_slow
                             );
 
@@ -16834,6 +17607,136 @@ struct llm_build_smollm3 : public llm_graph_context {
     }
 };
 
+struct llm_build_openai_moe_iswa : public llm_graph_context {
+    llm_build_openai_moe_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        auto * inp_attn = build_attn_inp_kv_unified_iswa();
+
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, nullptr,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_rot, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_rot, n_head_kv, n_tokens);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                cur = build_attn_with_sinks(inp_attn,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].attn_sinks, 1.0f/sqrtf(float(n_rot)), il);
+
+                cb(cur, "attn_out", il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            cur = ffn_inp;
+            cur = build_norm(cur,
+                    model.layers[il].attn_post_norm, nullptr,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_post_norm", il);
+
+            // MoE branch
+            cur = build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,  model.layers[il].ffn_gate_inp_b,
+                    model.layers[il].ffn_up_exps,   model.layers[il].ffn_up_exps_b,
+                    model.layers[il].ffn_gate_exps, model.layers[il].ffn_gate_exps_b,
+                    model.layers[il].ffn_down_exps, model.layers[il].ffn_down_exps_b,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SWIGLU_OAI_MOE, false,
+                    false, 0.0,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX_WEIGHT,
+                    il);
+            cb(cur, "ffn_moe_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+};
+
 struct llm_build_lfm2 : public llm_graph_context {
     const llama_model & model;
 
@@ -17011,6 +17914,127 @@ struct llm_build_lfm2 : public llm_graph_context {
     }
 };
 
+template <bool iswa>
+struct llm_build_smallthinker : public llm_graph_context{
+    llm_build_smallthinker(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params){
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_unified_iswa, llm_graph_input_attn_kv_unified>;
+        inp_attn_type * inp_attn = nullptr;
+
+        if constexpr (iswa) {
+            inp_attn = build_attn_inp_kv_unified_iswa();
+        } else {
+            inp_attn = build_attn_inp_kv_unified();
+        }
+
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * inpSA  = inpL;
+            ggml_tensor * probs  = nullptr;
+
+            probs = build_lora_mm(model.layers[il].ffn_gate_inp, inpL);  // [n_expert, n_tokens]
+            cb(probs, "ffn_moe_logits", il);
+
+            // norm
+            cur = build_norm(inpL,model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self_attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                if (hparams.n_no_rope_layer_step == n_layer || il % hparams.n_no_rope_layer_step != 0) {
+                    Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                     ext_factor, attn_factor, beta_fast, beta_slow);
+
+                    Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                     ext_factor, attn_factor, beta_fast, beta_slow);
+                }
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+            }
+
+            if (il == n_layer - 1 && inp_out_ids) {
+                cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+                probs = ggml_get_rows(ctx0, probs, inp_out_ids);
+            }
+
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // MoE branch
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            ggml_tensor * ffn_out =
+                build_moe_ffn(cur,
+                        nullptr,
+                        model.layers[il].ffn_up_exps,
+                        model.layers[il].ffn_gate_exps,
+                        model.layers[il].ffn_down_exps,
+                        nullptr,
+                        n_expert, n_expert_used,
+                        LLM_FFN_RELU, true,
+                        false, 0.0,
+                        static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func),
+                        il, probs);
+
+            cb(ffn_out, "ffn_out", il);
+            cur = ffn_out;
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+};
+
 llama_memory_i * llama_model::create_memory(const llama_memory_params & params, llama_cparams & cparams) const {
     llama_memory_i * res;
 
