Add --speculative-moe-runner-backend server arg (#10183)

python/sglang/srt/server_args.py

@@ -36,6 +36,7 @@ from sglang.srt.utils.common import (
     SUPPORTED_LORA_TARGET_MODULES,
     configure_ipv6,
     cpu_has_amx_support,
+    get_bool_env_var,
     get_device,
     get_device_memory_capacity,
     get_device_sm,
@@ -377,6 +378,7 @@ class ServerArgs:
     speculative_accept_threshold_acc: float = 1.0
     speculative_token_map: Optional[str] = None
     speculative_attention_mode: str = "prefill"
+    speculative_moe_runner_backend: Optional[str] = None
     # For ngram only
     speculative_ngram_min_match_window_size: int = 1
     speculative_ngram_max_match_window_size: int = 12
@@ -1379,6 +1381,19 @@ class ServerArgs:
                 "FlashInfer TRTLLM MoE is enabled. --disable-shared-experts-fusion is automatically set."
             )
 
+        if get_bool_env_var("SGLANG_CUTLASS_MOE"):
+            logger.warning(
+                "SGLANG_CUTLASS_MOE is deprecated, use --moe-runner-backend=cutlass and/or --speculative-moe-runner-backend=cutlass instead"
+            )
+            assert (
+                self.quantization == "fp8"
+            ), "cutlass MoE is only supported with fp8 quantization"
+            self.moe_runner_backend = "cutlass"
+        if self.moe_runner_backend == "cutlass" and self.quantization == "fp8":
+            assert (
+                self.ep_size == 1
+            ), "FP8 Cutlass MoE is only supported with ep_size == 1"
+
     def _handle_a2a_moe(self):
         if self.moe_a2a_backend == "deepep":
             if self.deepep_mode == "normal":
@@ -2722,6 +2737,13 @@ class ServerArgs:
             help="Attention backend for speculative decoding operations (both target verify and draft extend). Can be one of 'prefill' (default) or 'decode'.",
             default=ServerArgs.speculative_attention_mode,
         )
+        parser.add_argument(
+            "--speculative-moe-runner-backend",
+            type=str,
+            choices=MOE_RUNNER_BACKEND_CHOICES,
+            default=ServerArgs.speculative_moe_runner_backend,
+            help="Choose the runner backend for MoE in speculative decoding.",
+        )
         # Ngram speculative decoding
         parser.add_argument(
             "--speculative-ngram-min-match-window-size",