[Bugfix][V1] Clean up compiled-model bytecode hooks on VllmRunner exit (#45195)

Signed-off-by: Ting Sun <suntcrick@gmail.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Ting SUN
2026-06-16 20:31:17 -07:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 14b438a98b
commit 7b5d60cc37
4 changed files with 61 additions and 2 deletions
+1
View File
@@ -1294,6 +1294,7 @@ class VllmRunner:
# Ignore shutdown errors as cleanup will still proceed
pass
del self.llm
torch._dynamo.reset()
cleanup_dist_env_and_memory()
self._wait_for_rocm_memory_release(gpu_memory_utilization)
+28 -1
View File
@@ -4,7 +4,8 @@
import pytest
from tests.utils import wait_for_gpu_memory_to_clear
from tests.conftest import VllmRunner
from tests.utils import create_new_process_for_each_test, wait_for_gpu_memory_to_clear
from tests.v1.shutdown.utils import (
SHUTDOWN_TEST_THRESHOLD_BYTES,
SHUTDOWN_TEST_TIMEOUT_SEC,
@@ -106,3 +107,29 @@ def test_llm_delete(
devices=list(range(tensor_parallel_size)),
threshold_bytes=SHUTDOWN_TEST_THRESHOLD_BYTES,
)
@create_new_process_for_each_test("fork" if current_platform.is_cuda() else "spawn")
@pytest.mark.timeout(SHUTDOWN_TEST_TIMEOUT_SEC)
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("send_one_request", [False, True])
def test_llm_delete_inprocess(
monkeypatch,
model: str,
send_one_request: bool,
) -> None:
"""Test that VllmRunner frees GPU memory in in-process (no MP) mode."""
with monkeypatch.context() as m:
m.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
with VllmRunner(model) as vllm_model:
if send_one_request:
vllm_model.generate(
["Hello my name is"],
SamplingParams(max_tokens=1),
)
wait_for_gpu_memory_to_clear(
devices=[0],
threshold_bytes=SHUTDOWN_TEST_THRESHOLD_BYTES,
)