fix(moe): exclude padding tokens from dropless HybridEP routing

[seonjinn]

• I, the PR author, have personally reviewed every line of this PR.

What does this PR do?

Exclude physical padding rows from MoE routing before dropless HybridEP dispatch.

This ports the routing-mask behavior already merged to dev in #5542 onto main:

• zero probabilities and clear routing assignments for padding rows;
• apply the behavior only to flex + hybridep with no expert or rank capacity factor;
• preserve existing routing behavior for other dispatchers, backends, and capacity modes.

Why?

#6114 fixes padding-mask broadcasting for expert-bias token counts, but it does not remove physical padding rows from HybridEP dispatch. NeMo-RL sequence-packed THD workloads construct and pass this mask in NVIDIA-NeMo/RL#2964; MCore must consume it before dispatch to prevent padding rows from reaching experts.

After merging current main, this PR drops its older changes that are now owned by #6114 and the current sequence-parallel mask-alignment paths. The final diff is limited to the router behavior and focused unit coverage.

Validation

Validation	Result
H100, 8-rank focused test on unpatched main	Expected RED: 2 failed, 8 passed; fused and unfused HybridEP retained 64 nonzero padding routes
H100, 8-rank focused test with this change	GREEN: 10 passed, 54 deselected
Non-HybridEP and capacity-mode controls	All 6 parameterized cases passed
#6114 expert-bias masked/unmasked controls	Both cases passed
Static checks	py_compile, Black, isort, Ruff, and git diff --check passed

The focused H100 run used the final router and test contents. The branch then merged the latest upstream main; that upstream commit did not touch either changed file.

Related

• [codex] Exclude padding tokens from MoE routing #5542
• fix(moe): broadcast padding mask for expert bias counts #6114
• fix: resolve HybridEP sequence packing padding NVIDIA-NeMo/RL#2964

[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.

[github-actions]

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:

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

See the contribution guide for more details.