Fix FP4 MoE accuracy from missing routed_scaling_factor (#8333)

python/sglang/srt/layers/quantization/modelopt_quant.py

@@ -952,7 +952,6 @@ class ModelOptNvFp4FusedMoEMethod(FusedMoEMethodBase):
         tp_rank: Optional[int] = None,
         tp_size: Optional[int] = None,
     ) -> torch.Tensor:
-
         assert activation == "silu", "Only SiLU activation is supported."
 
         if self.enable_flashinfer_moe:
@@ -982,13 +981,15 @@ class ModelOptNvFp4FusedMoEMethod(FusedMoEMethodBase):
                 tp_size=tp_size,
                 tp_rank=tp_rank,
                 tune_max_num_tokens=next_power_of_2(x.shape[0]),
-            )
+            )[0]
-            return output[0]
+            if routed_scaling_factor is not None:
+                output *= routed_scaling_factor
+            return output
 
         from sglang.srt.layers.moe.cutlass_moe import cutlass_moe_fp4
 
         topk_weights, topk_ids, _ = topk_output
-        return cutlass_moe_fp4(
+        output = cutlass_moe_fp4(
             a=x,
             a1_gscale=layer.w13_input_scale_quant,
             w1_fp4=layer.w13_weight,
@@ -1003,3 +1004,6 @@ class ModelOptNvFp4FusedMoEMethod(FusedMoEMethodBase):
             params=layer.cutlass_moe_params,
             apply_router_weight_on_input=apply_router_weight_on_input,
         ).to(x.dtype)
+        if routed_scaling_factor is not None:
+            output *= routed_scaling_factor
+        return output

python/sglang/srt/server_args.py

@@ -433,10 +433,6 @@ class ServerArgs:
                 self.quantization == "modelopt_fp4"
             ), "modelopt_fp4 quantization is required for Flashinfer MOE"
             os.environ["TRTLLM_ENABLE_PDL"] = "1"
-            self.disable_shared_experts_fusion = True
-            logger.warning(
-                f"Flashinfer MoE is enabled. Shared expert fusion is disabled."
-            )
 
         # DeepEP MoE
         if self.enable_deepep_moe: