# 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