Draft

Author: XXXXRT666

.github/build_windows_packages.ps1

@@ -115,12 +115,17 @@ Remove-Item $ffDir.FullName -Recurse -Force
 Write-Host "[INFO] Installing PyTorch..."
 & ".\runtime\python.exe" -m ensurepip
 & ".\runtime\python.exe" -m pip install --upgrade pip --no-warn-script-location
+
 switch ($cuda) {
     "cu124" {
-        & ".\runtime\python.exe" -m pip install torch==2.6 torchaudio --index-url https://download.pytorch.org/whl/cu124 --no-warn-script-location
+        & ".\runtime\python.exe" -m pip install psutil ninja packaging wheel "setuptools>=42" --no-warn-script-location
+        & ".\runtime\python.exe" -m pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu124 --no-warn-script-location
+        & ".\runtime\python.exe" -m pip install flash-attn -i https://xxxxrt666.github.io/PIP-Index/ --no-build-isolation
     }
     "cu128" {
+        & ".\runtime\python.exe" -m pip install psutil ninja packaging wheel "setuptools>=42" --no-warn-script-location
         & ".\runtime\python.exe" -m pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu128 --no-warn-script-location
+        & ".\runtime\python.exe" -m pip install flash-attn -i https://xxxxrt666.github.io/PIP-Index/ --no-build-isolation
     }
     default {
         Write-Error "Unsupported CUDA version: $cuda"

.github/workflows/build_windows_packages.yaml

@@ -31,6 +31,15 @@ jobs:
       - name: Checkout
         uses: actions/checkout@v4
 
+      - name: Install Windows CUDA 12.9
+        if: ${{ runner.os == 'Windows' && matrix.torch_cuda == '12.8' }}
+        uses: Jimver/cuda-toolkit
+        id: cuda-toolkit-win-129
+        with:
+          cuda: 12.9.1
+          method: "network"
+          sub-packages: '["nvcc", "cudart", "visual_studio_integration"]'
+
       - name: Run Build and Upload Script
         shell: pwsh
         run: |

.gitignore

@@ -16,8 +16,8 @@ ffprobe*
 cfg.json
 speakers.json
 ref_audios
-tools/AP_BWE_main/24kto48k/*
-!tools/AP_BWE_main/24kto48k/readme.txt
+tools/AP_BWE/24kto48k/*
+!tools/AP_BWE/24kto48k/readme.txt
 
 # Byte-compiled / optimized / DLL files
 __pycache__/

Docker/miniconda_install.sh

@@ -23,8 +23,10 @@ fi
 
 if [ "$TARGETPLATFORM" = "linux/amd64" ]; then
     "${WGET_CMD[@]}" -O miniconda.sh https://repo.anaconda.com/miniconda/Miniconda3-py311_25.3.1-1-Linux-x86_64.sh
+    SYSROOT_PKG="sysroot_linux-64>=2.28"
 elif [ "$TARGETPLATFORM" = "linux/arm64" ]; then
     "${WGET_CMD[@]}" -O miniconda.sh https://repo.anaconda.com/miniconda/Miniconda3-py311_25.3.1-1-Linux-aarch64.sh
+    SYSROOT_PKG="sysroot_linux-aarch64>=2.28"
 else
     exit 1
 fi
@@ -45,20 +47,36 @@ rm miniconda.sh
 
 source "$HOME/miniconda3/etc/profile.d/conda.sh"
 
+"$HOME/miniconda3/bin/conda" init bash
+
+source "$HOME/.bashrc"
+
 "$HOME/miniconda3/bin/conda" config --add channels conda-forge
 
