[bugfix] Fix flashinfer cutlass EP moe after MoE refactor (#8630)

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

@@ -119,7 +119,8 @@ class FusedMoE(torch.nn.Module):
                 * self.num_local_experts : (self.moe_ep_rank + 1)
                 * self.num_local_experts
             ] = torch.arange(0, self.num_local_experts, dtype=torch.int32, device="cpu")
-            self.expert_map_gpu = self.expert_map_cpu.to(device="cuda")
+            if not self.enable_flashinfer_cutlass_moe:
+                self.expert_map_gpu = self.expert_map_cpu.to(device="cuda")
 
         self.routed_scaling_factor = routed_scaling_factor
         assert intermediate_size % self.moe_tp_size == 0

python/sglang/srt/server_args.py

@@ -437,6 +437,11 @@ class ServerArgs:
                 self.quantization == "modelopt_fp4"
             ), "modelopt_fp4 quantization is required for Flashinfer MOE"
             os.environ["TRTLLM_ENABLE_PDL"] = "1"
+            if self.enable_ep_moe:
+                self.ep_size = self.tp_size
+                logger.warning(
+                    f"Flashinfer cutlass MoE and EP MoE are enabled. The expert parallel size is adjusted to be the same as the tensor parallel size[{self.tp_size}]."
+                )
 
         if self.enable_flashinfer_trtllm_moe:
             assert self.enable_ep_moe, "EP MoE is required for Flashinfer TRTLLM MOE"