Spaces:
Running on Zero
Running on Zero
Commit History
-
optimize gpu memory usage
1fd58ee -
feat: add svc inference code and webui
2566adf
-
Fix NeMo CUDA error 35 on ZeroGPU by disabling CUDA graphs in decoding config
The previous fix disabled CUDA graphs on the decoding_computer object, but transcribe(timestamps=True) calls change_decoding_strategy() which rebuilds the decoder from cfg, re-enabling CUDA graphs. Setting cfg.decoding.greedy.use_cuda_graph_decoder=False persists across rebuilds. Also loads the English ASR model eagerly at startup so ZeroGPU can properly hijack CUDA calls at global scope, instead of lazy-loading inside a @spaces.GPU-decorated function. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>e36805d -
Disable NeMo CUDA Graphs to fix CUDA failure 35 on RNNT decoding
NeMo 2.6 RNNT decoding uses CUDA Graphs which fail with error 35 on CUDA 12.8 + PyTorch 2.9 (NVIDIA-NeMo/NeMo#15145). Disable them on the parakeet-tdt model after loading. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>a263769 -
Update preprocess/tools/lyric_transcription.py
Verified signatureb773a39 -
Update preprocess/tools/g2p.py
Verified signature0196c78 -
Fix model download reliability and silent import failures
Extract ensure_pretrained_models() into ensure_models.py and call it from webui.py before AppState init, so models are downloaded regardless of entry point. Narrow except clauses in preprocess/tools/__init__.py from Exception to ImportError so real errors propagate instead of silently setting classes to None. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>19b877f -
Upload folder using huggingface_hub
Verified signaturec7f3ffb
