FlashInfer NVFP4 MoE with EP & 2-stream shared expert (#7327)

Co-authored-by: JieXin Liang <Alcanderian@users.noreply.github.com>
Co-authored-by: alcanderian <alcanderian@gmail.com>

python/sglang/srt/server_args.py

@@ -152,6 +152,7 @@ class ServerArgs:
     ep_size: int = 1
     enable_ep_moe: bool = False
     enable_deepep_moe: bool = False
+    enable_flashinfer_moe: bool = False
     deepep_mode: Optional[Literal["auto", "normal", "low_latency"]] = "auto"
     ep_num_redundant_experts: int = 0
     ep_dispatch_algorithm: Optional[Literal["static", "dynamic", "fake"]] = None
@@ -244,7 +245,15 @@ class ServerArgs:
             logger.warning(
                 f"EP MoE is enabled. The expert parallel size is adjusted to be the same as the tensor parallel size[{self.tp_size}]."
             )
-
+        if self.enable_flashinfer_moe:
+            assert (
+                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."
+            )
         # Set missing default values
         if self.tokenizer_path is None:
             self.tokenizer_path = self.model_path
@@ -1162,6 +1171,11 @@ class ServerArgs:
             action="store_true",
             help="Enabling expert parallelism for moe. The ep size is equal to the tp size.",
         )
+        parser.add_argument(
+            "--enable-flashinfer-moe",
+            action="store_true",
+            help="Enable FlashInfer CUTLASS MoE backend for modelopt_fp4 quant on Blackwell. Supports MoE-EP with --enable-ep-moe",
+        )
         parser.add_argument(
             "--enable-deepep-moe",
             action="store_true",