Skip to content

feat: add native Cosmos workflow runtimes - #3

Open
ramanathan831 wants to merge 11 commits into
dev/ram/cosmos3-tao-reproducibilityfrom
dev/ram/cosmos-workflow-runtime
Open

ramanathan831 wants to merge 11 commits into
dev/ram/cosmos3-tao-reproducibilityfrom
dev/ram/cosmos-workflow-runtime

Conversation

@ramanathan831

@ramanathan831 ramanathan831 commented Sep 23, 2026 •

Copy link
Copy Markdown
Owner

Runtime-only scope

Native Cosmos Framework workflow capabilities, split out of #1. No skill entrypoints, workflow-support bundle, or skills CI are included.

Merge order: upstream parent #277, this runtime layer, then skills-only #1. The parent includes upstream main at cf5d68c; those upstream changes are outside this diff.

Framework ownership

  • Retain native dataset format conversion, checkpoint/LoRA helpers, reasoner inference/evaluation, annotation, quantization, and reasoner HTTP serving.
  • No alternate-runtime dependency, launcher, or converter bridge is required.
  • Shared training datasets, recipes, caching, and lifecycle integration now belong to the parent. This PR consumes them without recreating or replacing them.
  • Training-to-evaluation checkpoint handoffs use native Framework DCP export with merged PEFT weights and verified provenance.
  • Keep heavyweight workflow dependencies optional and quantization dependencies in an isolated image overlay.
  • Local image tags are build targets, not published images.
  • Cosmos-Embed and Predict 2.5 serving remain excluded. PAIDF Predict and all five execution-platform references remain in feat(skills): integrate Cosmos workflow capabilities #1.
  • Historical source attribution remains in the independently packaged runtime license/provenance resources; it is not a runtime dependency.

Focused history

80 changed files, 11 commits, down from 102 changed files before the parent consolidation:

  1. Native dataset format conversion.
  2. Dependency-free workflow lifecycle logging.
  3. Provenance-checked checkpoint/LoRA preparation.
  4. Native inference and evaluation.
  5. Video annotation pipeline and domain prompts.
  6. Quantization action and isolated overlay.
  7. Reasoner HTTP serving.
  8. Runtime packaging, deployment image, and tests.
  9. Installed-Framework inference execution fix.
  10. Annotation retry validation and terminal-error preservation.
  11. Framework license-header alignment.

Restacking preserves all runtime-only implementations and assertions, including the quantization-image test. Shared parent changes are inherited rather than duplicated in these commits.

Current validation at 393c4f8

  • 118 native runtime contract tests passed using the documented isolated test invocation.
  • 35 video recipe and Dockerfile/provenance tests passed with the repository test configuration.
  • Both original pre-commit configurations pass locally.
  • The parent separately passes 103 targeted regressions; the upper skills layer passes 557 CPU contracts and strict version stamps.
  • Both fresh hosted pre-commit runs passed on this head: run 36933483486, run 36933481514. The previous formatter-installation failure note is obsolete: the parent builds the genuine pinned release without changing the hook version.

No GPU jobs, container builds, model downloads, live serving, quantization runs, distributed execution, or paid API calls were launched for this cleanup. Existing tests and CI gates were not weakened.

The PR and source branch are unchanged. The history was restacked with explicit force-with-lease protection and local backup refs.

@ramanathan831
ramanathan831 force-pushed the dev/ram/cosmos-workflow-runtime branch 4 times, most recently from ea60695 to 9ca19bd Compare September 29, 2026 00:24
@ramanathan831
ramanathan831 force-pushed the dev/ram/cosmos3-tao-reproducibility branch from beaa459 to 31293c7 Compare September 29, 2026 00:59
@ramanathan831
ramanathan831 force-pushed the dev/ram/cosmos-workflow-runtime branch 8 times, most recently from 70ac74f to 6d7cc51 Compare October 1, 2026 19:20
@ramanathan831
ramanathan831 force-pushed the dev/ram/cosmos3-tao-reproducibility branch from c58ad84 to 0ec0e93 Compare October 1, 2026 21:29
@ramanathan831
ramanathan831 force-pushed the dev/ram/cosmos-workflow-runtime branch from 6d7cc51 to 99fa954 Compare October 1, 2026 21:29
@ramanathan831
ramanathan831 force-pushed the dev/ram/cosmos3-tao-reproducibility branch 2 times, most recently from 02b05df to 0ec0e93 Compare October 1, 2026 21:31
@ramanathan831
ramanathan831 force-pushed the dev/ram/cosmos-workflow-runtime branch from 99fa954 to 393c4f8 Compare October 1, 2026 22:11
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant