.

.github/build_windows_packages.ps1

@@ -92,6 +92,30 @@ Get-ChildItem -Path $x64Path -Directory | Where-Object {
     }
 }
 
+$ffmpegUrl = "https://github.com/BtbN/FFmpeg-Builds/releases/download/latest/ffmpeg-n7.1-latest-win64-gpl-shared-7.1.zip"
+$zipPath = Join-Path $tmpDir "ffmpeg.zip"
+
+Invoke-WebRequest -Uri $ffmpegUrl -OutFile $zipPath
+
+Expand-Archive -Path $zipPath -DestinationPath $tmpDir -Force
+
+$extractedDir = Get-ChildItem -Path $tmpDir -Directory | Where-Object { $_.Name -match "^ffmpeg.*win64.*gpl.*shared" } | Select-Object -First 1
+
+if (-not $extractedDir) {
+    Write-Error "Can Not Find FFmpeg Folder"
+    exit 1
+}
+
+$runtimeDir =  "$srcDir\runtime"
+New-Item -ItemType Directory -Force -Path $runtimeDir | Out-Null
+
+$src = Join-Path $extractedDir.FullName "bin"
+if (Test-Path $src) {
+    Move-Item -Path (Join-Path $src '*') -Destination $runtimeDir -Force
+}
+
+Write-Host "FFmpeg Downloaded and extracted to $runtimeDir"
+
 function DownloadAndUnzip($url, $targetRelPath) {
     $filename = Split-Path $url -Leaf
     $tmpZip = "$tmpDir\$filename"
@@ -157,7 +181,6 @@ switch ($cuda) {
 }
 
 Write-Host "[INFO] Installing dependencies..."
-& ".\runtime\python.exe" -m pip install --pre torchcodec --index-url https://download.pytorch.org/whl/nightly/cpu
 & ".\runtime\python.exe" -m pip install -r extra-req.txt --no-deps --no-warn-script-location
 & ".\runtime\python.exe" -m pip install -r requirements.txt --no-warn-script-location
 

GPT_SoVITS/Accelerate/PyTorch/__init__.py

@@ -7,6 +7,10 @@ from .structs import T2SRequest, T2SResult
 from .t2s_engine import T2SEngine as T2SEngineTorch
 
 torch.set_grad_enabled(False)
+torch.backends.cudnn.benchmark = True
+torch.backends.cudnn.enabled = True
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.backends.cudnn.allow_tf32 = True
 
 backends = ["torch_varlen"]
 if torch.cuda.is_available():

GPT_SoVITS/Accelerate/PyTorch/t2s_engine.py

@@ -184,7 +184,7 @@ class T2SEngine(T2SEngineProtocol):
                 case "mtia":
                     torch.mtia.empty_cache()
                 case "cpu":
-                    gc.collect()
+                    gc.collect(1)
 
             if request.use_cuda_graph and self.graphcache.is_applicable:
                 self.graphcache.release_graph(session)

GPT_SoVITS/inference_webui.py

@@ -1,5 +1,6 @@
 import argparse
 import contextlib
+import gc
 import logging
 import os
 import re
@@ -966,10 +967,11 @@ def get_tts_wav(
     gr.Info(f"{infer_speed_avg:.2f} Token/s", title="Infer Speed")
     gr.Info(f"{rtf_value:.2f}", title="RTF")
 
+    yield opt_sr, (audio_opt_n * 32767).astype(np.int16)
+
     if torch.cuda.is_available():
         torch.cuda.empty_cache()
-
+    gc.collect()
-    yield opt_sr, (audio_opt_n * 32767).astype(np.int16)
 
 
 def split(todo_text):

GPT_SoVITS/text/g2pw/onnx_api.py

@@ -91,7 +91,7 @@ class G2PWOnnxConverter:
         sess_options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
         sess_options.execution_mode = onnxruntime.ExecutionMode.ORT_SEQUENTIAL
         sess_options.intra_op_num_threads = 2 if torch.cuda.is_available() else 0
-        if "CUDAExecutionProvider" in onnxruntime.get_available_providers():
+        if "CUDAExecutionProvider" in onnxruntime.get_available_providers() and torch.cuda.is_available():
             self.session_g2pW = onnxruntime.InferenceSession(
                 os.path.join(uncompress_path, "g2pW.onnx"),
                 sess_options=sess_options,

README.md

@@ -45,12 +45,13 @@ Unseen speakers few-shot fine-tuning demo:
 
 ## Infer Speed
 
-| Device      | RTF   | Batch Size | Backend                     |
+| Device      | RTF   | TTFB   | Batch Size | Backend                     |
-| ----------- | ----- | ---------- | --------------------------- |
+| ----------- | ----- | ------ | ---------- | --------------------------- |
-| RTX 5090    | 0.05  | 1          | Flash Attn Varlen CUDAGraph |
+| RTX 5090    | 0.05  | 150 ms | 1          | Flash Attn Varlen CUDAGraph |
-| Apple M4    | 0.21  | 1          | MLX Quantized Affined       |
+| RTX 4090    | 0.014 | UNK    | 24         | Flash Attn Varlen CUDAGraph |
-| RTX 4090    | 0.014 | 24         | Flash Attn Varlen CUDAGraph |
+| RTX 4060 Ti | 0.07  | 460 ms | 1          | Flash Attn Varlen CUDAGraph |
-| RTX 4060 Ti | 0.028 | 28         | Flash Attn Varlen CUDAGraph |
+| RTX 4060 Ti | 0.028 | UNK    | 28         | Flash Attn Varlen CUDAGraph |
+| Apple M4    | 0.21  |        | 1          | MLX Quantized Affined       |
 
 **User guide: [简体中文](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e) | [English](https://rentry.co/GPT-SoVITS-guide#/)**
 

go-webui.bat

@@ -2,5 +2,5 @@ set "SCRIPT_DIR=%~dp0"
 set "SCRIPT_DIR=%SCRIPT_DIR:~0,-1%"
 cd /d "%SCRIPT_DIR%"
 set "PATH=%SCRIPT_DIR%\runtime"
-runtime\python.exe -I webui.py zh_CN
+runtime\python.exe -s webui.py zh_CN
 pause

go-webui.ps1

@@ -3,5 +3,5 @@ chcp 65001
 Set-Location $PSScriptRoot
 $runtimePath = Join-Path $PSScriptRoot "runtime"
 $env:PATH = "$runtimePath"
-& "$runtimePath\python.exe" -I "$PSScriptRoot\webui.py" zh_CN
+& "$runtimePath\python.exe" -s "$PSScriptRoot\webui.py" zh_CN
 pause

install.ps1

@@ -255,7 +255,6 @@ switch ($Device) {
 Write-Success "PyTorch Installed"
 
 Write-Info "Installing Python Dependencies From requirements.txt..."
-Invoke-Pip --pre torchcodec --index-url https://download.pytorch.org/whl/nightly/cpu
 Invoke-Pip -r extra-req.txt --no-deps
 Invoke-Pip -r requirements.txt
 Write-Success "Python Dependencies Installed"

requirements.txt

@@ -13,6 +13,7 @@ peft
 py-cpuinfo
 pypinyin
 split-lang
+torchaudio
 torchcodec
 transformers
 tensorboard