Add gpu memory wait before test_async_tp (#28893)

Signed-off-by: angelayi <yiangela7@gmail.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
This commit is contained in:
Angela Yi
2025-11-28 20:19:05 -08:00
committed by GitHub
co-authored by Luka Govedič
parent e23f665d83
commit 4b17ce6815
2 changed files with 32 additions and 3 deletions
+29
View File
@@ -1424,3 +1424,32 @@ def disable_deepgemm_ue8m0(monkeypatch):
# Clear cache so the next time it is used it is processed with the
# default VLLM_USE_DEEP_GEMM_E8M0 setting.
is_deep_gemm_e8m0_used.cache_clear()
@pytest.fixture(autouse=True)
def clean_gpu_memory_between_tests():
if os.getenv("VLLM_TEST_CLEAN_GPU_MEMORY", "0") != "1":
yield
return
# Wait for GPU memory to be cleared before starting the test
import gc
from tests.utils import wait_for_gpu_memory_to_clear
num_gpus = torch.cuda.device_count()
if num_gpus > 0:
try:
wait_for_gpu_memory_to_clear(
devices=list(range(num_gpus)),
threshold_ratio=0.1,
)
except ValueError as e:
logger.info("Failed to clean GPU memory: %s", e)
yield
# Clean up GPU memory after the test
if torch.cuda.is_available():
torch.cuda.empty_cache()
gc.collect()