Cosmos Dependencies

Repository for building and publishing Cosmos dependencies.

Index URLs:

• all
• cu126_torch29
• cu128_torch29
• cu129_torch28
• cu130_torch29

Setup

Install system dependencies:

uv

export XDG_BIN_HOME="${XDG_BIN_HOME:-$HOME/.local/bin}"
curl -LsSf https://astral.sh/uv/install.sh | sh
source $XDG_BIN_HOME/env

just

pushd "$XDG_BIN_HOME" && curl https://zyedidia.github.io/eget.sh | sh && popd
eget casey/just --to="$XDG_BIN_HOME"

If uploading wheels, gh:

eget cli/cli --asset=.tar.gz --to="$XDG_BIN_HOME"
gh auth login

Build Wheels

Run the docker container:

just docker-<cuda_version>

Optionally, test the environment:

just build-dummy <cuda_version>

Build a single package:

just build <package_name> <package_version> <python_version> <torch_version> <cuda_version>

Example:

just docker-cu128
just build-dummy 12.8
just build natten 0.21.0 3.12 2.7 12.8

On the host, fix file permissions:

just fix-permissions

Upload Wheels

1. Upload wheels

just upload <wheel_path>

1. Create and locally host the package index

just index-serve

1. Open a PR and merge to cosmos-dependencies.

Bump Version

gh release download --repo nvidia-cosmos/cosmos-dependencies v$(uv version --short) -D tmp/assets --pattern '*'
uv version --bump minor
gh release upload --repo nvidia-cosmos/cosmos-dependencies v$(uv version --short) tmp/assets/*

Version Constraints

• cuda
  • torch
• gcc
  • cuda
• flash-attn
  • transformer-engine FlashAttentionUtils.max_version

Contributing

We thrive on community collaboration! NVIDIA-Cosmos wouldn't be where it is without contributions from developers like you. Check out our Contributing Guide to get started, and share your feedback through issues.

Big thanks 🙏 to everyone helping us push the boundaries of open-source physical AI!

License

This repository is licensed under the Apache License 2.0.