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

Signed-off-by: Micah Williamson <micah.williamson@amd.com>
This commit is contained in:
Micah Williamson
2026-06-16 10:00:15 -05:00
committed by GitHub
parent 506ec6d656
commit 6f612fbedf
+15
View File
@@ -51,6 +51,7 @@ from vllm import LLM, SamplingParams, envs
from vllm.assets.audio import AudioAsset from vllm.assets.audio import AudioAsset
from vllm.assets.image import ImageAsset from vllm.assets.image import ImageAsset
from vllm.assets.video import VideoAsset from vllm.assets.video import VideoAsset
from vllm.config.cache import CacheConfig
from vllm.config.model import ConvertOption, RunnerOption, _get_and_verify_dtype from vllm.config.model import ConvertOption, RunnerOption, _get_and_verify_dtype
from vllm.connections import global_http_connection from vllm.connections import global_http_connection
from vllm.distributed import ( from vllm.distributed import (
@@ -924,6 +925,20 @@ class VllmRunner:
num_speculative_tokens + 1 num_speculative_tokens + 1
) )
from vllm.platforms import current_platform
if current_platform.is_rocm():
gpu_memory_utilization = kwargs.get(
"gpu_memory_utilization",
CacheConfig.gpu_memory_utilization,
)
# V1 startup requires free_memory >= total * gpu_memory_utilization.
# ROCm CI can hand a test a device that is still lazily releasing
# VRAM from a previous process, so wait before constructing LLM.
from tests.utils import wait_for_rocm_memory_to_settle
wait_for_rocm_memory_to_settle(threshold_ratio=1.0 - gpu_memory_utilization)
with init_ctx: with init_ctx:
self.llm = LLM( self.llm = LLM(
model=model_name, model=model_name,