@@ -17024,6 +18048,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
         case LLM_ARCH_NEO_BERT:
         case LLM_ARCH_WAVTOKENIZER_DEC:
         case LLM_ARCH_DREAM:
+        case LLM_ARCH_LLADA:
             {
                 res = nullptr;
             } break;
@@ -17059,6 +18084,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                         /* recurrent_kv_size */ std::max((uint32_t) 1, cparams.n_seq_max),
                         /* n_seq_max         */ cparams.n_seq_max,
                         /* offload           */ cparams.offload_kqv,
+                        /* unified           */ cparams.kv_unified,
                         /* filter_attn       */ (arch == LLM_ARCH_FALCON_H1) ? [&](int32_t) { return true; } : (llama_memory_hybrid::layer_filter_cb)nullptr,
                         /* filter_recr       */ (arch == LLM_ARCH_FALCON_H1) ? [&](int32_t) { return true; } : (llama_memory_hybrid::layer_filter_cb)nullptr);
                 } else {
@@ -17190,6 +18216,11 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
                 llm = std::make_unique<llm_build_dream>(*this, params);
             }
             break;
+        case LLM_ARCH_LLADA:
+            {
+                llm = std::make_unique<llm_build_llada>(*this, params);
+            }
+            break;
         case LLM_ARCH_QWEN2VL:
             {
                 llm = std::make_unique<llm_build_qwen2vl>(*this, params);
@@ -17332,6 +18363,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_glm4>(*this, params);
             } break;
+        case LLM_ARCH_GLM4_MOE:
+            {
+                llm = std::make_unique<llm_build_glm4_moe>(*this, params);
+            } break;
         case LLM_ARCH_BITNET:
             {
                 llm = std::make_unique<llm_build_bitnet>(*this, params);
@@ -17437,10 +18472,18 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_hunyuan_moe>(*this, params);
             } break;
+        case LLM_ARCH_HUNYUAN_DENSE:
+            {
+                llm = std::make_unique<llm_build_hunyuan_dense>(*this, params);
+            } break;
         case LLM_ARCH_SMOLLM3:
             {
                 llm = std::make_unique<llm_build_smollm3>(*this, params);
             } break;
+        case LLM_ARCH_OPENAI_MOE:
+            {
+                llm = std::make_unique<llm_build_openai_moe_iswa>(*this, params);
+            } break;
         case LLM_ARCH_FALCON_H1:
             {
                 llm = std::make_unique<llm_build_falcon_h1>(*this, params);
@@ -17449,6 +18492,14 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_lfm2>(*this, params);
             } break;
+        case LLM_ARCH_SMALLTHINKER:
+            {
+                if (hparams.swa_type == LLAMA_SWA_TYPE_STANDARD) {
+                    llm = std::make_unique<llm_build_smallthinker<true>> (*this, params);
+                } else {
+                    llm = std::make_unique<llm_build_smallthinker<false>>(*this, params);
+                }
+            } break;
         default:
             GGML_ABORT("fatal error");
     }
@@ -17478,6 +18529,7 @@ llama_model_params llama_model_default_params() {
         /*.use_mmap                    =*/ true,
         /*.use_mlock                   =*/ false,
         /*.check_tensors               =*/ false,
+        /*.use_extra_bufts             =*/ true,
     };
 
 #ifdef GGML_USE_METAL
@@ -17580,6 +18632,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
 
         // use what we call a normal RoPE, operating on pairs of consecutive head values
         case LLM_ARCH_LLAMA:
+        case LLM_ARCH_LLADA:
         case LLM_ARCH_LLAMA4:
         case LLM_ARCH_DECI:
         case LLM_ARCH_BAICHUAN:
@@ -17646,7 +18699,11 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_MINICPM3:
         case LLM_ARCH_DOTS1:
         case LLM_ARCH_HUNYUAN_MOE:
+        case LLM_ARCH_OPENAI_MOE:
+        case LLM_ARCH_HUNYUAN_DENSE:
         case LLM_ARCH_LFM2:
+        case LLM_ARCH_SMALLTHINKER:
+        case LLM_ARCH_GLM4_MOE:
             return LLAMA_ROPE_TYPE_NEOX;
 
         case LLM_ARCH_QWEN2VL:
@@ -17757,6 +18814,10 @@ bool llama_model_is_recurrent(const llama_model * model) {
     return llm_arch_is_recurrent(model->arch);
 }
 
