diff --git a/python/sglang/srt/layers/quantization/fp8.py b/python/sglang/srt/layers/quantization/fp8.py
index d6e13b4eb..0a03f5dc4 100644
--- a/python/sglang/srt/layers/quantization/fp8.py
+++ b/python/sglang/srt/layers/quantization/fp8.py
@@ -1320,10 +1320,3 @@ class Fp8KVCacheMethod(BaseKVCacheMethod):
 
     def __init__(self, quant_config: Fp8Config):
         super().__init__(quant_config)
-
-
-if _is_cuda:
-
-    @torch.library.register_fake("sgl_kernel::fp8_scaled_mm")
-    def _(mat_a, mat_b, scales_a, scales_b, out_dtype, bias):
-        return mat_a.new_empty((mat_a.shape[0], mat_b.shape[-1]), dtype=out_dtype)
diff --git a/python/sglang/srt/layers/quantization/fp8_utils.py b/python/sglang/srt/layers/quantization/fp8_utils.py
index 5cbde4c3e..f9578e4fe 100644
--- a/python/sglang/srt/layers/quantization/fp8_utils.py
+++ b/python/sglang/srt/layers/quantization/fp8_utils.py
@@ -52,6 +52,14 @@ if _use_aiter:
 if _is_cuda:
     from sgl_kernel import fp8_blockwise_scaled_mm, fp8_scaled_mm
 
+    @torch.library.register_fake("sgl_kernel::fp8_scaled_mm")
+    def _fp8_scaled_mm_abstract(mat_a, mat_b, scales_a, scales_b, out_dtype, bias=None):
+        # mat_a: [M, K], mat_b: [K, N] or [N, K] depending on callsite layout; output is [M, N].
+        M = mat_a.shape[-2]
+        N = mat_b.shape[-1]
+        return mat_a.new_empty((M, N), dtype=out_dtype)
+
+
 use_vllm_cutlass_w8a8_fp8_kernel = get_bool_env_var("USE_VLLM_CUTLASS_W8A8_FP8_KERNEL")
 use_triton_w8a8_fp8_kernel = get_bool_env_var("USE_TRITON_W8A8_FP8_KERNEL")
 
diff --git a/test/srt/run_suite.py b/test/srt/run_suite.py
index a83a0df20..b4ae69f54 100644
--- a/test/srt/run_suite.py
+++ b/test/srt/run_suite.py
@@ -94,7 +94,7 @@ suites = {
         TestFile("test_original_logprobs.py", 41),
         TestFile("test_page_size.py", 60),
         TestFile("test_penalty.py", 82),
-        TestFile("test_piecewise_cuda_graph.py", 450),
+        TestFile("test_piecewise_cuda_graph.py", 600),
         TestFile("test_priority_scheduling.py", 130),
         TestFile("test_pytorch_sampling_backend.py", 66),
         TestFile("test_radix_attention.py", 105),
diff --git a/test/srt/test_piecewise_cuda_graph.py b/test/srt/test_piecewise_cuda_graph.py
index 19723e5e6..3dc410ebf 100644
--- a/test/srt/test_piecewise_cuda_graph.py
+++ b/test/srt/test_piecewise_cuda_graph.py
@@ -209,5 +209,44 @@ class TestPiecewiseCudaGraphAWQ(CustomTestCase):
         self.assertGreaterEqual(metrics["score"], 0.65)
 
 
+class TestPiecewiseCudaGraphFP8(CustomTestCase):
+    """Test piecewise CUDA graph with FP8 quantized model"""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = "nvidia/Llama-3.1-8B-Instruct-FP8"
+        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",
+                "modelopt_fp8",
+                "--kv-cache-dtype",
+                "bfloat16",
+            ],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_mgsm_accuracy(self):
+        """Test MGSM accuracy with FP8 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)
+        self.assertGreaterEqual(metrics["score"], 0.85)
+        print(f"MGSM Accuracy: {metrics['score']:.3f}")
+
+
 if __name__ == "__main__":
     unittest.main()