Siddharth/fix ep sync by wdykas · Pull Request #4607 · NVIDIA/Megatron-LM

wdykas · 2026-05-04T17:03:06Z

What does this PR do ?

⚠️ For major changes (either in lines of code or in its impact), please make sure to first share a design doc with the team. If you're unsure what's the best way to do so, contact the @mcore-oncall.

Issue tracking

For PRs from open-source community contributors:

New features: a linked issue is required. Please open a feature request and reference it here before submitting the PR.
Small updates (bug fixes, minor improvements): a linked issue is recommended and will accelerate the PR review process.

Linked issue:

Contribution process

Pre-checks

I have added relevant unit tests
I have added relevant functional tests
I have added proper typing to my code Typing guidelines
I have added relevant documentation
I have run the autoformatter.sh on my PR

Code review

Feel free to message or comment the @mcore-oncall to help accelerate your merge into main. The less complex your PR is, the faster it will be approved and merged!

All PRs start as draft. If you open a non-draft PR, it will be automatically converted to draft.

Step 1: Mark PR as "Ready for Review"

When your PR is ready, click Ready for Review.
An oncall reviewer is auto-assigned and expert reviewers are notified based on your changes.
- Some PRs may jump straight to step 2. This is determined by .github/CODEOWNERS.

⚠️ Only mark as ready once merge-conflicts are resolved and the CI is passing.
Final Review might get declined if these requirements are not fulfilled.

Step 2: Final Review

For PRs that change megatron/core, once all expert reviewers have approved, the Final Review label is applied automatically and final reviewers are assigned.

For PRs outside megatron/core, this step is skipped.

Step 3: Approved

Once all required reviewers have approved, the Approved label is applied automatically.

Merge

Any member of mcore-engineers will be able to merge your PR.

For MRs into `dev` branch

The proposed review process for `dev` branch is under active discussion.

MRs are mergable after one approval by either eharper@nvidia.com or zijiey@nvidia.com.

The every-20-step consensus optimization caches `_last_ep_consensus`, but the cache is never invalidated on state transitions. After PAUSING -> PAUSED the cache holds (global_work>0, all_pausing=True); when UNPAUSING -> RUNNING the next loop iteration skips a fresh consensus (counter % 20 != 0 and global_work != 0), reads the stale all_pausing=True, and immediately re-pauses the engine. This wedges test_control_logic_lifecycle[tp2-pp2-ep2] at the post-unpause asyncio.gather (10s timeout). Reset the cache to (0, False) on UNPAUSING -> RUNNING so the first RUNNING iteration always re-runs consensus. Also add `--inference-disable-ep-consensus` (DynamicInferenceContextConfig .disable_ep_consensus) as an escape hatch: when set, the control loop skips _ep_establish_consensus and dummy_forward entirely and steps on local state. Safe only when EP coordination is not required. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

copy-pr-bot · 2026-05-04T17:03:10Z

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

github-actions · 2026-05-04T17:03:18Z

This PR has been automatically converted to draft because all PRs must start as drafts.

When you are ready for review, click Ready for Review to begin the review process. This will:

Add the oncall reviewer (optional reviewer)
Add required review teams based on your changes

See the contribution guide for more details.

wdykas · 2026-05-04T17:04:10Z

/ok to test 9a741d8

The disable_ep_consensus branch previously slept for 20ms when local_pending == 0. With EP > 1 a peer running a real forward blocks at its NCCL all-to-all kernel waiting for this rank's collective participation, so sleeping here deadlocks the EP group. Mirror the consensus path: when there is no local work, run controller. dummy_forward() instead. Trades the consensus all-reduce CPU cost for unconditional dummy forwards on idle ranks. Add `await asyncio.sleep(0)` after dummy_forward so the engine task yields to the asyncio loop — the consensus path naturally yields via _ep_establish_consensus and async_step, but the disable path otherwise has no awaits between steps and starves signal delivery. Add test_disable_ep_consensus that wraps controller.dummy_forward in a MagicMock spy and asserts the call count grows during a 200 ms idle window. Also exercises request submission and a pause/unpause cycle. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

wdykas · 2026-05-05T00:48:45Z

/ok to test 9637e6b

wdykas · 2026-05-05T01:37:40Z

/ok to test a8128e8

…ix-ep-sync # Conflicts: # megatron/inference/utils.py

wdykas · 2026-05-05T12:50:04Z

/ok to test ffd9ca5

tdene · 2026-05-05T15:13:50Z

For future work, I still believe we should move the EP sync to a side-task in the asyncio event loop - so that it runs in parallel with the forward step.

wdykas · 2026-05-05T16:08:01Z

/ok to test 478ed44

wdykas · 2026-05-05T17:38:26Z

/ok to test bca15ba

svcnvidia-nemo-ci · 2026-05-05T20:49:53Z

🔄 Merge queue validation started!

