Fix DSR1 accuracy for flashinfer_trtllm MoE with FP8 quantization (#11081)

python/sglang/srt/layers/moe/fused_moe_triton/layer.py

@@ -575,9 +575,9 @@ class FusedMoE(torch.nn.Module):
             )
 
         # Flashinfer assumes w31 format for w13_weight. Same for the scales.
-        if (
-            should_use_flashinfer_trtllm_moe()
-            and self.quant_method.__class__.__name__ == "ModelOptNvFp4FusedMoEMethod"
+        if should_use_flashinfer_trtllm_moe() and (
+            isinstance(self.quant_method, ModelOptNvFp4FusedMoEMethod)
+            or isinstance(self.quant_method, Fp8MoEMethod)
         ):
             shard_id = {"w1": "w3", "w3": "w1", "w2": "w2"}[shard_id]
 

python/sglang/srt/server_args.py

@@ -916,7 +916,7 @@ class ServerArgs:
         if self.moe_runner_backend == "flashinfer_trtllm":
             assert (
                 self.quantization == "modelopt_fp4" or self.quantization == "fp8"
-            ), "modelopt_fp4 quantization is required for Flashinfer TRTLLM MoE"
+            ), "modelopt_fp4 or fp8 quantization is required for Flashinfer TRTLLM MoE"
             self.disable_shared_experts_fusion = True
             logger.warning(
                 "FlashInfer TRTLLM MoE is enabled. --disable-shared-experts-fusion is automatically set."