Support piecewise cuda graph for dsv3 fp4 (#15531)

test/srt/run_suite.py

@@ -163,7 +163,7 @@ suites = {
         TestFile("test_disaggregation_dp_attention.py", 155),
     ],
     "per-commit-4-gpu-b200-stage-b": [
-        TestFile("test_deepseek_v3_fp4_4gpu.py", 1800),  # Stage B test
+        TestFile("test_deepseek_v3_fp4_4gpu.py", 2000),  # Stage B test
     ],
     "per-commit-4-gpu-b200": [
         TestFile("test_flash_attention_4.py", 90),

test/srt/test_deepseek_v3_fp4_4gpu.py

@@ -176,5 +176,72 @@ class TestDeepseekV3FP4MTP(CustomTestCase):
         self.assertGreater(speed, 150)
 
 
+class TestDeepseekV3FP4PiecewiseCudaGraph(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = FULL_DEEPSEEK_V3_FP4_MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        other_args = [
+            "--tp",
+            "4",
+            "--attention-backend",
+            "trtllm_mla",
+            "--moe-runner-backend",
+            "flashinfer_trtllm",
+            "--quantization",
+            "modelopt_fp4",
+            "--enable-piecewise-cuda-graph",
+            "--kv-cache-dtype",
+            "fp8_e4m3",
+            "--model-loader-extra-config",
+            '{"enable_multithread_load": true,"num_threads": 64}',
+        ]
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=SERVER_LAUNCH_TIMEOUT,
+            other_args=other_args,
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_a_gsm8k(
+        self,
+    ):
+        args = SimpleNamespace(
+            num_shots=8,
+            data_path=None,
+            num_questions=1319,
+            parallel=1319,
+            max_new_tokens=512,
+            host="http://127.0.0.1",
+            port=int(self.base_url.split(":")[-1]),
+        )
+        metrics = run_eval_few_shot_gsm8k(args)
+        print(f"{metrics=}")
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_gsm8k (deepseek-v3-fp4)\n" f'{metrics["accuracy"]=:.3f}\n'
+            )
+
+        self.assertGreater(metrics["accuracy"], 0.935)
+
+    def test_bs_1_speed(self):
+        args = BenchArgs(port=int(self.base_url.split(":")[-1]), max_new_tokens=2048)
+        _, speed = send_one_prompt(args)
+
+        print(f"{speed=:.2f}")
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_bs_1_speed (deepseek-v3-fp4)\n" f"{speed=:.2f} token/s\n"
+            )
+
+        self.assertGreater(speed, 120)
+
+
 if __name__ == "__main__":
     unittest.main()