Spaces:
Sleeping
Sleeping
Commit History
-
Fix deprecation warnings, CPU batch size, and root route
- torch_dtype → dtype in from_pretrained and pipeline (transformers 4.49+) - Move max_new_tokens to generate_kwargs in pipeline constructor to silence generation_config conflict warning - batch_size: 16 → 2 on CPU (16 parallel chunks wastes RAM, not faster), keep 8 on CUDA - Add GET / → redirect to /docs so HF Space health probes return 200 instead of 404 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>9308938 -
Fix HF Space startup: CPU wheels, auto-detect device
- Switch torch install to CPU wheels (default). CUDA wheels are 2.5 GB and cause build timeouts on HF Spaces free hardware. GPU deploys can pass --build-arg TORCH_INDEX=.../cu126 at build time. - Remove DEVICE=cuda from Dockerfile ENV: was crashing on CPU Spaces because no CUDA device exists. Now empty so Python auto-detects. - Change MODEL_PATH default to openai/whisper-large-v3 (a valid HF Hub ID) instead of a local path that doesn't exist in the container. - Remove --index-url from requirements-api.txt so pip doesn't try to pull all packages through the CUDA wheel index. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>27216ff -
Fix HuggingFace model inference hang
- Remove use_safetensors=True: auto-detect format to avoid hangs when a repo lacks .safetensors files - Only pass prompt_ids when initial_prompt is non-empty: passing the bare SOT token from get_prompt_ids("") conflicts with the pipeline's forced language/task tokens and causes infinite generation Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>54c1995 -
Add README with Space config
e333dd9 -
Deploy to HF Spaces
0db822c