MiMoV2: do not force flashinfer_trtllm when an EP a2a backend is active

[whn09]

Motivation

The MiMoV2 config-time override forces moe_runner_backend=flashinfer_trtllm on sm100 FP8
without checking whether an EP all-to-all backend is active:

if (
    get_platform().is_sm100
    and cfg.moe_runner_backend == "auto"
    and get_quantization_config(hf_config) == "fp8"
):
    overrides["moe_runner_backend"] = "flashinfer_trtllm"

FusedOpPool has no flashinfer_trtllm fused func registered for deepep or deepep_v2,
and flashinfer_trtllm has no runner_core (fused path only), so MoeRunner.__init__ hits
runner_core is None and fused_func is None:

NotImplementedError: Runner backend MoeRunnerBackend.FLASHINFER_TRTLLM requires a fused func
for a2a backend deepep_v2, but none is registered.

The user asked for auto. auto resolves to deep_gemm, which works. The override turns a
working default into a hard failure at model init, and because the substitution is silent the
traceback names a backend the command line never mentioned.

Sibling overrides already carry exactly this guard — glm4_moe.py:39, qwen3_moe.py:43,
deepseek_v4.py:62, nemotron_h.py:86 all test cfg.moe_a2a_backend == "none" before
forcing a runner. mimo_v2.py is the one that does not, so this brings it in line rather
than introducing a new convention.

Modifications

One condition, and cfg.moe_a2a_backend == "none", plus a comment explaining why (the next
person to read this will otherwise wonder whether the guard is load-bearing). No behaviour
change without an a2a backend: the sm100 FP8 fast path is untouched for the plain-TP case
the override was written for.

Accuracy Tests

No numerical change. With --moe-a2a-backend none the resolved runner is flashinfer_trtllm
before and after — verified via /get_server_info on 8× B200 with an FP8 MiMoV2 checkpoint.
With --moe-a2a-backend deepep_v2 --moe-runner-backend auto the resolved runner goes from
"init raises" to deep_gemm, whose numerics against the deepep (v1) path are 10/12 prompts
bit-identical on a fixed logprob set (details in the companion PR).

Speed Tests and Profiling

Unchanged where the override still fires. Where it no longer fires, the comparison is
NotImplementedError versus a running server.

Worth stating plainly for anyone weighing it: flashinfer_trtllm really is the faster runner
where it is usable — on decode at global bs=32 it is 24.58 ms/token versus 34.90 for
deep_gemm, ~30%. That is an argument for registering a deepep/deepep_v2 fused func, not
for forcing a runner that cannot load. (The one existing deepep+flashinfer_cutedsl fused
func takes CuteDslFp4MoeQuantInfo, i.e. NVFP4 only, so it does not help an FP8 checkpoint.)

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests. — happy to add one asserting _mimo_v2_overrides returns no moe_runner_backend when moe_a2a_backend != "none"; it needs no GPU, say the word
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

I cannot add the run-ci label myself, so this needs someone with write access to start CI.

---

Companions, independent of each other: #39080 (enable deepep_v2 for MiMoV2 — the PR that
makes this guard reachable), #39078 (ElasticBuffer vs CUDA graph capture), #37211
(ep_scatter_from_psum kernel args).

---

CI States

Latest PR Test (Base): ❌ Run #34590432083
Latest PR Test (Extra): ❌ Run #34590431824
Latest PR Test (AMD ROCm 10): ❌ Run #34590431968