fix(tensor-parallel): AUT-1177 deduplicate expert shards by expert TP

[svcnemo-autobot]

Background

Megatron-Bridge dev validation PR NVIDIA-NeMo/Megatron-Bridge#5174 exposed that MCore dev dropped expert-TP-aware duplicate filtering, causing expert parameter norm failures.

What changed

• Route allreduce=False parameters through the expert TP group.
• Preserve regular TP behavior for dense parameters.
• Add focused coverage for both paths.

Details

flowchart LR
 P[Parameter] --> A{allreduce?}
 A -->|false| E[Expert TP group]
 A -->|true| T[Regular TP group]
 E --> D[Duplicate filter]
 T --> D

Loading

The commit is based directly on Bridge's pinned dev SHA 95e4bafebaa799d166975ef82066a3c46648e004. Bridge PR #5174 will temporarily validate this PR head; the temporary pin will be replaced by the merged dev commit before Bridge lands.

Tested

• isort, Black, and Ruff checks passed.
• Focused regression: 4 passed.
• Complete eight-rank tensor-parallel module: 7 passed on the same logic before a formatter-only amendment.

[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.

[svcnvidia-nemo-ci]

/ok to test e2dc53a

[svcnvidia-nemo-ci]

/ok to test af91b6b

[svcnvidia-nemo-ci]

🔄 Merge queue validation started!

You can track the progress here: https://github.com/NVIDIA/Megatron-LM/actions/runs/30646819346