requirements.txt

--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