+bool llama_model_is_diffusion(const llama_model * model) {
+    return llm_arch_is_diffusion(model->arch);
+}
+
 const std::vector<std::pair<std::string, ggml_tensor *>> & llama_internal_get_tensor_map(const llama_model * model) {
     return model->tensors_by_name;
 }
diff --git a/examples/talk-llama/llama-model.h b/examples/talk-llama/llama-model.h
index 094e23808a8..46f7d0480fa 100644
--- a/examples/talk-llama/llama-model.h
+++ b/examples/talk-llama/llama-model.h
@@ -39,6 +39,7 @@ enum llm_type {
     LLM_TYPE_410M,
     LLM_TYPE_450M,
     LLM_TYPE_475M,
+    LLM_TYPE_537M,
     LLM_TYPE_700M,
     LLM_TYPE_770M,
     LLM_TYPE_780M,
@@ -101,8 +102,10 @@ enum llm_type {
     LLM_TYPE_A13B,
     LLM_TYPE_21B_A3B, // Ernie MoE small
     LLM_TYPE_30B_A3B,
+    LLM_TYPE_106B_A12B, // GLM-4.5-Air
     LLM_TYPE_235B_A22B,
     LLM_TYPE_300B_A47B, // Ernie MoE big
+    LLM_TYPE_355B_A32B, // GLM-4.5
     LLM_TYPE_E2B,
     LLM_TYPE_E4B,
 };
@@ -166,6 +169,15 @@ struct llama_layer_shortconv {
     struct ggml_tensor * out_proj = nullptr;
 };
 
+struct llama_layer_nextn {
+    struct ggml_tensor * eh_proj          = nullptr;
+    struct ggml_tensor * embed_tokens     = nullptr;
+    struct ggml_tensor * enorm            = nullptr;
+    struct ggml_tensor * hnorm            = nullptr;
+    struct ggml_tensor * shared_head_head = nullptr;
+    struct ggml_tensor * shared_head_norm = nullptr;
+};
+
 struct llama_layer {
     // normalization
     struct ggml_tensor * attn_norm       = nullptr;
@@ -241,10 +253,14 @@ struct llama_layer {
     struct ggml_tensor * ffn_up_enc   = nullptr;
 
     // ff MoE
-    struct ggml_tensor * ffn_gate_inp  = nullptr;
-    struct ggml_tensor * ffn_gate_exps = nullptr;
-    struct ggml_tensor * ffn_down_exps = nullptr;
-    struct ggml_tensor * ffn_up_exps   = nullptr;
+    struct ggml_tensor * ffn_gate_inp    = nullptr;
+    struct ggml_tensor * ffn_gate_exps   = nullptr;
+    struct ggml_tensor * ffn_down_exps   = nullptr;
+    struct ggml_tensor * ffn_up_exps     = nullptr;
+    struct ggml_tensor * ffn_gate_inp_b  = nullptr;
+    struct ggml_tensor * ffn_gate_exps_b = nullptr;
+    struct ggml_tensor * ffn_down_exps_b = nullptr;
+    struct ggml_tensor * ffn_up_exps_b   = nullptr;
 
     // ff shared expert (shexp)
     struct ggml_tensor * ffn_gate_inp_shexp = nullptr;
@@ -349,11 +365,16 @@ struct llama_layer {
     struct ggml_tensor * laurel_r             = nullptr;
     struct ggml_tensor * laurel_post_norm     = nullptr;
 
+    // openai-moe
+    struct ggml_tensor * attn_sinks = nullptr;
+
     struct llama_layer_posnet posnet;
 
     struct llama_layer_convnext convnext;
 
     struct llama_layer_shortconv shortconv;
+
+    struct llama_layer_nextn nextn;
 };
 
 struct llama_model {
diff --git a/examples/talk-llama/llama-quant.cpp b/examples/talk-llama/llama-quant.cpp
index a00af7a1d17..1d0361cc166 100644
--- a/examples/talk-llama/llama-quant.cpp
+++ b/examples/talk-llama/llama-quant.cpp
@@ -211,7 +211,10 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
             const int64_t nx = tensor->ne[0];
             const int64_t qk_k = ggml_blck_size(new_type);
 