 "$HOME/miniconda3/bin/conda" update -q --all -y 1>/dev/null
 
 "$HOME/miniconda3/bin/conda" install python=3.11 -q -y
 
-"$HOME/miniconda3/bin/conda" install gcc=14 gxx ffmpeg cmake make unzip -q -y
+"$HOME/miniconda3/bin/conda" install gcc=11 gxx ffmpeg cmake make unzip $SYSROOT_PKG "libstdcxx-ng>=11" -q -y
 
 if [ "$CUDA_VERSION" = "12.8" ]; then
     "$HOME/miniconda3/bin/pip" install torch torchaudio --no-cache-dir --index-url https://download.pytorch.org/whl/cu128
+    "$HOME/miniconda3/bin/conda" install cuda-nvcc=12.8 -c nvidia
 elif [ "$CUDA_VERSION" = "12.6" ]; then
-    "$HOME/miniconda3/bin/pip" install torch==2.6 torchaudio --no-cache-dir --index-url https://download.pytorch.org/whl/cu126
+    "$HOME/miniconda3/bin/pip" install torch torchaudio --no-cache-dir --index-url https://download.pytorch.org/whl/cu126
+    "$HOME/miniconda3/bin/conda" install cuda-nvcc=12.6 -c nvidia
 fi
 
+CUDA_PATH=$(echo "$HOME/miniconda3/targets/"*-linux | awk '{print $1}')
+
+export CUDA_HOME=$CUDA_PATH
+export PATH="$HOME/miniconda3/bin:$PATH"
+export PATH="$CUDA_HOME/bin:$PATH"
+export PATH="$CUDA_HOME/nvvm/bin:$PATH"
+
+"$HOME/miniconda3/bin/pip" install psutil ninja packaging wheel "setuptools>=42"
+"$HOME/miniconda3/bin/pip" install flash-attn -i https://xxxxrt666.github.io/PIP-Index/ --no-build-isolation
+
 "$HOME/miniconda3/bin/pip" cache purge
 
 rm $LOG_PATH

GPT_SoVITS/AR/data/data_module.py

@@ -3,8 +3,8 @@
 from pytorch_lightning import LightningDataModule
 from torch.utils.data import DataLoader
 
-from AR.data.bucket_sampler import DistributedBucketSampler
-from AR.data.dataset import Text2SemanticDataset
+from GPT_SoVITS.AR.data.bucket_sampler import DistributedBucketSampler
+from GPT_SoVITS.AR.data.dataset import Text2SemanticDataset
 
 
 class Text2SemanticDataModule(LightningDataModule):

GPT_SoVITS/AR/data/dataset.py

@@ -220,7 +220,7 @@ def __getitem__(self, idx: int) -> Dict:
 
         flag = 0
         path_bert = "%s/%s.pt" % (self.path3, item_name)
-        if os.path.exists(path_bert) == True:
+        if os.path.exists(path_bert) is True:
             bert_feature = torch.load(path_bert, map_location="cpu")
         else:
             flag = 1

GPT_SoVITS/AR/models/t2s_lightning_module.py

@@ -10,9 +10,9 @@
 import torch
 from pytorch_lightning import LightningModule
 
-from AR.models.t2s_model import Text2SemanticDecoder
-from AR.modules.lr_schedulers import WarmupCosineLRSchedule
-from AR.modules.optim import ScaledAdam
+from GPT_SoVITS.AR.models.t2s_model import Text2SemanticDecoder
+from GPT_SoVITS.AR.modules.lr_schedulers import WarmupCosineLRSchedule
+from GPT_SoVITS.AR.modules.optim import ScaledAdam
 
 
 class Text2SemanticLightningModule(LightningModule):
@@ -42,7 +42,7 @@ def __init__(self, config, output_dir, is_train=True):
     def training_step(self, batch: Dict, batch_idx: int):
         opt = self.optimizers()
         scheduler = self.lr_schedulers()
-        forward = self.model.forward if self.config["train"].get("if_dpo", False) == True else self.model.forward_old
+        forward = self.model.forward if self.config["train"].get("if_dpo", False) is True else self.model.forward_old
         loss, acc = forward(
             batch["phoneme_ids"],
             batch["phoneme_ids_len"],

GPT_SoVITS/AR/models/t2s_lightning_module_onnx.py

@@ -10,9 +10,9 @@
 import torch
 from pytorch_lightning import LightningModule
 
-from AR.models.t2s_model_onnx import Text2SemanticDecoder
-from AR.modules.lr_schedulers import WarmupCosineLRSchedule
-from AR.modules.optim import ScaledAdam
+from GPT_SoVITS.AR.models.t2s_model_onnx import Text2SemanticDecoder
+from GPT_SoVITS.AR.modules.lr_schedulers import WarmupCosineLRSchedule
+from GPT_SoVITS.AR.modules.optim import ScaledAdam
 
 
 class Text2SemanticLightningModule(LightningModule):

