Split test_piecewise_cuda_graph.py to optimize CI resource usage (#15290)

2025-12-17 21:20:21 -08:00
parent 3d42b7e7b0
commit 58c840db6b
4 changed files with 248 additions and 253 deletions
@@ -0,0 +1,57 @@
+import unittest
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.run_eval import run_eval
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    SimpleNamespace,
+    popen_launch_server,
+)
+
+
+class TestPiecewiseCudaGraphQwen3OmniMOE(CustomTestCase):
+    """Test piecewise CUDA graph with Qwen3-Omni-30B-A3B-Instruct model"""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = "Qwen/Qwen3-Omni-30B-A3B-Instruct"
+        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",
+                "--piecewise-cuda-graph-compiler",
+                "eager",
+                "--disable-radix-cache",
+                "--tp=2",
+            ],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_gsm8k_accuracy(self):
+        """Test GSM8K accuracy with 8-shot setting"""
+        num_examples = 2000
+
+        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"GSM8K Accuracy: {metrics['score']:.3f}")
+
+        self.assertGreaterEqual(metrics["score"], 0.70)
+
+
+if __name__ == "__main__":
+    unittest.main()