| # origin cluster versions (Python 3.12); vLLM with CUDA 12.8 wheels (Blackwell B200 / RTX 5090 need torch 2.8 cu128) | |
| vllm==0.11.0 | |
| torch==2.8.0 | |
| transformers==4.57.6 | |
| decord>=0.6.0 # frame decoder (fallbacks: opencv-python-headless, ffmpeg on PATH) | |
| opencv-python-headless>=4.10 | |
| pillow>=10 | |
| pyyaml>=6 | |
| requests>=2.32 | |
| huggingface_hub>=0.34 | |