You can track the progress here: https://github.com/NVIDIA/Megatron-LM/actions/runs/25401534202

…ync-2026-05 Upstream tip: 932d9ee (Add periodic GPU sniff tests, 2026-05-07). Conflict resolution: - 13 modify/delete conflicts in .github/* and skills/*: kept deletion (matches policy from c037427 removing upstream-NVIDIA CI machinery not relevant to this fork). - AGENTS.md content conflict: kept fork version (skills/contributing sections from upstream don't apply since the directories were removed). - New upstream additions matching the same removal policy were dropped: .claude/settings.json, .github/workflows/nightly-sync-main-to-dev.yml, skills/{cicd,linting-and-formatting,nightly-sync,run-on-slurm,testing, update-golden-values}/SKILL.md. - SECURITY.md (new from upstream): kept (generic security policy). Verified post-merge: - pretrain_gpt.py logging-patch hook intact (lines 405-408). - megatron/core/optimizer/__init__.py emerging_optimizers + muon logic intact. - megatron/training/arguments.py custom muon flags intact (lines 2297-2309). Notable upstream changes pulled in: - Removed legacy transformer + legacy GPT modules (NVIDIA#4207, NVIDIA#4322). - Docker base image bump to 26.04-py3 (NVIDIA#4611). - Inference fixes (vLLM grouped GEMM NVIDIA#4566, FlashInfer sampling NVIDIA#2456, EP sync NVIDIA#4607, MoE dispatcher fixes NVIDIA#4576). - Gradient corruption fix with layerwise param all-gather overlap (NVIDIA#4609). - Hybrid model + Flextron + GPU sniff tests + named validation sets + fault injection + InJob restart.

Co-authored-by: Siddharth Singh <sidsingh@nvidia.com> Co-authored-by: root <root@nvl72098-T11.cm.cluster> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> Co-authored-by: root <root@nvl72078-T18.cm.cluster> Co-authored-by: William Dykas <wdykas@oci-hsg-cs-001-vscode-03.cm.cluster> Co-authored-by: root <root@nvl72102-T05.cm.cluster> Signed-off-by: yhgalaxy <yhgalaxy@outlook.com>

Co-authored-by: Siddharth Singh <sidsingh@nvidia.com> Co-authored-by: root <root@nvl72098-T11.cm.cluster> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> Co-authored-by: root <root@nvl72078-T18.cm.cluster> Co-authored-by: William Dykas <wdykas@oci-hsg-cs-001-vscode-03.cm.cluster> Co-authored-by: root <root@nvl72102-T05.cm.cluster> Signed-off-by: Jon Barker <jbarker@aws-cmh-slurm-1-vscode-02.cm.cluster>

Co-authored-by: Siddharth Singh <sidsingh@nvidia.com> Co-authored-by: root <root@nvl72098-T11.cm.cluster> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> Co-authored-by: root <root@nvl72078-T18.cm.cluster> Co-authored-by: William Dykas <wdykas@oci-hsg-cs-001-vscode-03.cm.cluster> Co-authored-by: root <root@nvl72102-T05.cm.cluster>

