fix prebuild on cuda12.8 (#3857)

Author: lvhan028

builder/manywheel/Dockerfile_2014

@@ -1,25 +1,22 @@
 # syntax = docker/dockerfile:experimental
+# WARNING: CentOS 7 is out of date since 6/30/2024, we should use the following one in the future
+# FROM quay.io/pypa/manylinux_2_28_x86_64 as base
 FROM quay.io/pypa/manylinux2014_x86_64 as base
-ARG DEVTOOLSET_VERSION=9
 ARG BASE_CUDA_VERSION=11.8
 
 ENV LC_ALL en_US.UTF-8
 ENV LANG en_US.UTF-8
 ENV LANGUAGE en_US.UTF-8
 
-# gcc/g++
-RUN yum install -y \
-    devtoolset-${DEVTOOLSET_VERSION}-gcc \
-    devtoolset-${DEVTOOLSET_VERSION}-gcc-c++ \
-    devtoolset-${DEVTOOLSET_VERSION}-gcc-gfortran \
-    devtoolset-${DEVTOOLSET_VERSION}-binutils \
+RUN sed -i 's|^mirrorlist=|#mirrorlist=|g' /etc/yum.repos.d/CentOS-*.repo && \
+    sed -i 's|^#baseurl=http://mirror.centos.org|baseurl=https://vault.centos.org|g' /etc/yum.repos.d/CentOS-*.repo && \
+    yum install -y \
     wget \
     rapidjson-devel \
     glog-devel && \
     yum clean all
 
-ENV PATH=/opt/rh/devtoolset-${DEVTOOLSET_VERSION}/root/usr/bin:$PATH
-ENV LD_LIBRARY_PATH=/opt/rh/devtoolset-${DEVTOOLSET_VERSION}/root/usr/lib64:/opt/rh/devtoolset-${DEVTOOLSET_VERSION}/root/usr/lib:$LD_LIBRARY_PATH
+ENV LD_LIBRARY_PATH=/opt/rh/devtoolset-10/root/usr/lib64:/opt/rh/devtoolset-10/root/usr/lib:$LD_LIBRARY_PATH
 
 FROM base as cuda
 COPY manywheel/scripts/install_cuda.sh /tmp/install_cuda.sh

builder/manywheel/build_all_lmdeploy_builders.sh

@@ -4,6 +4,6 @@ set -eou pipefail
 
 TOPDIR=$(git rev-parse --show-toplevel)/builder
 
-for cuda_version in 11.8 12.4; do
+for cuda_version in 11.8 12.4 12.8; do
     MANY_LINUX_VERSION=2014 GPU_ARCH_VERSION="${cuda_version}" "${TOPDIR}/manywheel/build_lmdeploy_builder.sh"
 done

builder/manywheel/build_all_wheel.sh

@@ -4,7 +4,7 @@ set -eou pipefail
 
 TOPDIR=$(git rev-parse --show-toplevel)/builder
 
-CUDA_VER=${CUDA_VER:-12.4}
+CUDA_VER=${CUDA_VER:-12.8}
 
 PLAT_NAME=manylinux2014_x86_64
 for cuver in ${CUDA_VER}; do

src/turbomind/kernels/gemm/tma.cu

@@ -16,7 +16,8 @@ PFN_cuTensorMapEncodeTiled get_cuTensorMapEncodeTiled()
         cudaDriverEntryPointQueryResult driver_status;
         void*                           cuTensorMapEncodeTiled_ptr = nullptr;
 
-#if CUDA_VERSION >= 12050
+// https://github.com/NVIDIA/cutlass/pull/2086
+#if CUDA_VERSION >= 13000
         cudaGetDriverEntryPointByVersion(
             "cuTensorMapEncodeTiled", &cuTensorMapEncodeTiled_ptr, 12000, cudaEnableDefault, &driver_status);
 #else