[ROCm][CI] Explicitly tear down multimodal offline LLMs (#46999)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Andreas Karatzas
2026-06-29 07:59:24 +00:00
committed by GitHub
parent f6bb8682ee
commit db28ae2d07
6 changed files with 91 additions and 29 deletions
+2 -14
View File
@@ -1,19 +1,13 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import weakref
import pytest
from tests.entrypoints.multimodal.conftest import TEST_IMAGE_ASSETS
from vllm import LLM
from vllm.distributed import cleanup_dist_env_and_memory
@pytest.fixture(scope="function")
def vision_llm():
# pytest caches the fixture so we use weakref.proxy to
# enable garbage collection
llm = LLM(
def vision_llm(multimodal_llm_factory):
return multimodal_llm_factory(
model="microsoft/Phi-3.5-vision-instruct",
max_model_len=4096,
max_num_seqs=5,
@@ -23,12 +17,6 @@ def vision_llm():
seed=0,
)
yield weakref.proxy(llm)
del llm
cleanup_dist_env_and_memory()
@pytest.mark.parametrize(
"image_urls", [[TEST_IMAGE_ASSETS[0], TEST_IMAGE_ASSETS[1]]], indirect=True
@@ -69,6 +69,7 @@ def test_inject_into_mm_cache(
image_urls,
mm_processor_cache_type,
caplog_vllm,
multimodal_llm_factory,
):
"""Test that inject_into_mm_cache() injects pre-processed mm_kwargs into
the processor cache and MM cache hit metrics are updated correctly.
@@ -78,7 +79,7 @@ def test_inject_into_mm_cache(
2. Extract cached kwargs, call inject_into_mm_cache with a new hash,
then generate with a pre-rendered input -> verifies injection works
"""
llm = LLM(
llm = multimodal_llm_factory(
model="llava-hf/llava-1.5-7b-hf",
max_model_len=4096,
max_num_seqs=5,
@@ -145,11 +146,12 @@ def test_inject_into_mm_cache(
def test_inject_into_mm_cache_without_cache(
num_gpus_available,
image_urls,
multimodal_llm_factory,
):
"""Test that inject_into_mm_cache works gracefully when processor cache
is disabled (mm_processor_cache_gb=0). Should not crash.
"""
llm = LLM(
llm = multimodal_llm_factory(
model="llava-hf/llava-1.5-7b-hf",
max_model_len=4096,
max_num_seqs=5,
@@ -61,8 +61,9 @@ def test_mm_cache_stats(
image_urls,
mm_processor_cache_type,
caplog_vllm,
multimodal_llm_factory,
):
llm = LLM(
llm = multimodal_llm_factory(
model="llava-hf/llava-1.5-7b-hf",
max_model_len=4096,
max_num_seqs=5,
@@ -1,13 +1,11 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import weakref
import pytest
from tests.entrypoints.multimodal.conftest import managed_llm
from vllm import LLM, SamplingParams
from vllm.assets.image import ImageAsset
from vllm.distributed import cleanup_dist_env_and_memory
MODEL = "llava-hf/llava-1.5-7b-hf"
PROMPT = "USER: <image>\nDescribe this image briefly.\nASSISTANT:"
@@ -17,20 +15,15 @@ TEXT_ONLY_PROMPT = "USER: What is 2 + 2?\nASSISTANT:"
@pytest.fixture(scope="module")
def llm():
"""LLM with enable_mm_embeds=True and all modality limits zeroed out."""
llm = LLM(
with managed_llm(
model=MODEL,
max_model_len=2048,
enforce_eager=True,
gpu_memory_utilization=0.8,
enable_mm_embeds=True,
limit_mm_per_prompt={"image": 0},
)
yield weakref.proxy(llm)
del llm
cleanup_dist_env_and_memory()
) as llm:
yield llm
@pytest.mark.skip_global_cleanup