Repository navigation
Conversation
Update .main.commit and the 3rdparty/Megatron-LM gitlink from e0c5e18673bf4d4fd89cf87c4f1afbe521fe297f to e30d863a6d82a3c45f2d71355dc73fbf5610967b, align the TransformerEngine dependency metadata version with MCore (2.20.2+44f31bbe), and refresh uv.lock with uv lock --upgrade. Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com>
Requester-authorized pin revision: update .main.commit and the 3rdparty/Megatron-LM gitlink from e30d863a6d82a3c45f2d71355dc73fbf5610967b to a98aed553e25b4e86c59f50be54ac0ea4227b442 (NVIDIA/Megatron-LM main). MCore dependency metadata is unchanged; uv lock --check passes. .dev.commit unchanged. Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com>
…-261012-main Upstream main landed NVIDIA-NeMo#6414 (MCore main e30d863a). Resolve the .main.commit conflict to the session pin a98aed553e25b4e86c59f50be54ac0ea4227b442; the 3rdparty/Megatron-LM gitlink fast-forwards to the same commit. Keep the session uv lock --upgrade output (pyproject.toml is identical on both sides); uv lock --check passes on the merged tree. Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com>
L0_Launch_models_bert.sh hardcoded /opt/Megatron-Bridge for the coverage data file and source root. When the suite runs from a checkout other than /opt/Megatron-Bridge (or that path is read-only), coverage cannot write its data file and the row fails before pytest results are reported. Resolve REPO_ROOT from the script location and use it for --data-file and --source. Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com>
svcnvidia-nemo-ci
added
area:build
Contributor
|
/ok to test 33db331 |
Contributor
|
Automatic Claude reviews have been retired. To request a pull-request review, post a comment containing: Add |
…main-261012-a98aed55 Keep the MCore main pin a98aed55; main's 8c815a4c is its ancestor. Take main's uv.lock; uv lock --check passes for the merged tree. Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com>
Contributor
|
/ok to test 52c1a6d |
ko3n1g
approved these changes
Oct 13, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Background
Nightly Megatron-Core bump for the
mainlane, moving Bridge's MCore pin toa98aed55. Bridgemainhas since moved its own pin to8c815a4c, an ancestor ofa98aed55, and upstreammainwas merged in at1edea5bc.What changed
.main.commitand the3rdparty/Megatron-LMgitlink now point toa98aed553e25b4e86c59f50be54ac0ea4227b442..dev.commitis unchanged.uv.lockmatchesmain.uv lock --checkpasses with uv 0.7.2, thedocker/Dockerfile.cipin. MCore's dependency metadata is unchanged ata98aed55.L0_Launch_models_bert.shtakes its coverage--data-file/--sourcepaths from the repo root instead of a hardcoded/opt/Megatron-Bridge.Details
Tests were selected by tracing the MCore changes since
e30d863a: distributed-optimizer checkpoint layout, deterministic GDN l2norm, empty hybrid PP stages, and inference stop/sampling. That gave 77 rows: 4 unit rows (Core in 3 shards, plus Diffusion) and 73 H100 functional launchers. After themainmerge, 7 rows were re-run to cover its CUDA-graph validation/teardown and nemotronh perf-recipe changes.Tested
Run locally on H100 with the CI FlashAttention env (
NVTE_FLASH_ATTN_V2=1, V3/V4 off).52c1a6dewith imagesha256:68b61252df0c1a0eb6277494c99c2031fd0d042f08d82ae9864d8f7cef8d7e20, 7 rows passed with 0 failures: Core unit shards 0-2 (4045, 4017 and 4043 tests),L0_Launch_megatron_fsdp,L0_Launch_training_pretrain,L0_Launch_recipes_perf_configandL2_Launch_recipes_llama_cuda_graphs.sha256:65858877. Their results still apply: the MCore pin, Dockerfile and pyproject are unchanged, and no runtime code imports the packages bumped in the lock.L0_Launch_recipes_deepseek.shandL1_Launch_recipes_glm_vl.sh. Upstream CI skips them too.