perf(mdx): batch inference windows and reduce device transfers

.github/workflows/run-integration-tests.yaml

@@ -195,9 +195,35 @@ jobs:
           python-version: '3.12'
           cache: poetry
       - name: Install ffmpeg
-        run: choco install ffmpeg -y --no-progress
+        shell: pwsh
+        run: |
+          choco install ffmpeg -y --no-progress
+          if ($LASTEXITCODE -ne 0) { throw "FFmpeg installation failed with exit code $LASTEXITCODE" }
+          # Chocolatey updates the persistent PATH, not this runner process.
+          $env:Path = [Environment]::GetEnvironmentVariable('Path', 'Machine') + ';' + [Environment]::GetEnvironmentVariable('Path', 'User') + ';' + $env:Path
+          foreach ($tool in @('ffmpeg.exe', 'ffprobe.exe')) {
+            $executable = (Get-Command $tool -CommandType Application -ErrorAction Stop).Source
+            Split-Path -Parent $executable | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          }
       - name: Install Poetry dependencies (CPU)
         run: poetry install -E cpu
+      - name: Verify FFmpeg export in the Poetry environment
+        shell: pwsh
+        run: |
+          @'
+          import subprocess
+          import tempfile
+          from pathlib import Path
+          from pydub import AudioSegment
+
+          for executable in ('ffmpeg', 'ffprobe'):
+              subprocess.run([executable, '-version'], check=True)
+          with tempfile.TemporaryDirectory() as directory:
+              output = Path(directory) / 'ffmpeg-check.flac'
+              AudioSegment.silent(duration=100).export(str(output), format='flac').close()
+              assert output.stat().st_size > 0
+              assert len(AudioSegment.from_file(str(output))) == 100
+          '@ | poetry run python -
       - name: Cache separation models
         uses: actions/cache@v4
         with:

README.md

@@ -748,7 +748,7 @@ You can also rename specific stems:
 - **`use_native_fp16`:** (Optional) Convert a verified model to native float16 inference. Currently supported for MelBand RoFormer and BS-RoFormer on MPS and CUDA. Mutually exclusive with `use_autocast=True`; unsupported combinations warn and continue in float32. `Default: False`
 - **`use_torch_compile`:** (Optional) Compile verified repeated model blocks. This can be combined with float32 or autocast for MelBand RoFormer and BS-RoFormer on CPU, MPS, and CUDA, and with native float16 on MPS and CUDA. A fresh compiler cache can make the first run slower; unsupported combinations warn and continue in eager mode. `Default: False`
 - **`use_directml`:** (Optional) Flag to use DirectML for hardware-accelerated inference on Windows AMD/Intel GPUs (experimental; requires the `dml` extra and only takes effect when CUDA and Apple Silicon MPS are unavailable). `Default: False`
-- **`mdx_params`:** (Optional) MDX Architecture Specific Attributes & Defaults. `Default: {"hop_length": 1024, "segment_size": 256, "overlap": 0.25, "batch_size": 1, "enable_denoise": False}`
+- **`mdx_params`:** (Optional) MDX Architecture Specific Attributes & Defaults. `Default: {"hop_length": 1024, "segment_size": 256, "overlap": 0.25, "batch_size": 1, "enable_denoise": False}` `batch_size` groups independent windows for converted Torch models and dynamic-batch ONNX graphs; fixed-batch ONNX graphs retain their declared batch size. Larger batches can improve throughput at the cost of memory; overlap and denoise settings remain independent.
 - **`vr_params`:** (Optional) VR Architecture Specific Attributes & Defaults. `Default: {"batch_size": 1, "window_size": 512, "aggression": 5, "enable_tta": False, "enable_post_process": False, "post_process_threshold": 0.2, "high_end_process": False}`
 - **`demucs_params`:** (Optional) Demucs Architecture Specific Attributes & Defaults. `Default: {"segment_size": "Default", "shifts": 2, "overlap": 0.25, "segments_enabled": True}` _(Note: `segment_size` "Default" uses the model's internal default, typically 40 for older Demucs models and 10 for Demucs v4/htdemucs)_
 - **`mdxc_params`:** (Optional) MDXC Architecture Specific Attributes & Defaults. `Default: {"segment_size": 256, "override_model_segment_size": False, "batch_size": None, "overlap": None, "pitch_shift": 0}` (`None` uses the model YAML's `inference` value, falling back to `1` for batch size and `8` for overlap.)