--find-links https://download.pytorch.org/whl/torch_stable.html # we seem to have an issue with Torch 2.8 # I believe it works but it is incompatible with older weights formats? # it looks like they changed the checkpoint format or something # python3.10/site-packages/torch/distributed/checkpoint/default_planner.py", line 471, in create_default_local_load_plan # RuntimeError: Missing key in checkpoint state_dict: lr_scheduler._is_initial. # #torch==2.8.0 #torchvision==0.23.0 #torchdata==0.11.0 #torchao==0.12.0 #torchcodec # flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.8.3/flash_attn-2.8.3+cu12torch2.8cxx11abiFALSE-cp310-cp310-linux_x86_64.whl # # if we revert back to torch 2.7, then we get another error: # Missing key in checkpoint state_dict: optimizer.param_groups.scale_shift_table.decoupled_weight_decay. # #torch==2.7.1 #torchvision==0.22.1 #torchdata==0.11.0 #torchao==0.12.0 #torchcodec==0.5.0 #flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.7cxx11abiFALSE-cp310-cp310-linux_x86_64.whl # # so in the end, we have to revert back to the 2.6: # torch==2.6.0 torchvision==0.21.0 torchdata==0.10.1 torchao==0.9.0 # for torch 2.6, we must use torchcodec 0.2 torchcodec==0.2.1 --index-url=https://download.pytorch.org/whl/cu128 flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.6cxx11abiFALSE-cp310-cp310-linux_x86_64.whl # something broke in Transformers > 4.55.4 transformers==4.55.4 # For GPU monitoring of NVIDIA chipsets pynvml # Pin datasets to 3.6.0 to avoid VideoDecoder issues with 4.0.0 # see https://github.com/huggingface/finetrainers/issues/424#issuecomment-3255342554 datasets==3.6.0 # we are waiting for the next PyPI release #finetrainers==0.1.0 finetrainers @ git+https://github.com/huggingface/finetrainers.git@main # temporary fix for pip install bug: #finetrainers @ git+https://github.com/jbilcke-hf/finetrainers-patches.git@fix_missing_sft_trainer_files # it is recommended to always use the latest version diffusers @ git+https://github.com/huggingface/diffusers.git@main imageio imageio-ffmpeg # for youtube video download pytube pytubefix # for scene splitting scenedetect[opencv] # for llava video / captionning pillow pillow-avif-plugin polars einops open_clip_torch av==14.1.0 # for some reason LLaVA-NeXT has ceased to work, # but I think it it due to a breaking change in Transformers git+https://github.com/LLaVA-VL/LLaVA-NeXT.git # for our frontend gradio==5.33.1 gradio_toggle gradio_modal # used for the monitor matplotlib