Make the mxfp8 MoE runner backend list extensible (#30828)

python/sglang/srt/arg_groups/overrides.py

@@ -1928,13 +1928,10 @@ def _moe_runner_backend_quant_constraints(view: Any) -> dict:
                 "flashinfer_trtllm_routed."
             )
     if view.quantization == "mxfp8":
+        from sglang.srt.server_args import MXFP8_MOE_RUNNER_BACKEND_CHOICES
+
         is_gfx95_mxfp8 = is_hip() and is_gfx95_supported()
-        allowed = [
+        allowed = list(MXFP8_MOE_RUNNER_BACKEND_CHOICES)
-            "cutlass",
-            "deep_gemm",
-            "flashinfer_trtllm",
-            "flashinfer_trtllm_routed",
-        ]
         if is_gfx95_mxfp8:
             allowed.append("triton")
         mxfp8_default = "triton" if is_gfx95_mxfp8 else "flashinfer_trtllm"

python/sglang/srt/server_args.py

@@ -268,6 +268,13 @@ MOE_A2A_BACKEND_CHOICES = [
     "megamoe",
 ]
 
+MXFP8_MOE_RUNNER_BACKEND_CHOICES = [
+    "cutlass",
+    "deep_gemm",
+    "flashinfer_trtllm",
+    "flashinfer_trtllm_routed",
+]
+
 FP8_GEMM_RUNNER_BACKEND_CHOICES = [
     "auto",
     "deep_gemm",
@@ -367,6 +374,10 @@ def add_moe_runner_backend_choices(choices):
     MOE_RUNNER_BACKEND_CHOICES.extend(choices)
 
 
+def add_mxfp8_moe_runner_backend_choices(choices):
+    MXFP8_MOE_RUNNER_BACKEND_CHOICES.extend(choices)
+
+
 def add_fp8_gemm_runner_backend_choices(choices):
     FP8_GEMM_RUNNER_BACKEND_CHOICES.extend(choices)