The previous cu121 torchaudio linked against libcudart.so.13 which
doesn't exist on the ZeroGPU worker. Trying cu124 build of torchaudio
2.5.1 to match torch 2.5.1 preinstalled on ZeroGPU. Added diagnostic
print to confirm exact torch+CUDA versions in the logs.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add torchaudio via PyTorch cu121 extra-index-url in requirements.txt
pre-install.sh is not auto-executed by HF Spaces. Using
--extra-index-url in requirements.txt ensures pip fetches the
CUDA 12.1-compatible torchaudio wheel that matches the pre-installed
torch on the ZeroGPU worker.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Fix torchaudio CUDA mismatch: install via PyTorch cu121 index in pre-install.sh
torchaudio from plain pip installs CPU/wrong-CUDA build causing
libcudart.so.13 error on ZeroGPU. Install via the PyTorch CUDA 12.1
wheel index to match the pre-installed torch on the worker.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Fix torchaudio missing dep and TARO sys.path for bare imports
The "ValueError: When localhost is not accessible, a shareable link must
be created" is a known incompatibility between the spaces SDK proxy and
Gradio 4.44.x. Gradio 5 resolves this. Update sdk_version in README
and pin gradio==5.25.0 in requirements.txt.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Fix: pin gradio==4.44.1 to fix ZeroGPU localhost ValueError
Jack Wu
504eda5
Fix launch() call for ZeroGPU — remove server_name override
Jack Wu
78dbbf7
Fix: call get_random_seed(), add server_name to launch() for ZeroGPU
Jack Wu
d3011cb
Fix mmcv install: use mim via pre-install.sh instead of pip source build
Jack Wu
3b3c373
Fix: move mmcv to requirements.txt, remove runtime install that caused infinite rebuild loop
Jack Wu
c59ecb5
Fix HunyuanFoley checkpoint path and add README metadata
Jack Wu
1896e81
Add HuggingFace Spaces metadata to README
Jack Wu
657615c
Remove non-inference files from all three model folders
Jack Wu
22b2290
Keep only what is imported at runtime by app.py:
- TARO: remove dataset.py, infer.py, loss.py, train.py, train.sh, preprocess/, README.md
- MMAudio: remove train.py, batch_eval.py, eval_onsets.py, demo.py, gradio_demo.py,
config/, docs/, sets/, training/, README.md, LICENSE, .gitignore
- HunyuanFoley: remove infer.py, gradio_app.py, tests/, assets/, build_package.sh,
download_test_videos.sh, DEVELOPMENT.md, INSTALL.md, LICENSE, MANIFEST.in,
NOTICE, pytest.ini, README.md, .gitattributes, .gitignore, .pre-commit-config.yaml
Update .gitignore to permanently exclude all of the above.
Add .gitignore and untrack MMAudio binary training examples
Jack Wu
3886668
Remove MMAudio/training/example_audios/ and example_videos/ from the
git index — these .flac and .mp4 files are not needed in the Space
and were causing HF pre-receive hook rejection of binary files.
.
Jack Wu
c60109f
Restructure app.py: multi-model support (TARO, MMAudio, HunyuanFoley)
Jack Wu
6cf4573
- Split into three tabbed UI sections, one per model
- Updated checkpoint repo/folder paths to JackIsNotInTheBox/Generate_Audio_for_Video_Checkpoints
with TARO/, MMAudio/, HunyuanFoley/ subfolders
- TARO: preserve exact infer.py invocation (CAVP+VideoOnsetNet+MMDiT+AudioLDM2 decoder)
with sliding-window segmentation; fix samplers tuple return indexing
- MMAudio: use official load_video()+generate() pipeline from gradio_demo.py;
override model paths to our HF checkpoint repo; large_44k_v2 variant
- HunyuanFoley: use official load_model()+feature_process()+denoise_process()
pipeline; batch inference for multiple samples; xl/xxl size selector
- Add model-specific optimal duration constants from source configs
- Shared slot UI helper; MAX_SLOTS=8 maintained across all tabs
Show slot 0 by default; reveal extra slots on Generations slider drag