GPT_SoVITS/AR/models/t2s_model.py

@@ -9,7 +9,7 @@
 from torchmetrics.classification import MulticlassAccuracy
 from tqdm import tqdm
 
-from AR.models.utils import (
+from GPT_SoVITS.AR.models.utils import (
     dpo_loss,
     get_batch_logps,
     make_pad_mask,
@@ -18,8 +18,8 @@
     sample,
     topk_sampling,
 )
-from AR.modules.embedding import SinePositionalEmbedding, TokenEmbedding
-from AR.modules.transformer import LayerNorm, TransformerEncoder, TransformerEncoderLayer
+from GPT_SoVITS.AR.modules.embedding import SinePositionalEmbedding, TokenEmbedding
+from GPT_SoVITS.AR.modules.transformer import LayerNorm, TransformerEncoder, TransformerEncoderLayer
 
 default_config = {
     "embedding_dim": 512,
@@ -420,7 +420,7 @@ def forward(self, x, x_lens, y, y_lens, bert_feature):
             mask=xy_attn_mask,
         )
         x_len = x_lens.max()
-        logits = self.ar_predict_layer(xy_dec[:, x_len-1:])
+        logits = self.ar_predict_layer(xy_dec[:, x_len - 1 :])
 
         ###### DPO #############
         reject_xy_pos, reject_xy_attn_mask, reject_targets = self.make_input_data(
@@ -432,7 +432,7 @@ def forward(self, x, x_lens, y, y_lens, bert_feature):
             mask=reject_xy_attn_mask,
         )
         x_len = x_lens.max()
-        reject_logits = self.ar_predict_layer(reject_xy_dec[:, x_len-1:])
+        reject_logits = self.ar_predict_layer(reject_xy_dec[:, x_len - 1 :])
 
         # loss
         # from feiteng: 每次 duration 越多, 梯度更新也应该更多, 所以用 sum
@@ -502,7 +502,7 @@ def forward_old(self, x, x_lens, y, y_lens, bert_feature):
             (xy_pos, None),
             mask=xy_attn_mask,
         )
-        logits = self.ar_predict_layer(xy_dec[:, x_len-1:]).permute(0, 2, 1)
+        logits = self.ar_predict_layer(xy_dec[:, x_len - 1 :]).permute(0, 2, 1)
         # loss
         # from feiteng: 每次 duration 越多, 梯度更新也应该更多, 所以用 sum
         loss = F.cross_entropy(logits, targets, reduction="sum")
@@ -724,8 +724,8 @@ def infer_panel_batch_infer(
                 l1 = samples[:, 0] == self.EOS
                 l2 = tokens == self.EOS
                 l = l1.logical_or(l2)
-                removed_idx_of_batch_for_y = torch.where(l == True)[0].tolist()
-                reserved_idx_of_batch_for_y = torch.where(l == False)[0]
+                removed_idx_of_batch_for_y = torch.where(l is True)[0].tolist()
+                reserved_idx_of_batch_for_y = torch.where(l is False)[0]
                 # batch_indexs = torch.tensor(batch_idx_map, device=y.device)[removed_idx_of_batch_for_y]
                 for i in removed_idx_of_batch_for_y:
                     batch_index = batch_idx_map[i]

GPT_SoVITS/AR/models/t2s_model_onnx.py

@@ -5,8 +5,8 @@
 from torch.nn import functional as F
 from torchmetrics.classification import MulticlassAccuracy
 
-from AR.modules.embedding_onnx import SinePositionalEmbedding, TokenEmbedding
-from AR.modules.transformer_onnx import LayerNorm, TransformerEncoder, TransformerEncoderLayer
+from GPT_SoVITS.AR.modules.embedding_onnx import SinePositionalEmbedding, TokenEmbedding
+from GPT_SoVITS.AR.modules.transformer_onnx import LayerNorm, TransformerEncoder, TransformerEncoderLayer
 
 default_config = {
     "embedding_dim": 512,