-            if (arch == LLM_ARCH_FALCON || nx % qk_k != 0) {
+            if (ftype == LLAMA_FTYPE_MOSTLY_MXFP4_MOE) {
+                new_type = GGML_TYPE_Q8_0;
+            }
+            else if (arch == LLM_ARCH_FALCON || nx % qk_k != 0) {
                 new_type = GGML_TYPE_Q8_0;
             }
             else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
@@ -223,6 +226,14 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
                 new_type = GGML_TYPE_Q6_K;
             }
         }
+    } else if (ftype == LLAMA_FTYPE_MOSTLY_MXFP4_MOE) {
+        // MoE   tensors -> MXFP4
+        // other tensors -> Q8_0
+        if (tensor->ne[2] > 1) {
+            new_type = GGML_TYPE_MXFP4;
+        } else {
+            new_type = GGML_TYPE_Q8_0;
+        }
     } else if (name == "token_embd.weight" || name == "per_layer_token_embd.weight") {
         if (qs.params->token_embedding_type < GGML_TYPE_COUNT) {
             new_type = qs.params->token_embedding_type;
@@ -533,6 +544,8 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         case LLAMA_FTYPE_MOSTLY_BF16: default_type = GGML_TYPE_BF16; break;
         case LLAMA_FTYPE_ALL_F32:     default_type = GGML_TYPE_F32;  break;
 
+        case LLAMA_FTYPE_MOSTLY_MXFP4_MOE: default_type = GGML_TYPE_MXFP4; break;
+
         // K-quants
         case LLAMA_FTYPE_MOSTLY_Q2_K_S:
         case LLAMA_FTYPE_MOSTLY_Q2_K:    default_type = GGML_TYPE_Q2_K;    break;
@@ -875,9 +888,10 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
 
             // get more optimal quantization type based on the tensor shape, layer, etc.
             if (!params->pure && ggml_is_quantized(default_type)) {
+                int fallback = qs.n_fallback;
                 new_type = llama_tensor_get_type(qs, new_type, tensor, ftype);
-                // unless the user specifies a type
-                if (params->tensor_types) {
+                // unless the user specifies a type, and the tensor geometry will not require fallback quantisation
+                if (params->tensor_types && qs.n_fallback - fallback == 0) {
                     const std::vector<tensor_quantization> & tensor_types = *static_cast<const std::vector<tensor_quantization> *>(params->tensor_types);
                     const std::string tensor_name(tensor->name);
                     for (const auto & [tname, qtype] : tensor_types) {
@@ -890,7 +904,6 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
                     }
                 }
             }
-
             if (params->token_embedding_type < GGML_TYPE_COUNT && strcmp(tensor->name, "token_embd.weight") == 0) {
                 new_type = params->token_embedding_type;
             }
@@ -984,6 +997,29 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
                 const float * imatrix_03 = imatrix ? imatrix + i03 * n_per_row : nullptr;
 
