Repository navigation
build: AUT-2526 bump MCore main to 6911c751 - #6432
Closed
svcnemo-autobot wants to merge 2 commits into
Closed
svcnemo-autobot wants to merge 2 commits into
svcnemo-autobot wants to merge 2 commits into
Conversation
Bump the Megatron-LM main submodule and .main.commit from e0c5e186 to 6911c751. MCore's pin moves TransformerEngine to 44f31bbe (release_v2.20.2-nemo tip), so mirror its dependency metadata and regenerate uv.lock with uv 0.7.2. Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com>
svcnvidia-nemo-ci
added
area:build
Contributor
|
/ok to test ce1b60a |
Contributor
|
Automatic Claude reviews have been retired. To request a pull-request review, post a comment containing: Add |
Contributor
Bring in d8fe0b5 (NVIDIA-NeMo#6420, ci: use writable config locks for offline test containers) so this PR's CI uses a writable MEGATRON_CONFIG_LOCK_DIR. The merged tree is identical to the locally validated merge(ce1b60a, a2e8061). Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com>
Contributor
|
/ok to test 421fa9d |
Contributor
|
Closing as superseded by #6438, which is already merged. Its MCore pin a98aed553e25b4e86c59f50be54ac0ea4227b442 includes 6911c751fdab0b9ef55eb3d4efa22f294aa4682c and is seven commits newer. Current main also already contains the required TransformerEngine 2.20.2+44f31bbe metadata and lockfile pin. |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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 of the
mainlane. This moves the3rdparty/Megatron-LMsubmodule and.main.commitfrome0c5e186to6911c751..dev.commitis unchanged (4a4de865).What changed
.main.commitand the3rdparty/Megatron-LMgitlink now point at6911c751fdab0b9ef55eb3d4efa22f294aa4682c.pyproject.toml: the TransformerEngine dependency metadata matches MCore's pin,2.20.2+44f31bbe(tip ofrelease_v2.20.2-nemo).uv.lockregenerated withuv lock --upgradeusing uv 0.7.2, which matchesUV_VERSIONindocker/Dockerfile.ci.uv lock --checkpasses.Details
MCore commits in the bump:
megatron/core/inference.The TransformerEngine delta is a single Python-only change to the SM100 CuTe fused grouped MLP, which H100 never reaches. The other lock bumps are patch-level transitive dependencies.
Tests were selected from this impact: all Core unit shards, plus
L1_Launch_vlm_inference, which exercises SamplingParams and TextGenerationController. The diffusion unit tests and the other H100 functional launchers were excluded because nothing in them uses the changed modules.Tested
Local H100 runs on
merge(ce1b60ab, main@a2e80617), using an image built from that merge with the MCore pin verified:Launch_Unit_Tests_Coreshard 0/3: 4022 passedLaunch_Unit_Tests_Coreshard 1/3: 3995 passedLaunch_Unit_Tests_Coreshard 2/3: 4010 passedL1_Launch_vlm_inference: passedShard 1/3 first stalled in FlashAttention-4 backward in
test_real_packed_multimodal_optimizer_step. The local profile did not apply the FlashAttention policy that CI sets in.github/actions/test-template/action.yml(NVTE_FLASH_ATTN_V2=1,V3=0,V4=0). With those three variables set, the shard passed, including all 20 nemotron_omni tests. The other three runs used the profile's default FlashAttention auto-selection rather than CI's FA2 pin.The bump does not change MCore's attention code. Main CI on this head is pending.