[AMD] Fix Quark Shared Experts Fusion Gate after load-time-override Removal (#35200)

test/registered/unit/models/test_deepseek_v4_shared_expert_fusion.py

@@ -86,6 +86,29 @@ class TestDeepseekV4SharedExpertFusionPolicy(CustomTestCase):
                 SimpleNamespace(n_shared_experts=2), None
             )
 
+    def test_mixed_precision_quant_vetoes_even_when_enforced(self):
+        """A precision mismatch causes crash when shared expert fusion is enabled,
+        so --enforce-shared-experts-fusion must not override it. Guards the gap
+        where the enforce early-return skipped the quant check entirely."""
+        self._publish(enforce=True)
+        mixed = SimpleNamespace(
+            get_name=lambda: "quark", can_fuse_shared_expert=lambda: False
+        )
+        self.assertIn(
+            "higher precision",
+            DeepseekV4ForCausalLM.shared_experts_fusion_disable_reason(
+                SimpleNamespace(n_shared_experts=1), mixed
+            ),
+        )
+        matched = SimpleNamespace(
+            get_name=lambda: "quark", can_fuse_shared_expert=lambda: True
+        )
+        self.assertIsNone(
+            DeepseekV4ForCausalLM.shared_experts_fusion_disable_reason(
+                SimpleNamespace(n_shared_experts=1), matched
+            )
+        )
+
     def test_dspark_entry_class_uses_the_v4_gate(self):
         """A DSV4 DSpark draft must inherit the target's default fusion policy."""
         self._publish(enforce=False)

test/registered/unit/models/test_shared_experts_fusion_gates.py

@@ -117,6 +117,25 @@ class TestDeepseekV2Gate(_FusionGateCase):
             self._reason(DeepseekV2ForCausalLM, self._config(), moe_ep_size=2)
         )
 
+    def test_mixed_precision_quant_vetoes_even_when_enforced(self):
+        """A precision mismatch causes crash when shared expert fusion is enabled,
+        so --enforce-shared-experts-fusion must not override it. Guards the gap
+        where the enforce early-return skipped the quant check entirely."""
+        from sglang.srt.models.deepseek_v2 import DeepseekV2ForCausalLM
+
+        self._seed(enforce_shared_experts_fusion=True)
+        mixed = SimpleNamespace(
+            get_name=lambda: "quark", can_fuse_shared_expert=lambda: False
+        )
+        self.assertIn(
+            "higher precision",
+            self._reason(DeepseekV2ForCausalLM, self._config(), mixed),
+        )
+        matched = SimpleNamespace(
+            get_name=lambda: "quark", can_fuse_shared_expert=lambda: True
+        )
+        self.assertIsNone(self._reason(DeepseekV2ForCausalLM, self._config(), matched))
+
 
 class TestGlmMoeLiteGate(_FusionGateCase):
     def _config(self, **kw):