[Fix] Route asymmetric-KV models to fa4 on SM100 and pin MiMoV2 FP8 MoE to flashinfer_trtllm

[hnyls2002]

MiMo-V2.5 cannot be deployed on SM100/SM103 with default flags (#31243). The SM100 auto-dispatch picks trtllm_mha, whose paged-KV kernel requires equal K/V row widths, but MiMoV2 has asymmetric KV (head_dim 192, v_head_dim 128), so decode CUDA-graph capture aborts with Check failed: key_cache.size(i) == value_cache.size(i) (37074 vs. 24716). The reporter had to fall back to triton with CUDA graphs disabled.

Three changes:

1. ModelConfig.has_asymmetric_kv — new property (head_dim != v_head_dim or swa_head_dim != swa_v_head_dim). A property rather than an __init__ field because the MLA special-casing further down still rewrites v_head_dim.
2. _get_default_attn_backend — on SM100, asymmetric-KV MHA models now resolve to fa4, which carries v_head_dim through, instead of trtllm_mha. This is a capability check, not a model special case; MLA models are asymmetric by construction but resolve through the separate MLA branch and are unaffected.
3. _mimo_v2_overrides — on SM100 with an FP8 checkpoint and moe_runner_backend still auto, pin flashinfer_trtllm. Left on auto the MoE falls through to the triton fused-MoE runner (auto only selects DeepGEMM when the a2a backend is DeepEP / Mooncake / NIXL). Gated on fp8 because the FP4 checkpoints run through flashinfer_mxfp4 instead.

Also updates the cookbook, whose base-variant Blackwell recipe emitted neither flag while the Pro recipe emitted both.

Measured on MiMo-V2.5 (310B) TP=4, bs=1 decode, --speculative-algorithm EAGLE --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 --enable-multi-layer-eagle:

MoE runner	tok/s	accept_length
triton (previous auto result)	315.7	2.88
flashinfer_trtllm	353.0	2.94

+11.8% output throughput (~+9.6% after discounting the 2.1% higher accept length).

Verification: a launch command carrying no backend flags at all now resolves to attention_backend='fa4', moe_runner_backend='flashinfer_trtllm', page_size=128 and serves correctly, with generated tokens identical to passing both flags explicitly. test_model_overrides.py 67 passed.

---

CI States

Latest PR Test (Base): ❌ Run #30495495475
Latest PR Test (Extra): ❌ Run #30495495333

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
lmsysorg	🟢 Ready	View Preview	Jul 29, 2026, 10:13 PM

💡 Tip: Enable Workflows to automatically generate PRs for you.

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
lmsysorg	🟡 Building	–	Jul 29, 2026, 10:11 PM

💡 Tip: Enable Workflows to automatically generate PRs for you.

[hnyls2002]

/rerun-test test_model_overrides.py test_model_config_shapes.py test/registered/models_e2e/test_mimo_v2.py test/registered/models_e2e/test_mimo_v2_flash.py

[github-actions]

Results for /rerun-test test_model_overrides.py test_model_config_shapes.py test/registered/models_e2e/test_mimo_v2.py test/registered/models_e2e/test_mimo_v2_flash.py:

🚀 ubuntu-latest (2 tests): ✅ View workflow run

cd test/ && python3 registered/unit/test_model_overrides.py
cd test/ && python3 registered/unit/configs/test_model_config_shapes.py

🚀 8-gpu-h200 (2 tests): ✅ View workflow run

cd test/ && python3 registered/models_e2e/test_mimo_v2.py
cd test/ && python3 registered/models_e2e/test_mimo_v2_flash.py