[None][feat] Add kimi_k2/glm_5 grouped routing and fused router to bench_moe

[guqiqi]

Dev Engineer Review

• Updates tests/microbenchmarks/bench_moe/mapping.py to construct DeepSeekV3MoeRoutingMethod with is_fused=True, enabling the fused DeepSeek-V3 routing path for this benchmark configuration.
• Fixes grouped-routing configuration for Kimi-K2 and adds GLM-5: tests/microbenchmarks/bench_moe/specs.py updates kimi_k2 to n_group=1/topk_group=1 and adds a new glm_5 ModelSpec with routing_method="DEEPSEEK_V3", n_group=1/topk_group=1, and quant_algo=None.
• Prevents sweep failures by pruning unsupported MegaMoE-DeepGemm + TEP combinations early: tests/microbenchmarks/bench_moe/search.py rejects MEGAMOE_DEEPGEMM when enable_dp is false (TEP) and world_size > 1, recording the candidate as skipped instead of hard-failing during backend initialization.
• Scope is limited to bench_moe microbenchmark code/specs; no unrelated config/test-list changes were made.

QA Engineer Review

• Updated test/benchmark code under tests/microbenchmarks/bench_moe/:
  • mapping.py: _create_routing_method (DeepSeek-V3 routing construction now uses is_fused=True).
  • search.py: is_candidate_valid (adds early prune for MEGAMOE_DEEPGEMM + TEP when enable_dp is false and world_size > 1).
  • specs.py: BUILT_IN_MODELS registry (adds glm_5, updates kimi_k2 routing grouping to n_group=1/topk_group=1).
• No corresponding entries were found in tests/integration/test_lists/ (e.g., test-db/, qa/, waives.txt) for bench_moe/microbenchmarks/bench_moe.
• Verdict: needs follow-up.

Description

This PR adds bench_moe support for two additional MoE models and fixes a routing
configuration issue for Kimi-K2:

• specs.py: set kimi_k2 n_group=1/topk_group=1. Real Kimi-K2 uses a single
  expert group; the previous multi-group value tripped the TRT-LLM routing kernel
  assertion topk_group <= 4 and failed the build. Also adds a glm_5 model spec
  (256 experts, top-8, DeepSeek-V3-style single-group routing).
• mapping.py: use the fused DeepSeek-V3 routing path (is_fused=True) for these specs.
• search.py: mark the MEGAMOE_DEEPGEMM + TEP combination as skipped rather than a
  hard build failure. That kernel assumes a partitioned MoE input, whereas TEP replicates
  the tensor TP-wide, so the combination is not valid and should be pruned from the sweep.

The change is scoped entirely to the bench_moe benchmarking utility and does not touch
runtime/serving code paths.

Test Coverage

Validated by running the bench_moe sweep locally for kimi_k2 and glm_5:
the previously-failing Kimi-K2 routing build now succeeds, glm_5 is enumerated and
benchmarked, and MEGAMOE_DEEPGEMM+TEP entries are reported as skipped instead of aborting
the sweep.

PR Checklist

• Please check this after reviewing the above items as appropriate for this PR.

GitHub Bot Help

To see a list of available CI bot commands, please comment /bot help.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 14941af4-4ac8-4750-9b0b-9700686b74a1

📥 Commits

Reviewing files that changed from the base of the PR and between 201960c and f73f203.

📒 Files selected for processing (3)

• tests/microbenchmarks/bench_moe/mapping.py
• tests/microbenchmarks/bench_moe/search.py
• tests/microbenchmarks/bench_moe/specs.py

🚧 Files skipped from review as they are similar to previous changes (3)

• tests/microbenchmarks/bench_moe/mapping.py
• tests/microbenchmarks/bench_moe/search.py
• tests/microbenchmarks/bench_moe/specs.py

---

Walkthrough

The MoE microbenchmark adds a glm_5 model specification, updates kimi_k2 routing groups, enables fused DeepSeek V3 routing, and rejects unsupported multi-rank MEGAMOE_DEEPGEMM candidates.

Changes

MoE benchmark updates

Layer / File(s)	Summary
Model and routing setup tests/microbenchmarks/bench_moe/specs.py, tests/microbenchmarks/bench_moe/mapping.py	Adds the glm_5 model spec, sets kimi_k2 routing groups to 1, and enables fused DeepSeek V3 routing.
Backend candidate validation tests/microbenchmarks/bench_moe/search.py	Rejects MEGAMOE_DEEPGEMM candidates when attention data parallelism is disabled and world_size > 1.

Estimated code review effort: 2 (Simple) | ~10 minutes

Suggested reviewers: leslie-fang25

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title matches the PR's main changes and follows the required [None][feat] format.
Description check	✅ Passed	The description includes the required Description, Test Coverage, and PR Checklist sections with relevant details.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands.}

[guqiqi]

/bot run

[tensorrt-cicd]

PR_Github #61523 [ run ] triggered by Bot. Commit: f73f203 Link to invocation

[tensorrt-cicd]

PR_Github #61523 [ run ] completed with state SUCCESS. Commit: f73f203
/LLM/main/L0_MergeRequest_PR pipeline #49739 completed with status: 'SUCCESS'

CI Report

Link to invocation