diff --git a/python/sglang/test/test_utils.py b/python/sglang/test/test_utils.py index 0e26b57ca3..48e98eb88f 100644 --- a/python/sglang/test/test_utils.py +++ b/python/sglang/test/test_utils.py @@ -2385,6 +2385,9 @@ def enter_scope(test_case, scope): class CustomTestCase(unittest.TestCase): + # Fault-injection tests must inspect the failure after their server exits. + expect_server_exit = False + def __init_subclass__(cls, **kwargs): super().__init_subclass__(**kwargs) @@ -2417,6 +2420,9 @@ class CustomTestCase(unittest.TestCase): cls.setUpClass = classmethod(safe_setUpClass) def _callTestMethod(self, method): + if self.expect_server_exit: + # Retrying would reuse the server that the test intentionally stopped. + return super()._callTestMethod(method) max_retry = envs.SGLANG_TEST_MAX_RETRY.get() if max_retry is None: max_retry = 1 if is_in_ci() else 0 diff --git a/test/registered/kv_canary/test_self_e2e_perturb_raise.py b/test/registered/kv_canary/test_self_e2e_perturb_raise.py index fc66f0d2fe..659c595d1b 100644 --- a/test/registered/kv_canary/test_self_e2e_perturb_raise.py +++ b/test/registered/kv_canary/test_self_e2e_perturb_raise.py @@ -12,6 +12,7 @@ register_amd_ci(est_time=50, stage="extra-a", runner_config="1-gpu-small-amd") class TestPerturbRaiseMha(CanaryE2EBase): + expect_server_exit = True model_mode = "mha" kv_canary_mode = CanaryMode.RAISE extra_server_args = ("--kv-canary-real-data", "partial", "--skip-server-warmup")