Add Qwen3 MoE tests for PP compatibility with CP and DP (#29640)

Signed-off-by: Shangming Cai <csmthu@gmail.com>
2026-06-29 18:52:03 +08:00
parent a2b5ce2ed1
commit 93926b9a95
1 changed files with 111 additions and 0 deletions
@@ -0,0 +1,111 @@
+import unittest
+from types import SimpleNamespace
+
+from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.run_eval import run_eval
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    kill_process_tree,
+    popen_launch_server,
+)
+
+register_cuda_ci(est_time=600, stage="extra-b", runner_config="4-gpu-h100")
+
+QWEN3_MOE_MODEL_PATH = "Qwen/Qwen3-30B-A3B-FP8"
+
+GSM8K_BASELINE_ACCURACY = 0.93
+
+
+class _Qwen3MoePPCompatMixin:
+    """Launch a Qwen3 MoE server combining PP with another parallel strategy and
+    check GSM8K accuracy. Concrete subclasses set ``parallel_args`` (and
+    optionally ``server_env``); this mixin is not a TestCase so it is never
+    collected on its own.
+    """
+
+    model = QWEN3_MOE_MODEL_PATH
+    parallel_args: list = []
+    server_env = None
+
+    @classmethod
+    def setUpClass(cls):
+        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=[
+                *cls.parallel_args,
+                "--cuda-graph-max-bs-decode",
+                "32",
+                "--max-running-requests",
+                "32",
+                "--trust-remote-code",
+                "--disable-piecewise-cuda-graph",
+                "--model-loader-extra-config",
+                '{"enable_multithread_load": true, "num_threads": 64}',
+            ],
+            env=cls.server_env,
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        if hasattr(cls, "process") and cls.process:
+            kill_process_tree(cls.process.pid)
+
+    def test_gsm8k(self):
+        args = SimpleNamespace(
+            model=self.model,
+            eval_name="gsm8k",
+            num_shots=5,
+            num_examples=200,
+            max_tokens=16000,
+            num_threads=128,
+            repeat=1,
+            temperature=0.6,
+            top_p=0.95,
+            top_k=20,
+            base_url=self.base_url,
+        )
+        metrics = run_eval(args)
+        print(f"{metrics=}")
+        self.assertGreaterEqual(metrics["score"], GSM8K_BASELINE_ACCURACY)
+
+
+class TestQwen3MoePPxCP(_Qwen3MoePPCompatMixin, CustomTestCase):
+    """PP x CP: pp_size=2 x attn_cp_size=2 (tp_size=2, moe_dp_size=1)."""
+
+    parallel_args = [
+        "--tp-size",
+        "2",
+        "--pp-size",
+        "2",
+        "--ep-size",
+        "2",
+        "--attn-cp-size",
+        "2",
+        "--enable-prefill-cp",
+        "--cp-strategy",
+        "zigzag",
+    ]
+    server_env = {"SGLANG_ENABLE_CP_V2": "1"}
+
+
+class TestQwen3MoePPxDP(_Qwen3MoePPCompatMixin, CustomTestCase):
+    """PP x DP: pp_size=2 x dp_size=2 attention (tp_size=2)."""
+
+    parallel_args = [
+        "--tp-size",
+        "2",
+        "--pp-size",
+        "2",
+        "--dp-size",
+        "2",
+        "--enable-dp-attention",
+    ]
+
+
+if __name__ == "__main__":
+    unittest.main()