Add ag group expert parallelism

[jeffnvidia]

What does this PR do ?

This PR extends the all-gather/reduce-scatter overlap optimization to support Expert Parallelism (EP), enabling communication overlap for MoE (Mixture-of-Experts) models. It creates a separate process group for expert all-gather operations to avoid head-of-line blocking with gradient reduce-scatter, following the same pattern as the base implementation for regular data parallelism.

Background

This PR builds on #2663 (add all-gather process-group for overlapping), which implemented AG/RS overlap for regular data parallelism. That PR explicitly excluded expert parameters from the optimization (via if not group.is_expert_param check). This PR removes that limitation and adds full support for expert parallelism.

Changes

Core Changes:

• Add _EXPERT_DATA_PARALLEL_GROUP_AG process group in parallel_state.py
• Create expert AG group during initialize_model_parallel() when --create-all-gather-group is enabled
• Update get_expert_data_parallel_group() with independent_all_gather parameter
• Add has_separate_expert_all_gather_group() helper function

FSDP Integration:

• Extend FSDPDistributedIndex to store expt_fsdp_group_ag
• Update get_fsdp_group() to handle all 4 combinations: (is_expert_parallel × independent_all_gather)
• Update buffer initialization logic to use expert AG group for expert parameters
• Add UBR (User Buffer Registration) for expert AG group

Testing:

• Add test_separate_expert_all_gather_group() unit test

---

Contribution process

Pre-checks

• I want this PR in a versioned release and have added the appropriate Milestone (e.g., Core 0.8)
• 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

The following process is enforced via the CODEOWNERS file for changes into megatron/core. For changes outside of megatron/core, it is up to the PR author whether or not to tag the Final Reviewer team.

For MRs into `main` branch

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!

(Step 1): Add PR label Expert Review

(Step 2): Collect the expert reviewers reviews

1. Attach the Expert Review label when your PR is ready for review.
2. GitHub auto-assigns expert reviewers based on your changes. They will get notified and pick up your PR soon.

⚠️ Only proceed to the next step once all reviewers have approved, merge-conflict are resolved and the CI is passing.
Final Review might get declined if these requirements are not fulfilled.

(Step 3): Final Review

1. Add Final Review label
2. GitHub auto-assigns final reviewers based on your changes. They will get notified and pick up your PR soon.

(Optional Step 4): Cherry-pick into release branch

If this PR also needs to be merged into core_r* release branches, after this PR has been merged, select Cherry-pick to open a new PR into the release branch.

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.

Merging your PR

Any member of core-adlr and core-nemo will be able to merge your PR.

[copy-pr-bot]

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.

[youngeunkwon0405]

Have you tested this PR working fine with EP > 1 case? Did it show expected behavior?

[jeffnvidia]

Have you tested this PR working fine with EP > 1 case? Did it show expected behavior?

Hi, I just tested it with EP=4 on GAIA cluster on 2 nodes and it does show expected behaviour
EP=1: Expert parameters correctly use all 16 GPUs (no sharding)
EP=4: Expert parameters correctly use the expert-specific DP group (size 4)

[jeffnvidia]

could you review this PR @youngeunkwon0405 @shjwudp ? Thanks a lot

[youngeunkwon0405]

/ok to test 33ede31

[youngeunkwon0405]

Seems to have a lint error.

[jeffnvidia]

Seems to have a lint error.

I fixed the linting, could you retrigger the CI ? It should be fine now. Thanks !

[youngeunkwon0405]

/ok to test ce44736

[jeffnvidia]

/ok to test ce44736

could we push this PR also to final review ? @youngeunkwon0405
Thanks

[shjwudp]

Hi @jeffnvidia , this PR depends on PR-2663, so we’ll need to wait for that one to be merged first. Thanks for your patients.

[jaredcasper]

Please resolve conflicts now that 2663 has gone in.

[jeffnvidia]

Please resolve conflicts now that 2663 has gone in.

hey,

Please resolve conflicts now that 2663 has gone in.

so actually, I added everything in the new PR that uses process group collection, I think I can close this current PR : #3249