Piecewise Cuda Graph set default (#16331)

2026-03-02 23:18:07 +08:00
parent 468e3dc56b
commit c64274c746
34 changed files with 505 additions and 143 deletions
@@ -111,7 +111,14 @@ class TestMLADeepseekV3Fa3Fp8Kvcache(CustomTestCase):
        ]
        if is_cuda():
            other_args.extend(
-                ["--attention-backend", "fa3", "--cuda-graph-max-bs", "2"]
+                [
+                    "--attention-backend",
+                    "fa3",
+                    "--mem-fraction-static",
+                    "0.8",
+                    "--cuda-graph-max-bs",
+                    "2",
+                ]
            )
        cls.process = popen_launch_server(
            cls.model,