Cosmos3 Codebase Navigation

SkillFiles & storage

Use when navigating the Cosmos3 integration in NPA or locating upstream Cosmos3 framework files, defaults, scripts, configs, recipes, and docs.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Cosmos3 Codebase Navigation skill

What this skill tells your AI

The instructions your AI receives, as published by nebius/nebius-physical-ai in skills/atomic/cosmos3-codebase-nav/SKILL.md and read by ahel’s review.

Source And Attribution

Adapted from NVIDIA cosmos-framework skills/atomic/cosmos3-codebase-nav/SKILL.md.

Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. Used under OpenMDW-1.1. See skills/LICENSE-NVIDIA-COSMOS3-OPENMDW-1.1 and skills/NOTICE-NVIDIA-COSMOS3.

When To Use

Use this skill when the user asks where Cosmos3 setup, access checks, inference defaults, parser settings, workflows, or post-training configs live. Also use it when reviewing or modifying PRs that touch Cosmos3 support in NPA.

NPA Integration Map

Primary NPA files:

NeedFile
Cosmos3 access, fetch, redaction, parser config, inference arg buildernpa/src/npa/workbench/cosmos/cosmos3.py
Real Cosmos 3 generation runner (CLI/SDK/YAML share it)npa/src/npa/workbench/cosmos/generate.py
Containerized runtime (framework pinned, no weights baked)npa/docker/workbench/cosmos3/Dockerfile
Build-time proof the generate deps resolve for every modenpa/docker/workbench/cosmos3/verify_env.py
Generate CLI / SDKnpa/src/npa/cli/workbench/cosmos3.py, npa/src/npa/sdk/workbench/cosmos3.py
Nano vLLM-Omni continuation and source-conditioned augmentationnpa/src/npa/workbench/cosmos/nano_video.py, npa/src/npa/workbench/cosmos/nano_video_augment.py
Nano augmentation artifact handoff and recoverynpa/src/npa/workbench/cosmos/nano_video_augment_client.py
Nano diffusion serving, routing and measured deployment recipenpa/src/npa/workbench/cosmos/nano_video_server.py, npa/src/npa/workbench/cosmos/nano_video_router.py, npa/deploy/cosmos3-nano-video/README.md
Generate workflowworkflows/testing/cosmos3-generate.yaml
Human CLI commands for real workflowsnpa/src/npa/cli/cosmos/__init__.py
SDK compatibility exportsnpa/src/npa/workbench/cosmos/__init__.py, npa/src/npa/sdk/workbench/cosmos.py
Fetch workflowworkflows/testing/cosmos-fetch.yaml
Text-to-image H100 smoke workflowworkflows/testing/cosmos3-text-to-image.yaml
Unit tests for access, fetch, inference YAML, and skill artifactsnpa/tests/workbench/test_cosmos3_access.py
CLI tests for check and fetchnpa/tests/cli/test_cosmos3_cli.py

The agent-skill artifacts live in this repository under:

  • skills/atomic/cosmos3-setup/SKILL.md
  • skills/atomic/cosmos3-codebase-nav/SKILL.md
  • skills/atomic/cosmos3-env-troubleshoot/SKILL.md
  • skills/workflows/cosmos3-inference/SKILL.md
  • skills/workflows/cosmos3-post-training/SKILL.md

What Not To Look For

There is no supported Cosmos skill-display subcommand. Cosmos3 skills are not platform commands, job templates, or SDK callables. They are SKILL.md files for coding agents to read.

The retained Cosmos3 workflow YAMLs are real executable specs:

  • workflows/testing/cosmos-fetch.yaml
  • workflows/testing/cosmos3-generate.yaml
  • workflows/testing/cosmos3-text-to-image.yaml

Do not restore cosmos3-setup.yaml, cosmos3-codebase-nav.yaml, cosmos3-env-troubleshoot.yaml, cosmos3-post-training.yaml, or retired raw SkyPilot task templates as skill launchers unless the product direction changes to a real executable workflow.

Upstream Cosmos Framework Map

When a task needs upstream package internals, inspect a clone of https://github.com/NVIDIA/cosmos-framework.git.

Common upstream paths:

NeedUpstream path
Sampling args, setup args, override modelscosmos_framework/inference/args.py
Per-modality defaultscosmos_framework/inference/defaults/<mode>/sample_args.json
Inference scriptcosmos_framework/scripts/inference.py
Training scriptcosmos_framework/scripts/train.py
Ray serving presetscosmos_framework/inference/ray/configs/latency.yaml, throughput.yaml
SFT recipe TOMLsexamples/toml/sft_config/<recipe>.toml
SFT launch shellsexamples/launch_sft_<recipe>.sh
SFT schemacosmos_framework/configs/toml_config/sft_config.py
Setup docsdocs/setup.md
Inference docsdocs/inference.md
Training docsdocs/training.md

Navigation Routine

  1. Start with rg in NPA:

    rg -n "NPA_COSMOS3|Cosmos3|cosmos3-text-to-image|guardrails" npa
    
  2. If the answer depends on upstream behavior, inspect the upstream file rather than guessing from NPA wrappers.

    For the Nano diffusion deployment, inspect the installed vLLM-Omni video parser and Cosmos3OmniDiffusersPipeline; the framework inference argument map does not establish what that image accepts. The deployment README links the pinned implementation and records the effective transfer contract.

  3. Keep NPA and upstream responsibilities separate. NPA owns CLI/SDK wrappers, SkyPilot workflows, image defaults, and tests. NVIDIA's repo owns Cosmos3 framework semantics, inference arguments, recipes, and training internals.

  4. When adding docs or tests, reference SKILL.md artifacts as repository files, not as commands.

Signals

GitHub stars
28
Forks
15
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
cosmos3-codebase-nav
Source
github.com/nebius/nebius-physical-ai