[ROCm][CI] Patch conftest to resolve occasional OOMs (#45722)

Signed-off-by: Micah Williamson <micah.williamson@amd.com>

Micah Williamson committed Jun 16, 2026 at 10:00 UTC 6f612fbedff718af2dabb93692f00044e66a9b4b

1 file changed +15

tests/conftest.py

+15

@@ -51,6 +51,7 @@ from vllm import LLM, SamplingParams, envs

51 from vllm.assets.audio import AudioAsset

52 from vllm.assets.image import ImageAsset

53 from vllm.assets.video import VideoAsset

54 +from vllm.config.cache import CacheConfig

55 from vllm.config.model import ConvertOption, RunnerOption, _get_and_verify_dtype

56 from vllm.connections import global_http_connection

57 from vllm.distributed import (

@@ -924,6 +925,20 @@ class VllmRunner:

925 num_speculative_tokens + 1

926 )

927

928 + from vllm.platforms import current_platform

929 +

930 + if current_platform.is_rocm():

931 + gpu_memory_utilization = kwargs.get(

932 + "gpu_memory_utilization",

933 + CacheConfig.gpu_memory_utilization,

934 + )

935 + # V1 startup requires free_memory >= total * gpu_memory_utilization.

936 + # ROCm CI can hand a test a device that is still lazily releasing

937 + # VRAM from a previous process, so wait before constructing LLM.

938 + from tests.utils import wait_for_rocm_memory_to_settle

939 +

940 + wait_for_rocm_memory_to_settle(threshold_ratio=1.0 - gpu_memory_utilization)

941 +

942 with init_ctx:

943 self.llm = LLM(

944 model=model_name,