Accept MXFP8 dispatch in FlashInfer A2A TRT-LLM MoE (#39613)

Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai>

python/sglang/srt/layers/moe/moe_runner/flashinfer_trtllm.py

@@ -1707,6 +1707,17 @@ def fused_experts_flashinfer_to_flashinfer_trtllm(
             use_routed_topk=True,
         )
     elif isinstance(quant_info, FlashInferTrtllmFp8MoeQuantInfo):
+        mxfp8_dispatch = (
+            quant_info.use_mxfp8
+            and dispatch_output.hidden_states.dtype == torch.float8_e4m3fn
+        )
+        if mxfp8_dispatch:
+            if dispatch_output.hidden_states_scale is None:
+                raise ValueError(
+                    "FlashInfer A2A + TRT-LLM Gen MXFP8 MoE requires activation "
+                    "scales alongside the FP8 dispatch payload."
+                )
+        else:
             if dispatch_output.hidden_states.dtype != torch.bfloat16:
                 raise TypeError(
                     "FlashInfer A2A + TRT-LLM Gen FP8 MoE requires a BF16 "