[ishovkun]

@coderabbitai summary

Description

Cosmos3 checkpoints spread their settings over several files (model_index.json, the transformer
and VAE config.json, checkpoint.json, the scheduler config), and TensorRT-LLM read them from
a dozen places in the pipeline: some keys from one file, some from another, some with a hard-coded
fallback, with no record of which value won. The released checkpoints disagree with each other on
a few keys (system prompt, native flow schedule), and a silent fallback is how a wrong setting
goes unnoticed.

This PR puts every checkpoint- and request-level setting behind one precedence table, and
makes it the only path: the executor defaults, warmup and infer() all resolve through it.

Checkpoint settings. Cosmos3ConfigResolver(checkpoint_dir).resolve() walks a rule table.
Each rule names the sources for one setting in order of precedence (a key in a checkpoint file,
or an explicit default with a reason), and says what to do when the sources disagree: error,
warn, or ignore. The result is a Cosmos3Config with
sections (architecture, vae, audio, sampling, prompt, action); every value carries its source,
and describe() logs the table at load. A setting with no source and no declared default is an
error naming the file and path, not a guess. Only the four transformer compatibility keys and
native_flow_schedule=False have declared defaults; prompt.use_system_prompt and
vae.spatial_compression are required. The transformer's compatibility defaults now come from
the same table (fill_pretrained_config), so they are declared once.

Request settings. request_resolver.py does the same for per-request values (resolution,
frames, frame rate, steps, guidance, action mode, domain, action width): a request value wins,
then the checkpoint's distilled or policy contract, then the checkpoint's embodiment table, then
the family's tables. Disagreements with a recommendation warn once per request. A golden fixture
pins the resolved defaults for every published checkpoint so a change is visible in review.

Pipeline. pipeline_cosmos3.py reads self.cosmos3_config and the request resolver instead
of its own key lookups; default_generation_params, warmup and infer() share that path, so
the sampling policy's separate default overrides and the pipeline's own VAE/transformer
compression check are gone (the resolver's rules cover both). The constants that were defined
in two or three places (V2V flow shift, prompt format strings, aspect-ratio buckets, resolution
table) are defined once; the output geometry has its own leaf module, geometry.py.

Behaviour changes for shipped checkpoints: none intended. The golden fixtures were generated
from the current behaviour and reviewed by hand. The deliberate differences: a checkpoint
declaring neither prompt key now fails to load instead of silently running without a system
prompt, and a checkpoint without vae/config.json fails instead of assuming a spatial
compression of 16.

The rules for the autoregressive (Sim) checkpoints, their serving-config overrides and the
checkpoint embodiment table come with the Sim pipeline PR; this PR carries only the settings the
current pipeline consumes.

Test Coverage

• tests/unittest/_torch/visual_gen/test_cosmos3_config_resolver.py (CPU): all 10 published
  checkpoints resolve (test_data/cosmos3_checkpoint_configs.json); conflict policies; required
  keys; source recording; the transformer compatibility defaults.
• tests/unittest/_torch/visual_gen/test_cosmos3_request_resolver.py (CPU): golden request
  defaults per checkpoint, mode and workflow (test_data/cosmos3_request_defaults.json);
  override precedence; disagreement warnings.
• Existing test_cosmos3_edge.py, test_cosmos3_pipeline.py, test_cosmos3_distilled.py,
  test_cosmos3_action.py, test_cosmos3_transfer.py, test_cosmos3_example_prompts.py,
  test_cosmos3_transformer.py run on the new entry points (GPU).
• Test lists: the two CPU tests are added to l0_cpu.yml; the GPU suites are already listed.

PR Checklist

Please review the following before submitting your PR:

• PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.

• PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.

• Test cases are provided for new code paths (see test instructions)

• If PR introduces API changes, an appropriate PR label is added - either api-compatible or api-breaking. For api-breaking, include BREAKING in the PR title.

• Any new dependencies have been scanned for license and vulnerabilities

• CODEOWNERS updated if ownership changes

• Documentation updated as needed

• Update tava architecture diagram if there is a significant design change in PR.

• The reviewers assigned automatically/manually are appropriate for the PR.

• Please check this after reviewing the above items as appropriate for this PR.

GitHub Bot Help

To see a list of available CI bot commands, please comment /bot help.

[ishovkun]

/bot run --disable-fail-fast

[ishovkun]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #81788 [ run ] triggered by Bot. Commit: 60c8832 Link to invocation

[ishovkun]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #81791 [ run ] triggered by Bot. Commit: f8d7739 Link to invocation

[tensorrt-cicd]

PR_Github #81788 [ run ] completed with state ABORTED. Commit: 60c8832

Link to invocation

[tensorrt-cicd]

PR_Github #81791 [ run ] completed with state FAILURE. Commit: f8d7739
/LLM/main/L0_MergeRequest_PR pipeline #67448 completed with status: 'UNSTABLE'

CI Report

⚠️ Multi-GPU Label Required:
Multi-GPU tests require the ci: full pre-merge approved label on this PR. Either:

• Wait for the PR to be fully approved — the label is added automatically once approval is complete. Having unresolved open comments is fine, or
• If needed, ask a member of NVIDIA/trt-llm-ci-approvers to add the label manually.
  Then re-trigger CI with the same bot command (no rebase needed).

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

Link to invocation

[ishovkun]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #81868 [ run ] triggered by Bot. Commit: e27621e Link to invocation

[tensorrt-cicd]

PR_Github #81868 [ run ] completed with state FAILURE. Commit: e27621e
/LLM/main/L0_MergeRequest_PR pipeline #67519 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation