gradio numpy torch==2.4.0 torchvision==0.19.0 Pillow transformers matplotlib decord timm einops accelerate bitsandbytes peft tensorboardX flash_attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.6.3/flash_attn-2.6.3+cu123torch2.4cxx11abiFALSE-cp310-cp310-linux_x86_64.whl spaces pypandoc fastapi wheel imageio