NVIDIA/cosmos-framework

Latest recorded source activity: October 31, 2026 at 00:08

SkillsMP catalog refreshed: November 2, 2026 at 18:02

skills collected: 29

GitHub stars: 588
GitHub forks: 188

Install with an AI assistant

Copy this prompt into the AI assistant you're using.

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/NVIDIA/cosmos-framework.

Repository README

Repository README source: README.md

Read the source README

Expand full README

<img src="https://github.com/user-attachments/assets/28f2d612-bbd6-44a3-8795-833d05e9f05f" width="274" alt="NVIDIA Cosmos"/> <a href="https://github.com/NVIDIA/Cosmos">NVIDIA Cosmos</a> | 🤗 <a href="https://huggingface.co/collections/nvidia/cosmos3">Cosmos 3 </a> Part of the <a href="https://github.com/NVIDIA/Cosmos">NVIDIA Cosmos</a> project family — the training and serving framework repository. # Cosmos-Framework **Cosmos-Framework** is an end-to-end framework for training and serving world models, including the **Cosmos3** model family. Everything lives in a single top-level [`cosmos_framework/`](./cosmos_framework) Python package: - **Training** — distributed FSDP / TP / CP / PP trainer, native DCP checkpoints with HuggingFace `safetensors` import/export, JSONL / WebDataset / LeRobot dataset adapters. Entry point: `cosmos_framework.scripts.train`. See [`docs/training.md`](./docs/training.md). - **Inference** — Diffusers / Transformers / vLLM backends with offline batch generation and online serving (Ray + Gradio). Entry point: `cosmos_framework.scripts.inference`. Ecosystem-facing shim libraries (lightweight standalone wrappers for downstream projects) live under [`packages/`](./packages). For Cosmos3-Nano generation on GPUs with limited VRAM, see the standalone [Diffusers CPU-GPU offloading adapter](./packages/diffusers-cpu-offloading/README.md). - **Policy serving** — a server-client pipeline for deploying Cosmos3 policies. See [`docs/action_policy_droid_server.md`](./docs/action_policy_droid_server.md). - **Post-training and RL** — supervised, preference-based, reinforcement-learning, distillation, diffusion, and embodied-policy optimization under `cosmos_framework.posttrain`. See [`docs/posttrain/`](./docs/posttrain/index.md). ## Cosmos 3 **Cosmos 3** is our newest model family [[Report]](https://research.nvidia.com/labs/cosmos-lab/cosmos3/technical-report.pdf) [[Website]](https://research.nvidia.com/labs/cosmos-lab/cosmos3/). It is a suite of omnimodal world models designed to jointly process and generate language, images, video, audio, and action sequences within a unified Mixture-of-Transformers architecture. By supporting highly flexible input-output configurations, it seamlessly unifies critical modalities for Physical AI — effectively subsuming vision-language models, video generators, world simulators, and world-action models into a single framework. For a guided experience to test out Cosmos3, please visit [[Cosmos]](https://github.com/nvidia/cosmos). ## Framework Documentation - [Quickstart](#setup) - [Setup](./docs/setup.md) - [Training (Supervised Fine-Tuning)](./docs/training.md) - [JSONL Dataset](./docs/dataset_jsonl.md) - [Inference](./docs/inference.md) - [Evaluation](./docs/evaluation.md) - [Policy Server](./docs/action_policy_droid_server.md) - [Post-training and RL](./docs/posttrain/index.md) - [Step-by-step container run guide](./docs/posttrain/howto/run-posttrain.md) - [Agent Skills](#agent-skills) - Reference - [Code Structure](./docs/code_structure.md) - [Environment Variables](./docs/environment_variables.md) - [FAQ](./docs/faq.md) - [AGENTS.md — repo map for coding agents](./AGENTS.md) ## Setup For more details and alternative installation methods, see [Setup](./docs/setup.md#installation). Before installing, make sure your machine meets the [System Requirements](./docs/setup.md#system-requirements). If you want a curated PyTorch + CUDA environment, start from the [recommended NVIDIA NGC base image](./docs/setup.md#recommended-base-image). Install system dependencies: ```shell sudo apt-get install -y --no-install-recommends curl ffmpeg git-lfs libgl1 libglib2.0-0 libx11-dev tree wget ``` Install the package with the required `uv` version, `0.11.29` (pick a dependency group supported by your driver — see [CUDA Variants](./docs/setup.md#cuda-variants)): ```shell # CUDA 13.0 (recommended) uv sync --all-extras --group=cu130-train # Or, for CUDA 12.8: # uv sync --all-extras --group=cu128-train source .venv/bin/activate ``` Run the [GPU and media checks](./docs/setup.md#verify-the-gpu-and-media-runtime) before downloading weights. Preserve any CUDA compatibility libraries required by your driver; do not clear `LD_LIBRARY_PATH` unconditionally. If you are starting from the recommended NGC image (`nvcr.io/nvidia/pytorch:26.06-py3`), see the [one-shot quickstart](./docs/setup.md#quickstart-from-the-recommended-base-image). ## Training For the full guide (data preparation, base-checkpoint conversion, parallelism strategies, mixed precision, resuming), see [Training](./docs/training.md). The number of GPUs required depends on the recipe; the shipped recipes under [`examples/`](./examples/README.md) are 8-GPU configurations (tested on 8× H100 80 GB) launched via their paired launch shells, e.g.: ```shell bash examples/launch_sft_vision_nano.sh ``` Users may adjust the GPU count to match their model and underlying hardware architecture — tune `NPROC_PER_NODE` and the parallelism degrees (DP/CP/FSDP shard) in the recipe accordingly. ### Transfer-Auto and Sim-Auto - **Cosmos3-Nano-Transfer-Auto:** Start from **Cosmos3-Nano** to reproduce the training curriculum, or directly fine-tune the **released Transfer-Auto checkpoint** on your own data. See the [Transfer-Auto training skill](./.agents/skills/cosmos3-nano-transfer-auto/post-training/SKILL.md) for both paths. - **Cosmos3-Nano-Sim-Auto:** For domain or sensor adaptation, fine-tune Transfer-Auto first, then distill that adapted checkpoint through **TF → TF-dCM → SF-DMD**. See the [Sim-Auto training skill](./.agents/skills/cosmos3-nano-sim-auto/post-training/SKILL.md) and [training guide](./docs/training_sim_auto.md). We recommend completing adaptation before distillation, rather than applying ordinary supervised fine-tuning (SFT) to the released distilled Sim-Auto checkpoint, to preserve its few-step generation behavior. The released Sim-Auto checkpoint can be used directly for [inference](./docs/inference_sim_auto.md). ## Inference See [Inference](./docs/inference.md) for the full guide — launch commands, supported modes, parallelism presets, and troubleshooting. Quick single-GPU launch: ```shell python -m cosmos_framework.scripts.inference \ --parallelism-preset=latency \ -i "inputs/omni/t2v.json" \ -o outputs/omni_nano \ --checkpoint-path Cosmos3-Nano \ --seed=0 ``` ## Policy Server See [Policy Server](./docs/action_policy_droid_server.md) for the full guide. ## Post-training and RL `cosmos_framework.posttrain` provides a common runtime for offline and online post-training, from single-host recipes to distributed Ray execution. Start with the [Posttrain documentation](./docs/posttrain/index.md) or follow the [step-by-step container run guide](./docs/posttrain/howto/run-posttrain.md). The repository has two SFT paths. Use the main [Training](./docs/training.md) path for Cosmos3 generator, reasoner, and policy recipes built on the framework's native configuration, checkpoint, and distributed-training stack. Use [Posttrain SFT](./docs/posttrain/guides/sft.md) for Hugging Face-based models outside the Cosmos3 recipes. It currently supports Qwen text models, dense Qwen3.5/3.8 multimodal models, and Qwen3-VL/Qwen3-VL-MoE, with a common Hugging Face foundation for extending support to additional compatible models. | Algorithm | Description | Guide | | ------------ | ------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------- | | DPO | Offline preference optimization | [DPO](./docs/posttrain/guides/dpo.md) | | GRPO | Group-relative online policy optimization | [GRPO](./docs/posttrain/guides/grpo.md) | | PPO | Actor-critic policy optimization, including embodied-agent training | [PPO](./docs/posttrain/guides/ppo.md) · [Embodied PPO examples](./docs/posttrain/howto/run-posttrain.md#8-embodied-ppo) | | OPD | On-policy distillation from a teacher model | [OPD](./docs/posttrain/guides/opd.md) | | DiffusionNFT | Online reinforcement learning for diffusion generators | [DiffusionNFT](./docs/posttrain/guides/nft.md) | | SFT | Supervised fine-tuning | — | Model integrations include Cosmos3, GR00T-N1.7, π0.5, Alpamayo through AlpaGym, Qwen and Qwen-VL families, and general Hugging Face text and vision-language backends. Algorithm support is defined by runtime and data contracts rather than a fixed model allowlist; shipped recipes demonstrate qualified combinations, while additional compatible models can use the documented model, DataPacker, reward, and simulator extension interfaces. ## Agent Skills Coding agents (Claude Code, Codex CLI, Cursor, and other AGENTS.md-aware tools) can load task-specific instructions for this repo from the bundled AgentSkills. Each skill is a self-contained `SKILL.md` that the agent invokes automatically when the user's request matches its description. Skills live in [`.agents/skills/`](./.agents/skills) (canonical) and are mirrored under [`.claude/skills/`](./.claude/skills) for Claude Code: | Skill | When it activates | | -------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------- | | [`cosmos3-setup`](./.agents/skills/cosmos3-setup/SKILL.md) | Installation, environment setup, checkpoint downloads, Docker. | | [`cosmos3-codebase-nav`](./.agents/skills/cosmos3-codebase-nav/SKILL.md) | "Where is X" / "where do I change parameter Y" questions across `cosmos_framework/`. | | [`cosmos3-inference`](./.agents/skills/cosmos3-inference/SKILL.md) | Running offline or online inference, parallelism, sampling parameters. | | [`cosmos3-post-training`](./.agents/skills/cosmos3-post-training/SKILL.md) | SFT post-training end-to-end: data prep, DCP conversion, launch, export. | | [`cosmos-automl`](./.agents/skills/cosmos-automl/SKILL.md) | Reasoner hyperparameter search and GEPA prompt optimization; [main/development-branch setup](./docs/automl.md#installation-from-source). | | [`cosmos3-env-troubleshoot`](./.agents/skills/cosmos3-env-troubleshoot/SKILL.md) | Diagnosing install/runtime errors (ImportError, CUDA, Docker, checkpoint failures). | See [`AGENTS.md`](./AGENTS.md) for the canonical repo map that agents load first; the skills above are referenced from there. ## Reference | Topic | What it covers | | ------------------------------------------------------------ | ----------------------------------------------------------------------------------------------------------------------------------------- | | [Setup](./docs/setup.md) | Hardware/software prerequisites, `uv` install paths, CUDA variants, Docker base image, and base-checkpoint downloading. | | [Code Structure](./docs/code_structure.md) | Repository layout and a per-subpackage tour of `cosmos_framework/` — where each concern lives and where to add new code. | | [Training](./docs/training.md) | Launching multi-GPU and multi-node runs; parallelism strategies; mixed precision; resuming. | | [Inference (from a trained checkpoint)](./docs/inference.md) | Loading a trained checkpoint into one of the inference backends. | | [Policy Server](./docs/action_policy_droid_server.md) | Running the server-client pipeline for Cosmos3-Policy-DROID. | | [Post-training and RL](./docs/posttrain/index.md) | SFT, preference optimization, online RL, distillation, diffusion RL, embodied policies, and extension guides. | | [FAQ](./docs/faq.md) | Troubleshooting (OOM, NCCL hangs, slow training), environment variables, and common pitfalls. | | [AGENTS.md](./AGENTS.md) + [Agent Skills](#agent-skills) | Repo map and task-specific `SKILL.md` files loaded automatically by AGENTS.md-aware coding agents (Claude Code, Codex CLI, Cursor, etc.). |

Loading the repository README…

Skills in this repository

Software Developers

1 occupation category · 17% classified

skill

2026-10-21

cosmos3-codebase-nav

Software Developers

Navigate the Cosmos3 package codebase to find where parameters, configs, defaults, scripts, and documentation live. Use when the user asks "where is X in cosmos3", "how do I find the config for Y", "where are the defaults", "where do I change a parameter", or…

2026-07-28

Software Developers

2026-06-01

Showing 29 of 29 collected skills.