Support piecewise cuda graph for fused marlin moe (#15100)

test/srt/test_piecewise_cuda_graph.py

@@ -214,6 +214,41 @@ class TestPiecewiseCudaGraphAWQ(CustomTestCase):
         self.assertGreaterEqual(metrics["score"], 0.65)
 
 
+class TestPiecewiseCudaGraphGPTQ(CustomTestCase):
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = "Qwen/Qwen3-30B-A3B-GPTQ-Int4"
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=["--enable-piecewise-cuda-graph"],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_mgsm_accuracy(self):
+        num_examples = 1319
+
+        args = SimpleNamespace(
+            base_url=self.base_url,
+            model=self.model,
+            eval_name="mgsm_en",
+            num_examples=num_examples,
+            num_threads=min(num_examples, 1024),
+        )
+
+        metrics = run_eval(args)
+        print(f"MGSM Accuracy: {metrics['score']:.3f}")
+
+        # Expected accuracy: 0.948, allow some variance
+        self.assertGreaterEqual(metrics["score"], 0.92)
+
+
 class TestPiecewiseCudaGraphFP8(CustomTestCase):
     """Test piecewise CUDA graph with FP8 quantized model"""