Disable Hopper GLM shared-expert fusion for modelopt_fp4 Marlin (#37325)

python/sglang/srt/models/deepseek_v2.py

@@ -200,6 +200,7 @@ from sglang.srt.utils import (
     LazyValue,
     add_prefix,
     is_non_idle_and_non_empty,
+    is_sm90_supported,
     make_layers,
     use_intel_amx_backend,
 )
@@ -2973,6 +2974,17 @@ class DeepseekV2ForCausalLM(nn.Module, DeepseekV2WeightLoaderMixin):
             )
         if get_exec().moe.enforce_shared_experts_fusion:
             return None
+        if (
+            quant_config is not None
+            and quant_config.get_name() == "modelopt_fp4"
+            and is_sm90_supported()
+            and get_moe_runner_backend().is_marlin()
+        ):
+            return (
+                "Hopper modelopt_fp4 with moe_runner_backend=marlin: "
+                "fusion off by default until the shared-expert fused load path "
+                "is validated."
+            )
         if is_sbo_enabled() or is_tbo_enabled():
             return "SBO/TBO enabled: incompatible with fusing shared expert into MoE kernel."
         if is_deepep_class_backend():

test/registered/unit/models/test_shared_experts_fusion_gates.py

@@ -159,6 +159,55 @@ class TestDeepseekV2Gate(_FusionGateCase):
         )
         self.assertIsNone(self._reason(DeepseekV2ForCausalLM, self._config(), matched))
 
+    def test_hopper_modelopt_fp4_marlin_disables_fusion_by_default(self):
+        import sglang.srt.models.deepseek_v2 as deepseek_v2
+        from sglang.srt.layers.moe.utils import MoeRunnerBackend
+        from sglang.srt.models.deepseek_v2 import DeepseekV2ForCausalLM
+
+        self._seed()
+        with (
+            unittest.mock.patch.object(
+                deepseek_v2, "is_sm90_supported", return_value=True
+            ),
+            unittest.mock.patch.object(
+                deepseek_v2,
+                "get_moe_runner_backend",
+                return_value=MoeRunnerBackend.MARLIN,
+            ),
+        ):
+            self.assertIn(
+                "fusion off by default",
+                self._reason(
+                    DeepseekV2ForCausalLM,
+                    self._config(),
+                    _quant("modelopt_fp4"),
+                ),
+            )
+
+    def test_hopper_modelopt_fp4_marlin_can_still_be_forced(self):
+        import sglang.srt.models.deepseek_v2 as deepseek_v2
+        from sglang.srt.layers.moe.utils import MoeRunnerBackend
+        from sglang.srt.models.deepseek_v2 import DeepseekV2ForCausalLM
+
+        self._seed(enforce_shared_experts_fusion=True)
+        with (
+            unittest.mock.patch.object(
+                deepseek_v2, "is_sm90_supported", return_value=True
+            ),
+            unittest.mock.patch.object(
+                deepseek_v2,
+                "get_moe_runner_backend",
+                return_value=MoeRunnerBackend.MARLIN,
+            ),
+        ):
+            self.assertIsNone(
+                self._reason(
+                    DeepseekV2ForCausalLM,
+                    self._config(),
+                    _quant("modelopt_fp4"),
+                )
+            )
+
 
 class TestGlmMoeLiteGate(_FusionGateCase):
     def _config(self, **kw):