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

[FDecaYed]

What

Exclude padding-token rows from the routing probabilities and sparse routing map only for dropless HybridEP. Dispatchers that still require fixed route counts keep their existing behavior.

This is a self-contained correctness prerequisite for the HybridModel-only DSv4 merge effort.

Provenance

• Refactors the DSv4 feature set captured in [Tracking] DeepSeek-v4 HybridModel merge plan #5795.
• Ports the focused fix from [codex] Exclude padding tokens from MoE routing #5542 to main.
• Original implementation by @xiaoyao0115; DSv4 integration work by @hxbai and the source authors.
• Reconstructed against current main; feature cutoff: 2026-07-21.
• GPTModel integration is intentionally out of scope.

Testing

• uv run isort --check-only megatron/core/transformer/moe/router.py tests/unit_tests/transformer/moe/test_routers.py
• uv run ruff check megatron/core/transformer/moe/router.py tests/unit_tests/transformer/moe/test_routers.py
• uv run python -m compileall -q megatron/core/transformer/moe/router.py tests/unit_tests/transformer/moe/test_routers.py
• CUDA unit tests require the project CUDA/Triton environment and are left to CI.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[FDecaYed]

/ok to test 7fd4938

[FDecaYed]

Closing this draft until its real dependency round. Its logical successor is P10 (HybridEP packed-token correctness), to be rebuilt after #5008, P4, and P7 merge. The publication diff will use the current #5008 flag name and will reference tracking PR #5795.