[CI][Bugfix] Rerun test_engine_log_metrics_ray on Ray GCS startup timeout (#47208)
Signed-off-by: pei.zhang <pei.zhang@amd.com> Co-authored-by: Claude <noreply@anthropic.com>
peizhang56 committed
Jul 1, 2026 at 19:32 UTC
7fe7fa9cda6f38d27b2f727e65f3aa172021c389
1 file changed
+13
tests/v1/metrics/test_ray_metrics.py
+13
@@ -7,6 +7,7 @@ import pytest
7
import ray
8
9
from vllm.config.model import ModelDType
10
+from vllm.platforms import current_platform
11
from vllm.sampling_params import SamplingParams
12
from vllm.v1.engine.async_llm import AsyncEngineArgs, AsyncLLM
13
from vllm.v1.metrics.ray_wrappers import (
@@ -22,6 +23,18 @@ MODELS = [
23
]
24
25
26
+# The first .remote() call starts a local Ray cluster via ray.init(), whose
27
+# GCS server occasionally fails to start within Ray's fixed 30s bootstrap
28
+# window on ROCm CI (RuntimeError: "Timed out waiting for file
29
+# .../gcs_server_port_..."). That timeout is not configurable, so retry the
30
+# whole test: a fresh ray.init() gets a new GCS process. Scoped to that error
31
+# so real failures still fail immediately.
32
+@pytest.mark.flaky(
33
+ reruns=2,
34
+ reruns_delay=5,
35
+ only_rerun="Timed out waiting for file",
36
+ condition=current_platform.is_rocm(),
37
+)
38
@pytest.mark.parametrize("model", MODELS)
39
@pytest.mark.parametrize("dtype", ["half"])
40
@pytest.mark.parametrize("max_tokens", [16])