Co-authored-by: Siddharth Singh <sidsingh@nvidia.com> Co-authored-by: root <root@nvl72098-T11.cm.cluster> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> Co-authored-by: root <root@nvl72078-T18.cm.cluster> Co-authored-by: William Dykas <wdykas@oci-hsg-cs-001-vscode-03.cm.cluster> Co-authored-by: root <root@nvl72102-T05.cm.cluster> Signed-off-by: Dmytro Pykhtar <dpykhtar@nvidia.com>

sidsingh-nvidia and others added 6 commits April 24, 2026 11:41

simplify TP broadcast

3dd1bc6

only run establish consensus every 20 engine steps

0647385

format

ff1120b

Merge branch 'main' into siddharth/fix-ep-sync

5dd57cd

Merge branch 'main' into siddharth/fix-ep-sync

d161240

wdykas requested review from a team as code owners May 4, 2026 17:03

svcnvidia-nemo-ci marked this pull request as draft May 4, 2026 17:03

copy-pr-bot Bot temporarily deployed to test May 4, 2026 17:05 Inactive

sidsingh-nvidia reviewed May 4, 2026

View reviewed changes

Comment thread megatron/core/inference/engines/dynamic_engine.py

fix

a8128e8

copy-pr-bot Bot temporarily deployed to test May 5, 2026 01:38 Inactive

Merge branch 'main' of github.com:wdykas/Megatron-LM into siddharth/f…

ffd9ca5

…ix-ep-sync # Conflicts: # megatron/inference/utils.py

wdykas marked this pull request as ready for review May 5, 2026 12:49

svcnvidia-nemo-ci requested a review from a team May 5, 2026 12:49

copy-pr-bot Bot temporarily deployed to test May 5, 2026 12:51 Inactive

svcnvidia-nemo-ci added the complexity: low label May 5, 2026

sidsingh-nvidia approved these changes May 5, 2026

View reviewed changes

Comment thread megatron/core/inference/engines/dynamic_engine.py

svcnvidia-nemo-ci added the Final Review PR is in the "final review" stage label May 5, 2026

tdene approved these changes May 5, 2026

View reviewed changes

kvareddy approved these changes May 5, 2026

View reviewed changes

shanmugamr1992 approved these changes May 5, 2026

View reviewed changes

svcnvidia-nemo-ci added Approved All necessary approvals have been made and removed Final Review PR is in the "final review" stage labels May 5, 2026

Merge branch 'main' into siddharth/fix-ep-sync

478ed44

wdykas enabled auto-merge May 5, 2026 16:07

copy-pr-bot Bot had a problem deploying to test May 5, 2026 16:09 Error

shanmugamr1992 approved these changes May 5, 2026

View reviewed changes

Merge branch 'main' into siddharth/fix-ep-sync

bca15ba

copy-pr-bot Bot temporarily deployed to test May 5, 2026 17:39 Inactive

wdykas added this pull request to the merge queue May 5, 2026

Merged via the queue into NVIDIA:main with commit 190c833 May 5, 2026
66 of 67 checks passed

wdykas deleted the siddharth/fix-ep-sync branch May 5, 2026 22:14

Navigation Menu

Uh oh!

Siddharth/fix ep sync - #4607

Siddharth/fix ep sync#4607
wdykas merged 11 commits into
NVIDIA:mainfrom
wdykas:siddharth/fix-ep-sync

wdykas commented May 4, 2026

Uh oh!

copy-pr-bot Bot commented May 4, 2026

Uh oh!

github-actions Bot commented May 4, 2026

Uh oh!

wdykas commented May 4, 2026

Uh oh!

Uh oh!

wdykas commented May 5, 2026

Uh oh!

wdykas commented May 5, 2026

Uh oh!

wdykas commented May 5, 2026

Uh oh!

Uh oh!

tdene commented May 5, 2026

Uh oh!

wdykas commented May 5, 2026

Uh oh!

wdykas commented May 5, 2026

Uh oh!

svcnvidia-nemo-ci commented May 5, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

Uh oh!

Conversation

wdykas commented May 4, 2026

What does this PR do ?

Issue tracking

Contribution process

Pre-checks

Code review

Step 1: Mark PR as "Ready for Review"

Step 2: Final Review

Step 3: Approved

Merge

Uh oh!

copy-pr-bot Bot commented May 4, 2026

Uh oh!

github-actions Bot commented May 4, 2026

Uh oh!

wdykas commented May 4, 2026

Uh oh!

Uh oh!

wdykas commented May 5, 2026

Uh oh!

wdykas commented May 5, 2026

Uh oh!

wdykas commented May 5, 2026

Uh oh!

Uh oh!

tdene commented May 5, 2026

Uh oh!

wdykas commented May 5, 2026

Uh oh!

wdykas commented May 5, 2026

Uh oh!

svcnvidia-nemo-ci commented May 5, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants