[Fix] Route asymmetric-KV models to fa4 on SM100 and pin MiMoV2 FP8 MoE to flashinfer_trtllm (#32818)

python/sglang/srt/arg_groups/overrides.py

@@ -448,10 +448,21 @@ def _deepseek_family_overrides(server_args: Any, hf_config: Any) -> dict:
 # Keep in sync with MIMO_V2_MODEL_ARCHS (server_args.py / configs/hf_config.py).
 @_register_for("MiMoV2ForCausalLM", "MiMoV2FlashForCausalLM")
 def _mimo_v2_overrides(server_args: Any, hf_config: Any) -> dict:
+    overrides: Dict[str, Any] = {}
     if server_args.speculative_algorithm == "EAGLE":
         logger.info("Enable multi-layer EAGLE speculative decoding for MiMoV2 model.")
-        return {"enable_multi_layer_eagle": True}
-    return {}
+        overrides["enable_multi_layer_eagle"] = True
+
+    # On Blackwell "auto" falls through to the triton fused-MoE runner, ~12%
+    # slower at bs=1 decode. FP4 checkpoints use flashinfer_mxfp4 instead.
+    if (
+        is_sm100_supported()
+        and server_args.moe_runner_backend == "auto"
+        and get_quantization_config(hf_config) == "fp8"
+    ):
+        overrides["moe_runner_backend"] = "flashinfer_trtllm"
+        logger.info("MiMoV2 FP8 on SM100: moe_runner_backend=flashinfer_trtllm.")
+    return overrides
 
 
 @_register_for("MiniMaxM2ForCausalLM")

python/sglang/srt/configs/model_config.py

@@ -713,6 +713,17 @@ class ModelConfig:
     def linear_attn_registry_result(self) -> Any:
         return get_linear_attn_config(self.hf_config)
 
+    @property
+    def has_asymmetric_kv(self) -> bool:
+        """Whether K and V rows differ in width (MiMoV2 is 192 / 128).
+
+        Not an ``__init__`` field because the MLA special-casing below still
+        rewrites ``v_head_dim``.
+        """
+        return (
+            self.head_dim != self.v_head_dim or self.swa_head_dim != self.swa_v_head_dim
+        )
+
     def _detect_attention_sinks(self) -> bool:
         """Check whether the model uses learned attention sinks.
 

python/sglang/srt/server_args.py

@@ -5565,6 +5565,10 @@ class ServerArgs:
                     or self.speculative_eagle_topk is not None
                 )
             ):
+                # trtllm_mha requires equal K/V row widths; fa4 carries
+                # v_head_dim through.
+                if model_config.has_asymmetric_kv:
+                    return "fa4"
                 return "trtllm_mha"
             elif is_hip():
                 return "aiter"