Cosmos3 Codebase Nav

Cosmos3 Codebase Navigation

When to use this skill

• Use this skill when an agent is navigating the Cosmos3 package
• Use this skill to answer "where is X", "how do I find the config for Y", or any file-location question
• Use this skill when the user opens or edits cosmos3 files and needs orientation

Path convention

All paths below are relative to this file's location (.claude/skills/cosmos3-codebase-nav/). The repo is laid out as:

• cosmos_framework/ — main training package (data, model, trainer, callbacks, checkpoint, utils, …).
• cosmos_framework/configs/base/experiment/ — vfm (generator) experiment SKUs referenced by [train.train_policy].experiment in the recipe TOMLs.
• cosmos_framework/configs/base/reasoner/experiment/ — vlm (reasoner) experiment SKUs.
• cosmos_framework/inference/ — inference subpackage (args, model, inference engine, defaults, Ray serving, common helpers).
• cosmos_framework/scripts/ — top-level entry-point scripts (train, inference, eval, export_model, convert_model_to_dcp, upsample_prompts, caption_from_video, captions_to_sft_jsonl, action_policy_server, …). Invoked as python -m cosmos_framework.scripts.<name>.
• examples/toml/sft_config/<recipe>.toml + examples/launch_sft_<recipe>.sh — paired SFT recipes (training entry-point input). The shell sources examples/_sft_launcher_common.sh, which forwards into cosmos_framework.scripts.train --sft-toml=....
• cosmos_framework/configs/toml_config/ — pydantic schemas (sft_config.py) and helpers that validate the recipe TOML at load time.

Quick Reference

Where parameters and defaults live

Available modality modes for defaults: text2image, text2video, image2video, image2image, video2video, forward_dynamics, inverse_dynamics, wam.

Config defaults resolution chain

When a user runs inference, default parameter values are resolved in this order:

cosmos_framework/inference/defaults/<mode>/sample_args.json     # 1. Per-modality JSON defaults (num_steps, guidance, shift, fps, etc.)
        ↓
_load_modality_defaults() in cosmos_framework/inference/args.py # 2. Loaded and cached at import time
        ↓
SamplingArgs / SamplingOverrides                      # 3. Pydantic models with field-level validation
        ↓
OmniSampleOverrides.build_sample()                    # 4. Merges user overrides → final resolved args
        ↓
_RESOLUTION_SHIFT_DEFAULTS[model_size, resolution]    # 5. Model+resolution shift override (if user didn't set shift)
        ↓
CLI flags (--guidance, --shift, etc.)                 # 6. User overrides from command line

The _RESOLUTION_SHIFT_DEFAULTS table in ../../../cosmos_framework/inference/args.py (on OmniSampleOverrides) overrides the default shift based on model size and resolution, unless the user explicitly specified --shift.

Action and video2video modes also have defaults under cosmos_framework/inference/defaults/{image2image,video2video,forward_dynamics,inverse_dynamics,policy}/sample_args.json.

Users can also supply a custom defaults file per-request via the defaults_file field in sample arguments (see ../../../docs/inference.md).

Where to make changes

Key entry points

Documentation