[CI] Deflake pooling shards with GPU teardown fixtures between tests (#57371)

Signed-off-by: khluu <khluu000@gmail.com>
Co-authored-by: Kimi <noreply@moonshot.cn>
This commit is contained in:
Kevin H. Luu
2026-09-19 11:30:02 +08:00
committed by GitHub
co-authored by Kimi
parent 2fd6268e6d
commit 9d75dd4544
5 changed files with 119 additions and 2 deletions
@@ -0,0 +1,34 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Pytest configuration for vLLM multimodal pooling tests."""
import gc
import pytest
import torch
from tests.utils import wait_for_memory_to_settle
from vllm.logger import init_logger
logger = init_logger(__name__)
@pytest.fixture(autouse=True)
def release_gpu_memory_between_tests():
"""Reclaim cached GPU memory between tests in this shard.
These tests start many vLLM/HF engines sequentially in one pytest
process; without explicit reclamation between tests, VRAM fragmentation
accumulates and later tests fail at engine startup ("Engine core
initialization failed").
"""
yield
gc.collect()
if torch.accelerator.is_available():
torch.accelerator.empty_cache()
try:
wait_for_memory_to_settle()
except ValueError as e:
# Longer-lived (class/module-scoped) engine fixtures may legitimately
# still hold VRAM; nothing further to reclaim in that case.
logger.info("Failed to clean GPU memory: %s", e)