                 new_size += llama_tensor_quantize_impl(new_type, f32_data_03, new_data_03, chunk_size, nrows, n_per_row, imatrix_03, workers, nthread_use);
+
+                // TODO: temporary sanity check that the F16 -> MXFP4 is lossless
+#if 0
+                if (new_type == GGML_TYPE_MXFP4) {
+                    auto * x = f32_data_03;
+
+                    //LLAMA_LOG_INFO("nrows = %d, n_per_row = %d\n", nrows, n_per_row);
+                    std::vector<float> deq(nrows*n_per_row);
+                    const ggml_type_traits * qtype = ggml_get_type_traits(new_type);
+                    qtype->to_float(new_data_03, deq.data(), deq.size());
+
+                    double err = 0.0f;
+                    for (int i = 0; i < (int) deq.size(); ++i) {
+                        err += fabsf(deq[i] - x[i]);
+                        //if (fabsf(deq[i] - x[i]) > 0.00001 && i < 256) {
+                        if (deq[i] != x[i]) {
+                            LLAMA_LOG_INFO("deq[%d] = %f, x[%d] = %f\n", i, deq[i], i, x[i]);
+                        }
+                    }
+                    //LLAMA_LOG_INFO("err = %f\n", err);
+                    GGML_ASSERT(err == 0.00000);
+                }
+#endif
             }
             LLAMA_LOG_INFO("size = %8.2f MiB -> %8.2f MiB\n", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0);
         }
diff --git a/examples/talk-llama/llama-vocab.cpp b/examples/talk-llama/llama-vocab.cpp
index e8bae645088..de5d1681dff 100644
--- a/examples/talk-llama/llama-vocab.cpp
+++ b/examples/talk-llama/llama-vocab.cpp
@@ -307,6 +307,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                 };
                 break;
             case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM:
+            case LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE:
                 regex_exprs = {
                     "\\p{N}{1,3}",
                     "[一-龥぀-ゟ゠-ヿ]+",
@@ -1855,7 +1856,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     tokenizer_pre == "gigachat"   ||
                     tokenizer_pre == "jina-v2-es" ||
                     tokenizer_pre == "jina-v2-de" ||
-                    tokenizer_pre == "a.x-4.0") {
+                    tokenizer_pre == "a.x-4.0" ||
+                    tokenizer_pre == "mellum") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_GPT2;
             } else if (
                     tokenizer_pre == "jina-v1-en" ||
@@ -1964,6 +1966,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 tokenizer_pre == "hunyuan") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_HUNYUAN;
                 clean_spaces = false;
+            } else if (
+                tokenizer_pre == "hunyuan-dense") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE;
+                clean_spaces = false;
             } else if (
                 tokenizer_pre == "kimi-k2") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_KIMI_K2;
@@ -2185,6 +2191,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<｜fim▁begin｜>" // DeepSeek
                         || t.first == "<PRE>"
                         || t.first == "▁<PRE>"          // CodeLlama
+                        || t.first == "<|code_prefix|>" // GLM-4.5
                         ) {
                     special_fim_pre_id = t.second;
                     if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -2204,6 +2211,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<｜fim▁hole｜>" // DeepSeek
                         || t.first == "<SUF>"
                         || t.first == "▁<SUF>"         // CodeLlama
+                        || t.first == "<|code_suffix|>" // GLM-4.5
                         ) {
                     special_fim_suf_id = t.second;
                     if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -2223,6 +2231,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<｜fim▁end｜>"  // DeepSeek
                         || t.first == "<MID>"
                         || t.first == "▁<MID>"         // CodeLlama
+                        || t.first == "<|code_middle|>" // GLM-4.5
                         ) {
                     special_fim_mid_id = t.second;
                     if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -2305,6 +2314,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     || t.first == "<|eot_id|>"
                     || t.first == "<|im_end|>"
                     || t.first == "<|end|>"
+                    || t.first == "<|return|>" // o200k_harmony
+                    || t.first == "<|call|>"   // o200k_harmony
                     || t.first == "<end_of_turn>"
                     || t.first == "<|endoftext|>"
                     || t.first == "<|eom_id|>"
@@ -2328,6 +2339,13 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             }
         }
 
+        // @ngxson : quick hack for gpt-oss, always render these tokens
+        for (const auto & t : token_to_id) {
+            if (t.first == "<|channel|>" || t.first == "<|message|>" || t.first == "<|start|>" || t.first == "<|constrain|>") {
+                id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_USER_DEFINED;
+            }
+        }
+
         // sanity checks
         if (special_eos_id != LLAMA_TOKEN_NULL && special_eog_ids.count(special_eos_id) == 0) {
             special_eog_ids.insert(special_eos_id);
@@ -2343,6 +2361,37 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             special_eog_ids.insert(special_eom_id);
             LLAMA_LOG_WARN("%s: special_eom_id is not in special_eog_ids - the tokenizer config may be incorrect\n", __func__);
         }
+
+        // TODO: workaround for o200k_harmony tokenizer: the "<|end|>" token should not be EOG
+        //       we don't have a good way to detect this, so for now, if we have "<|return|>" and "<|call|>" tokens,
+        //       we remove the "<|end|>" token from the EOG list
+        {
+            bool has_return = false;
+            bool has_call   = false;
+            bool has_end    = false;
+
+            llama_token end_id = LLAMA_TOKEN_NULL;
+
+            LLAMA_LOG_INFO("%s: printing all EOG tokens:\n", __func__);
+            for (auto tid : special_eog_ids) {
+                LLAMA_LOG_INFO("%s:   - %d ('%s')\n", __func__, tid, id_to_token[tid].text.c_str());
+
+                if (id_to_token[tid].text == "<|return|>") {
+                    has_return = true;
+                } else if (id_to_token[tid].text == "<|call|>") {
+                    has_call = true;
+                } else if (id_to_token[tid].text == "<|end|>") {
+                    has_end = true;
+                    end_id = tid;
+                }
+            }
+
+            if (has_return && has_call && has_end) {
+                special_eog_ids.erase(end_id);
+                id_to_token[end_id].attr = LLAMA_TOKEN_ATTR_USER_DEFINED;
+                LLAMA_LOG_WARN("%s: special_eog_ids contains both '<|return|>' and '<|call|>' tokens, removing '<|end|>' token from EOG list\n", __func__);
+            }
+        }
     }
 
