# Notes: # The base image `vllm/vllm-openai:v0.8.5` already ships: # - torch 2.6 / torchaudio / torchvision (CUDA 12.4 build) # - vllm 0.8.5 # - transformers (compatible with vllm 0.8.5) # - flash_attn 2.7.x (used by vllm's flash-attn backend) # Re-installing any of them here triggers a full source build of flash_attn # (no PyPI wheel available), which OOM-kills the HF Space *build* # container. Keep this list to lightweight, app-only dependencies. # # IMPORTANT version pin: # The transformers shipped in vllm 0.8.5 requires huggingface-hub<1.0. # Without an explicit upper bound here, pip happily upgrades # huggingface-hub to 1.x (released 2025-Q4), which then makes # `import transformers` -> `import vllm` fail at container start with: # ImportError: huggingface-hub>=0.30.0,<1.0 is required ... huggingface-hub>=0.30.0,<1.0 librosa>=0.11.0 nnAudio>=0.3.3 gradio>=4.44 wheel>=0.45.1 # Required by megatron/tokenizer.py (OmegaConf.load on the vocab file). # Pure-Python, no native deps -> safe for the locked transformers/torch tree. omegaconf>=2.3,<3.0 # Required by: # - mucodec/descript_quantize3.py # - mucodec/.../musicfm/model/musicfm_25hz.py # - mucodec/.../musicfm/modules/{conv,random_quantizer}.py # Pure-Python, zero deps. The vllm base image often already bundles it, # but listing it explicitly removes the guesswork. einops>=0.7