From 4792ab1e901032c85682d4d1e81150ec8a165430 Mon Sep 17 00:00:00 2001 From: Liangsheng Yin Date: Sat, 8 Aug 2026 20:07:19 -0700 Subject: [PATCH] [CI] Pin the rust frontend parity test to eager prefill (#34146) --- .../openai_server/basic/test_openai_completion_rust.py | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-) diff --git a/test/registered/openai_server/basic/test_openai_completion_rust.py b/test/registered/openai_server/basic/test_openai_completion_rust.py index 7c51f67b0..87c698098 100644 --- a/test/registered/openai_server/basic/test_openai_completion_rust.py +++ b/test/registered/openai_server/basic/test_openai_completion_rust.py @@ -26,13 +26,19 @@ class TestOpenAICompletionRustParity(CustomTestCase): api_key = "sk-123456" def _get_logprobs(self, *, rust_frontend): + # Prefill CUDA graph pads the batch, so numerics follow whichever + # requests share the forward pass; the assertions below need equality. process = popen_launch_server( self.model, DEFAULT_URL_FOR_TEST, timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH, api_key=self.api_key, env={"SGLANG_RUST_SERVER": "1" if rust_frontend else "0"}, - other_args=["--random-seed", "42"], + other_args=[ + "--random-seed", + "42", + "--disable-prefill-cuda-graph", + ], ) try: response = requests.post(