fix(minimax): use routed TRT-LLM for NVFP4 MoE auto on SM100 (#32229)

Co-authored-by: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com>

python/sglang/srt/arg_groups/overrides.py

@@ -731,10 +731,21 @@ def _mimo_v2_overrides(server_args: Any, hf_config: Any) -> dict:
 
 @_register_for("MiniMaxM2ForCausalLM")
 def _minimax_m2_overrides(server_args: Any, hf_config: Any) -> dict:
+    overrides = {"enable_tf32_matmul": True}
     logger.info(
         "Enable TF32 matmul for MiniMaxM2ForCausalLM model to improve gate gemm performance."
     )
-    return {"enable_tf32_matmul": True}
+    if (
+        is_sm100_supported()
+        and server_args.moe_runner_backend == "auto"
+        and server_args.get_model_config().quantization == "modelopt_fp4"
+    ):
+        overrides["moe_runner_backend"] = "flashinfer_trtllm_routed"
+        logger.info(
+            "Use flashinfer_trtllm_routed as MoE runner backend on SM10X "
+            "for MiniMaxM2ForCausalLM with modelopt_fp4."
+        )
+    return overrides
 
 
 @_register_for("MiniMaxM3SparseForCausalLM", "MiniMaxM3SparseForConditionalGeneration")