Skip to content

build: update Transformer Engine to release_v2.20.2-nemo tip - #6384

Closed
balasaajay wants to merge 1 commit into
NVIDIA-NeMo:mainfrom
balasaajay:build/te-2.20.2-nemo-20261009
Closed

balasaajay wants to merge 1 commit into
NVIDIA-NeMo:mainfrom
balasaajay:build/te-2.20.2-nemo-20261009

Conversation

@balasaajay

Copy link
Copy Markdown
Contributor

What does this PR do?

Pin Transformer Engine to 44f31bbeaf99bd30a67b928000d4b980026debd2, the tip of release_v2.20.2-nemo resolved on 2026-10-09. The effective TE source advances from 6ea2a74a to 44f31bbe, picking up the fused grouped MLP deterministic-dprob fix in NVIDIA/TransformerEngine#3635.

Changelog

  • Advance the MCore submodule from 441a987409b8fdd2b90b463f7999bac970109a43 to 4e7f45066b37a88060058ff012eeacb37c90f7d0, the signed companion NVIDIA/Megatron-LM#8036 revision. Bridge resolves TE through MCore and its dependency-layer check requires their locked TE sources to agree. This advance also includes the 15 MCore main commits between the old gitlink and the companion PR's base; merge the companion PR first.
  • Align Bridge's stale standalone TE source declaration with the effective MCore source, and set source metadata to 2.20.2+44f31bbe.
  • Regenerate uv.lock using uv 0.7.2 in the existing nvcr.io/nvidia/pytorch:26.08-py3 base. All 345 package names remain unchanged and no unrelated package versions move. The resolver also normalizes existing platform markers; forcing re-resolution on the unchanged base reproduces those same marker changes.

Validation

  • uv lock and uv lock --check in the supported PyTorch base container.
  • uv run --no-sync pre-commit run --all-files in a Bridge CI container.
  • .github/scripts/test_docker_dependency_layers.sh, including exact Bridge/MCore TE source equality.
  • Compare the regenerated base and update lock graphs: only the MCore TE requirement and TE package change.
  • git diff --check.
  • Native TE rebuild/import and H100/GB200 unit/functional CI. Local checks above do not establish GPU runtime correctness; needs-more-tests and full-test-suite request the expanded CI matrix.

Before ready for review

  • Followed the contributor guidelines; commit signed and signed off.
  • Existing dependency validation covers the pin/lock change; no new product tests or documentation are needed.
  • Review the companion MCore PR and final CI results before merging.

Signed-off-by: Ajay Balasa <abalasa@nvidia.com>
@balasaajay balasaajay added ci CI, automation, test queue, or workflow infrastructure work area:build Dependencies, packaging, images, and environment setup needs-more-tests Requires additional L0 and L1 test coverage before merge full-test-suite labels Oct 9, 2026
@copy-pr-bot

copy-pr-bot Bot commented Oct 9, 2026

Copy link
Copy Markdown

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

@balasaajay

Copy link
Copy Markdown
Contributor Author

/ok to test 3aed45c

This branch was successfully deployed

1 active deployment
test — 3aed45cb Deployed Oct 9, 2026 by copy-pr-bot[bot] via cicd-wait-in-queue #22699
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

area:build Dependencies, packaging, images, and environment setup ci CI, automation, test queue, or workflow infrastructure work full-test-suite needs-more-tests Requires additional L0 and L1 test coverage before merge

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant