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

2026-09-08 06:16:36 -07:00
parent 2d339ddef1
commit 5097f9ac95
2 changed files with 61 additions and 0 deletions
@@ -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):