AGENTS.md

Read this file first — it is the canonical map for navigating the Cosmos repository and stays up to date.

Cosmos is a framework for training and serving world foundation models. Everything lives in a single top-level cosmos_framework/ Python package:

• Training infrastructure — top-level subpackages under cosmos_framework/ (data, model, trainer, callbacks, checkpoint, …).
• Inference infrastructure — cosmos_framework/inference/ (Diffusers / Transformers / vLLM-friendly inference core, online serving via Ray + Gradio).
• Backend packages — packages/{diffusers,transformers,vllm}-cosmos3/ provide library-style shims that load Cosmos3 checkpoints into the respective ecosystems.
• Entry-point scripts — cosmos_framework/scripts/ (train.py, inference.py, export_model.py, …) invoked as python -m cosmos_framework.scripts.<name>. Primary training entry point: cosmos_framework.scripts.train driven by a structured, pydantic-validated TOML interface (--sft-toml=<recipe-toml>); the schema lives at cosmos_framework/configs/toml_config/sft_config.py and the canonical recipe pattern is documented in examples/README.md.

All paths below are relative to the repository root (the directory containing pyproject.toml, the cosmos_framework/ Python package, and packages/).

Commands

Task	Command
Lint	uv run ruff check .
Format check	uv run ruff format --check .
Auto-fix lint + format	uv run ruff check --fix . && uv run ruff format .
Type-check	uv run pyrefly check
Test (all)	uv run pytest
Test (single file)	uv run pytest --capture=no <path>

Config files: .ruff.toml (ruff), pyrefly.toml (pyrefly), .pytest.toml (pytest), conftest.py (pytest fixtures).

A justfile is provided at the root with longer recipes (just install, just lint, just test, just docker-cu130).

Rules

• Always answer questions with references to code or documentation in file:line format.
• When unsure, point the user to the closest doc rather than guessing.
• Keep this file short. Link out to skills and docs for detail — this file is included in every prompt.
• Inference code belongs under cosmos_framework/inference/; training infrastructure belongs under the other cosmos_framework/ subpackages. Don't blur the two — if you find yourself adding training-time imports inside cosmos_framework/inference/ (or vice versa), reconsider.

Key File Locations

Training (cosmos_framework/)

What	Where
Algorithms (losses, RL, reward)	cosmos_framework/algorithm/{loss,reward,rl}
Training loop	cosmos_framework/trainer/
Models + parallelism	cosmos_framework/model/
Datasets / data loading	cosmos_framework/data/
Checkpoint I/O	cosmos_framework/checkpoint/
Callbacks (logging, eval)	cosmos_framework/callbacks/
RL workers (rollout, reward, reference, simulations)	cosmos_framework/workers/
Controller / orchestrator	cosmos_framework/controller/
Launchers (Slurm, torchrun, k8s)	cosmos_framework/launcher/
Evaluation harness	cosmos_framework/evaluation/
CLI tools	cosmos_framework/tools/, tools/ (repo root)

For a per-subpackage tour with descriptions, see docs/code_structure.md.

Inference (cosmos_framework/inference/)

What	Where
CLI entry point	cosmos_framework/scripts/inference.py
Args / param definitions	cosmos_framework/inference/args.py
Per-modality defaults	cosmos_framework/inference/defaults/<mode>/sample_args.json
Model / inference core	cosmos_framework/inference/model.py, cosmos_framework/inference/inference.py
Ray serving	cosmos_framework/inference/ray/
Backend packages	packages/{diffusers,transformers,vllm}-cosmos3/
Example inputs	inputs/omni/*.json, inputs/reasoner/*.json

Documentation

Doc	What it covers
docs/setup.md	Install, NGC base image, CUDA variants, base-checkpoint download.
docs/code_structure.md	Repo layout and per-subpackage tour of cosmos_framework/.
docs/training.md	Single- and multi-node launches, parallelism, mixed precision.
docs/inference.md	Sample arguments, parallelism, schemas, troubleshooting.
docs/faq.md	Troubleshooting (OOM, NCCL, slow training) + env vars.

Agent skills (codebase navigation, env troubleshooting, inference, post-training, setup) live in .agents/skills/ and .claude/skills/.

Common Tasks

Training

Task	Command
Single-GPU train (smoke)	python -m cosmos_framework.scripts.train --sft-toml=examples/toml/sft_config/<recipe>.toml
Multi-GPU train	IMAGINAIRE_OUTPUT_ROOT=outputs/train torchrun --nproc-per-node=8 -m cosmos_framework.scripts.train --sft-toml=examples/toml/sft_config/<recipe>.toml
Resume from checkpoint	Re-run the same train --sft-toml=<recipe>.toml against the same IMAGINAIRE_OUTPUT_ROOT (auto-resume from latest DCP).
Export DCP → HF	python -m cosmos_framework.scripts.export_model --src <dcp> --dst <hf>
Run a config sweep	just run python -m cosmos_framework.scripts.train --sft-toml=examples/toml/sft_config/<recipe>.toml -- key.path=value ...

Inference

Task	Command
Single-GPU inference	python -m cosmos_framework.scripts.inference -i inputs/omni/t2v.json -o outputs/ --checkpoint-path Cosmos3-Nano
Multi-GPU inference	torchrun --nproc-per-node=4 -m cosmos_framework.scripts.inference --parallelism-preset=latency -i ... -o outputs/ ...
Start online Ray server	python -m cosmos_framework.inference.ray.serve --parallelism-preset=latency -o outputs/ray_serve --checkpoint-path Cosmos3-Nano
Launch Gradio UI	python -m cosmos_framework.inference.ray.gradio --port=8080
See all CLI flags	python -m cosmos_framework.scripts.inference --help

Gotchas

• NGC / PyTorch containers: run export LD_LIBRARY_PATH='' before any python call or you'll hit a torch._C import error. See docs/setup.md.
• Reproducibility: always pass --seed <int>. Without it a random seed is used each run.
• JSON paths: relative paths inside input JSON files resolve relative to the JSON file's directory, not the working directory.
• Resume: re-running the same inference command skips already-generated outputs automatically.
• Separation of concerns: keep training-time imports out of cosmos_framework/inference/, and keep heavyweight inference-only deps (vLLM, Ray Serve, Gradio) gated behind optional extras so plain training installs stay slim.