syncworld-simulator / requirements.txt
multimodalart's picture
multimodalart HF Staff
add qwen-vl-utils (imported by cosmos processors package)
7e77941 verified
Raw
History Blame Contribute Delete
1.08 kB
# FlashAttention-2 for sm_120 (Blackwell). MANDATORY: cosmos_framework's attention
# dispatcher offers only ["flash2", "natten"] on sm>=80 -- there is no SDPA fallback.
# torch is pinned to 2.11.0 so the prebuilt wheel's cell matches exactly.
torch==2.11.0
torchvision==0.26.0
https://huggingface.co/datasets/multimodalart/zerogpu-blackwell-wheels/resolve/main/wheels/pt211-cu130-cp312/flash_attn-2.8.3-cp312-cp312-linux_x86_64.whl
# The authors' inference stack (NVIDIA cosmos-framework fork shipped with SyncWorld).
# Pinned to the commit the checkpoint was released with.
cosmos-framework @ git+https://github.com/UMass-Embodied-AGI/SyncWorld.git@f026490eb5b15640087b575e13952c86c3d9b7b3
# Imported by the modules the demo touches but not declared by cosmos-framework's
# base extra (they live in its `train` / `dev` extras, which also pull gradio).
attrs
einx
iopath
qwen-vl-utils
multi-storage-client==0.44.0
boto3
pandas
pyarrow
matplotlib
opencv-python-headless
mediapy
ftfy
webdataset
scikit-image
tensorstore
zarr
h5py
lz4
more-itertools
wandb
imageio
imageio-ffmpeg
av