diff --git a/python/sglang/srt/layers/quantization/w8a8_int8.py b/python/sglang/srt/layers/quantization/w8a8_int8.py
index 7c145e310..7c9ff6385 100644
--- a/python/sglang/srt/layers/quantization/w8a8_int8.py
+++ b/python/sglang/srt/layers/quantization/w8a8_int8.py
@@ -28,6 +28,7 @@ from sglang.srt.utils import (
     set_weight_attrs,
     use_intel_amx_backend,
 )
+from sglang.srt.utils.patch_torch import register_fake_if_exists
 
 if TYPE_CHECKING:
     from sglang.srt.layers.moe.token_dispatcher import StandardDispatchOutput
@@ -39,6 +40,20 @@ _is_cpu = is_cpu()
 if _is_cuda:
     from sgl_kernel import int8_scaled_mm
 
+    @register_fake_if_exists("sgl_kernel::int8_scaled_mm")
+    def _int8_scaled_mm_abstract(
+        mat_a,
+        mat_b,
+        scales_a,
+        scales_b,
+        out_dtype,
+        bias=None,
+    ):
+        M = mat_a.shape[-2]
+        N = mat_b.shape[-1]
+        return mat_a.new_empty((M, N), dtype=out_dtype)
+
+
 logger = logging.getLogger(__name__)
 
 
diff --git a/test/srt/run_suite.py b/test/srt/run_suite.py
index 91c224e89..225c512f4 100644
--- a/test/srt/run_suite.py
+++ b/test/srt/run_suite.py
@@ -151,7 +151,7 @@ suites = {
         TestFile("test_local_attn.py", 411),
         TestFile("test_multi_instance_release_memory_occupation.py", 64),
         TestFile("test_pp_single_node.py", 500),
-        TestFile("test_piecewise_cuda_graph.py", 1200),
+        TestFile("test_piecewise_cuda_graph.py", 1260),
         TestFile("test_epd_disaggregation.py", 400),
     ],
     "per-commit-8-gpu-h200": [
diff --git a/test/srt/test_piecewise_cuda_graph.py b/test/srt/test_piecewise_cuda_graph.py
index 4fc5934d6..75d80cf20 100644
--- a/test/srt/test_piecewise_cuda_graph.py
+++ b/test/srt/test_piecewise_cuda_graph.py
@@ -288,6 +288,43 @@ class TestPiecewiseCudaGraphFP8(CustomTestCase):
         print(f"MGSM Accuracy: {metrics['score']:.3f}")
 
 
+class TestPiecewiseCudaGraphW8A8Int8(CustomTestCase):
+    """Test piecewise CUDA graph with W8A8 INT8 quantized model"""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = "RedHatAI/Llama-3.2-1B-Instruct-quantized.w8a8"
+        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",
+                "--quantization",
+                "w8a8_int8",
+            ],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_mgsm_accuracy(self):
+        """Test MGSM accuracy with W8A8 INT8 model"""
+        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}")
+        self.assertGreaterEqual(metrics["score"], 0.40)
+
+
 class TestPiecewiseCudaGraphQwen25VL(CustomTestCase):
     """Test piecewise CUDA graph with Qwen2.5-VL-7B-Instruct model"""