torch
torchvision
numpy
pillow
huggingface_hub
accelerate
qwen-vl-utils
bitsandbytes
optimum
transformers>=4.57.1
