[BCG][GLM5] perf: BCG support and prefill enhancements (#27053)

2026-06-24 13:13:35 -07:00
parent d46afbf8b4
commit d5e9176f65
7 changed files with 695 additions and 225 deletions
@@ -0,0 +1,75 @@
+import unittest
+from types import SimpleNamespace
+
+from sglang.srt.utils import kill_process_tree
+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,
+    popen_launch_server,
+)
+
+register_cuda_ci(est_time=900, stage="base-c", runner_config="8-gpu-h200")
+
+GLM5_FP8_MODEL = "zai-org/GLM-5-FP8"
+
+
+class TestBCGGlm5Fp8TP8(CustomTestCase):
+    """Breakable CUDA graph prefill on GLM-5-FP8 (DSA model, TP=8, H200)."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = GLM5_FP8_MODEL
+        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=[
+                "--tp-size",
+                "8",
+                "--trust-remote-code",
+                "--reasoning-parser",
+                "glm45",
+                "--tool-call-parser",
+                "glm47",
+                "--mem-fraction-static",
+                "0.8",
+                "--disable-flashinfer-autotune",
+                "--cuda-graph-backend-prefill=breakable",
+                # Small chunks => many prefill iterations, each <= the 2048
+                # capture max, so every prefill batch replays the BCG graph and
+                # exercises the DSA split-op / dual-stream / MLA-fusion paths.
+                "--chunked-prefill-size",
+                "512",
+                "--model-loader-extra-config",
+                '{"enable_multithread_load": true, "num_threads": 64}',
+            ],
+            env={
+                "SGLANG_ENABLE_PCG_DSV2_DUAL_STREAM": "1",
+            },
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        if hasattr(cls, "process") and cls.process:
+            kill_process_tree(cls.process.pid)
+
+    def test_gsm8k(self):
+        args = SimpleNamespace(
+            base_url=self.base_url,
+            model=self.model,
+            eval_name="gsm8k",
+            num_examples=200,
+            num_threads=200,
+            max_tokens=4096,
+        )
+        metrics = run_eval(args)
+        print(f"{metrics=}")
+        self.assertGreater(metrics["score"], 0.92)
+
+
+if __name__ == "__main__":
+    unittest.main()