[Dev] add support for deepep/hybridep dispatcher under thd format training

[HaochenYuan]

What does this PR do ?

Previous PR added support for thd format in training, but for MoE dispatcher, only all2all type is supported. This PR adds the deepep & hybridep backend.

⚠️ 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"

1. When your PR is ready, click Ready for Review.
2. 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.

[Victarry]

@HaochenYuan Could please you add an UT to guard this and submit a PR to main.

cc @Autumn1998 for review

[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.

[HaochenYuan]

@HaochenYuan Could please you add an UT to guard this and submit a PR to main.

cc @Autumn1998 for review

UT added.

[yaox12]

/claude strict-review

[claude]

Code Review Summary

PR: Add support for deepep/hybridep dispatcher under THD format training

Overview

This PR enables HybridEP and DeepEP MoE dispatcher backends to work with THD-format sequence packing (variable-length token counts per rank). The core approach — all-reduce MAX to find the group-wide padded token count, pad routing metadata and hidden states before dispatch, trim after combine — is sound. The gradient flow through pad/trim is correct since torch.cat and slicing are differentiable with the right zero-gradient semantics.

The assertion update in transformer_config.py correctly allows "flex" (which wraps deepep/hybridep) alongside "alltoall" for sequence packing. The fused_a2a.py change to skip RDMA buffer allocation for node-local EP groups is a reasonable fix for DeepEP builds without internode support.

Findings

Severity	Count
CRITICAL	0
IMPORTANT	2
SUGGESTION	2

IMPORTANT:

1. CPU-GPU sync in _HybridEPManager.setup_metadata — int(nt.item()) forces a host-device sync on every forward pass when THD packing is active. This is functionally necessary but worth documenting, and worth considering whether a config-derived upper bound could avoid it.
2. Test skip condition too weak — test_sequence_packing_thd_e2e_proxy_model needs 16 ranks (tp=2, pp=2, cp=2, ep=2) but only guards with world_size % 8 != 0, which passes on 8-GPU nodes where initialize_model_parallel would then fail. Should be world_size < 16.

SUGGESTION:

1. The hardcoded 64 alignment for padded token count should be a named constant with a reference to the HybridEP kernel requirement.
2. The torch.cuda.device_count() heuristic in fused_a2a.py assumes all local GPUs are visible; a comment noting this assumption would help.

Risk Assessment: Low-Medium

The core dispatcher changes are well-contained in _HybridEPManager and correctly gated behind sequence_packing_scheduler is not None, so existing non-packing paths are unaffected. The fused_a2a.py RDMA change is additive (only skips unnecessary allocation). Main risk is the test skip condition allowing false passes on standard 8-GPU CI nodes.

[yaox12]

LGTM. Please fix the CI failure.

[Autumn1998]

LGTM

[svcnvidia-nemo-ci]

🔄 Merge queue validation started!

You can track the progress here: https://github.com/NVIDIA/Megatron-LM/actions/runs/26200974444

[svcnvidia-nemo-ci]

🔄 Merge queue validation started!

You can track the progress here: https://github.com/NVIDIA/Megatron-LM/actions/runs/26206185854