[11/N] MoE Refactor: Simplifying SBO Implementation with Dispatcher Hooks (#13327)

Co-authored-by: Baizhou Zhang <sobereddiezhang@gmail.com>

test/srt/run_suite.py

@@ -180,7 +180,7 @@ suites = {
     "per-commit-8-gpu-b200": [],
     "per-commit-4-gpu-gb200": [
         TestFile("test_cutedsl_moe.py", 300),
-        TestFile("test_deepseek_v3_cutedsl_4gpu.py", 3600),
+        TestFile("test_deepseek_v3_cutedsl_4gpu.py", 590),
     ],
     "per-commit-4-gpu-deepep": [
         TestFile("ep/test_deepep_small.py", 531),

test/srt/test_deepseek_v3_cutedsl_4gpu.py

@@ -14,6 +14,7 @@ from sglang.test.test_utils import (
 )
 
 
+@unittest.skip("See https://github.com/sgl-project/sglang/issues/12533")
 class TestDeepseekR1Nvfp4CuteDSLDeepEP(CustomTestCase):
     @classmethod
     def setUpClass(cls):
@@ -83,5 +84,79 @@ class TestDeepseekR1Nvfp4CuteDSLDeepEP(CustomTestCase):
         self.assertGreater(metrics["accuracy"], 0.92)
 
 
+class TestDummyWithSBO(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = try_cached_model(DEFAULT_DEEPSEEK_NVFP4_MODEL_FOR_TEST)
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        other_args = [
+            "--trust-remote-code",
+            "--disable-radix-cache",
+            "--mem-fraction-static",
+            "0.05",
+            "--max-prefill-tokens",
+            "16384",
+            "--max-running-requests",
+            "256",
+            "--chunked-prefill-size",
+            "1024",
+            "--cuda-graph-bs",
+            "64",
+            "--tp",
+            "4",
+            "--dp",
+            "4",
+            "--ep",
+            "4",
+            "--moe-dense-tp-size",
+            "1",
+            "--enable-dp-attention",
+            "--quantization",
+            "modelopt_fp4",
+            "--attention-backend",
+            "trtllm_mla",
+            "--moe-a2a-backend",
+            "deepep",
+            "--moe-runner-backend",
+            "flashinfer_cutedsl",
+            "--deepep-mode",
+            "low_latency",
+            "--json-model-override-args",
+            '{"num_hidden_layers": 1, "first_k_dense_replace": 0, "n_routed_experts": 24}',
+            "--enable-single-batch-overlap",
+            "--load-format",
+            "dummy",
+        ]
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=other_args,
+            env={
+                **os.environ,
+                "SGLANG_DEEPEP_BF16_DISPATCH": "1",
+                "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "256",
+                "SGLANG_CUTEDSL_MOE_NVFP4_DISPATCH": "0",
+            },
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_gsm8k(self):
+        args = SimpleNamespace(
+            num_shots=0,
+            data_path=None,
+            num_questions=512,
+            parallel=512,
+            max_new_tokens=16,
+            host="http://127.0.0.1",
+            port=int(self.base_url.split(":")[-1]),
+        )
+        metrics = run_eval_few_shot_gsm8k(args)
+        print(f"Eval accuracy of GSM8K: {metrics=}")
+
+
 if __name__ == "__main__":
     unittest.main()