     // build special tokens cache
diff --git a/examples/talk-llama/llama-vocab.h b/examples/talk-llama/llama-vocab.h
index 842b129e861..61b81242168 100644
--- a/examples/talk-llama/llama-vocab.h
+++ b/examples/talk-llama/llama-vocab.h
@@ -46,6 +46,7 @@ enum llama_vocab_pre_type {
     LLAMA_VOCAB_PRE_TYPE_SEED_CODER     = 35,
     LLAMA_VOCAB_PRE_TYPE_HUNYUAN        = 36,
     LLAMA_VOCAB_PRE_TYPE_KIMI_K2        = 37,
+    LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE  = 38,
 };
 
 struct LLM_KV;
diff --git a/examples/talk-llama/llama.h b/examples/talk-llama/llama.h
index 6f454a508a0..135eaf1b655 100644
--- a/examples/talk-llama/llama.h
+++ b/examples/talk-llama/llama.h
@@ -152,6 +152,7 @@ extern "C" {
         //LLAMA_FTYPE_MOSTLY_Q4_0_8_8      = 35, // removed from gguf files, use Q4_0 and runtime repack
         LLAMA_FTYPE_MOSTLY_TQ1_0         = 36, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_TQ2_0         = 37, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_MXFP4_MOE     = 38, // except 1d tensors
 
         LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
     };
@@ -284,10 +285,11 @@ extern "C" {
         const struct llama_model_kv_override * kv_overrides;
 
         // Keep the booleans together to avoid misalignment during copy-by-value.
-        bool vocab_only;    // only load the vocabulary, no weights
-        bool use_mmap;      // use mmap if possible
-        bool use_mlock;     // force system to keep model in RAM
-        bool check_tensors; // validate model tensor data
+        bool vocab_only;      // only load the vocabulary, no weights
+        bool use_mmap;        // use mmap if possible
+        bool use_mlock;       // force system to keep model in RAM
+        bool check_tensors;   // validate model tensor data
+        bool use_extra_bufts; // use extra buffer types (used for weight repacking)
     };
 
     // NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations
@@ -537,6 +539,9 @@ extern "C" {
     // Returns true if the model is recurrent (like Mamba, RWKV, etc.)
     LLAMA_API bool llama_model_is_recurrent(const struct llama_model * model);
 
+    // Returns true if the model is diffusion-based (like LLaDA, Dream, etc.)
+    LLAMA_API bool llama_model_is_diffusion(const struct llama_model * model);
+
     // Returns 0 on success
     LLAMA_API uint32_t llama_model_quantize(
             const char * fname_inp,
@@ -865,6 +870,29 @@ extern "C" {
                           size_t   n_token_capacity,
                           size_t * n_token_count_out);
 
+#define LLAMA_STATE_SEQ_FLAGS_SWA_ONLY 1
+
+    typedef uint32_t llama_state_seq_flags;
+
+    LLAMA_API size_t llama_state_seq_get_size_ext(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+           llama_state_seq_flags   flags);
+
+    LLAMA_API size_t llama_state_seq_get_data_ext(
+            struct llama_context * ctx,
+                         uint8_t * dst,
+                          size_t   size,
+                    llama_seq_id   seq_id,
+           llama_state_seq_flags   flags);
+
+    LLAMA_API size_t llama_state_seq_set_data_ext(
+            struct llama_context * ctx,
+                   const uint8_t * src,
+                          size_t   size,
+                    llama_seq_id   dest_seq_id,
+           llama_state_seq_flags   flags);
+
     //
     // Decoding
     //
@@ -1432,6 +1460,8 @@ extern "C" {
 
         ggml_opt_get_optimizer_params get_opt_pars; // callback for calculating optimizer parameters
         void * get_opt_pars_ud;                     // userdata for calculating optimizer parameters
+
+        enum ggml_opt_optimizer_type optimizer_type;
     };
 
     LLAMA_API void llama_opt_init(struct llama_context * lctx, struct llama_model * model, struct llama_opt_params lopt_params);