[ROCm] Run init test engine in-process to avoid KV-cache OOM (#48527)
Signed-off-by: Djordje Ramic <djoramic@amd.com> Co-authored-by: Andreas Karatzas <akaratza@amd.com>
djramic committed
Jul 15, 2026 at 08:39 UTC
b6770d7b544febb9636f1750b77cde8904024728
1 file changed
+7
-1
tests/models/test_initialization.py
+7
-1
@@ -14,7 +14,7 @@ from vllm.v1.core.kv_cache_utils import (
14
)
15
from vllm.v1.engine.core import EngineCore as V1EngineCore
16
17
-from ..utils import create_new_process_for_each_test
17
+from ..utils import create_new_process_for_each_test, requires_spawn_multiprocessing
18
from .registry import (
19
_TRANSFORMERS_BACKEND_MODELS,
20
AUTO_EXAMPLE_MODELS,
@@ -129,6 +129,12 @@ def can_initialize(
129
patch.object(V1EngineCore, "_initialize_kv_caches", _initialize_kv_caches_v1),
130
monkeypatch.context() as m,
131
):
132
+ if requires_spawn_multiprocessing():
133
+ # The EngineCore subprocess re-imports the class and does not
134
+ # inherit the KV-cache patch above, so it OOMs. Run in-process
135
+ # so the patch applies.
136
+ m.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
137
+
138
# FIXME: A hack to bypass FA3 assertion because our CI's L4 GPU
139
# has cc==8.9 which hasn't supported FA3 yet. Remove this hack when
140
# L4 supports FA3.