[CI] Remove VLLM_TEST_CLEAN_GPU_MEMORY to avoid environment variable pollution that causes unnecessary GPU detection, thereby slowing down test case execution. (#3446)

Signed-off-by: wangyu <410167048@qq.com>
Signed-off-by: [Your Name] <your.email@example.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>

.buildkite/test-amd-merge.yml

@@ -123,7 +123,6 @@ steps:
   grade: Blocking
   commands:
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - export VLLM_TEST_CLEAN_GPU_MEMORY=1
     - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py tests/e2e/online_serving/test_qwen3_omni.py tests/e2e/online_serving/test_mimo_audio.py -m "advanced_model" --run-level "advanced_model"
 
 - label: "Qwen3-TTS CustomVoice E2E Test"
@@ -164,7 +163,6 @@ steps:
 #   grade: Blocking
 #   commands:
 #     - export GPU_ARCHS=gfx942
-#     - export VLLM_TEST_CLEAN_GPU_MEMORY=1
 #     - export VLLM_LOGGING_LEVEL=DEBUG
 #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
 #     - export VLLM_ROCM_USE_AITER_RMSNORM=0
@@ -177,7 +175,6 @@ steps:
 #   grade: Blocking
 #   commands:
 #     - export GPU_ARCHS=gfx942
-#     - export VLLM_TEST_CLEAN_GPU_MEMORY=1
 #     - export VLLM_LOGGING_LEVEL=DEBUG
 #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
 #     - export VLLM_ROCM_USE_AITER_RMSNORM=0
@@ -190,7 +187,6 @@ steps:
 #   grade: Blocking
 #   commands:
 #     - export GPU_ARCHS=gfx942
-#     - export VLLM_TEST_CLEAN_GPU_MEMORY=1
 #     - export VLLM_IMAGE_FETCH_TIMEOUT=60
 #     - export VLLM_LOGGING_LEVEL=DEBUG
 #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn

.buildkite/test-amd-ready.yaml

@@ -114,7 +114,6 @@ steps:
 #   commands:
 #     - export VLLM_LOGGING_LEVEL=DEBUG
 #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-#     - export VLLM_TEST_CLEAN_GPU_MEMORY=1
 #     - timeout 10m pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py
 #     - timeout 20m pytest -s -v tests/e2e/online_serving/test_qwen3_omni.py -m "core_model" --run-level "core_model"
 
@@ -165,7 +164,6 @@ steps:
 #   grade: Blocking
 #   commands:
 #     - export GPU_ARCHS=gfx942
-#     - export VLLM_TEST_CLEAN_GPU_MEMORY=1
 #     - export VLLM_LOGGING_LEVEL=DEBUG
 #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
 #     - export VLLM_ROCM_USE_AITER_RMSNORM=0
@@ -178,7 +176,6 @@ steps:
 #   grade: Blocking
 #   commands:
 #     - export GPU_ARCHS=gfx942
-#     - export VLLM_TEST_CLEAN_GPU_MEMORY=1
 #     - export VLLM_LOGGING_LEVEL=DEBUG
 #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
 #     - export VLLM_ROCM_USE_AITER_RMSNORM=0
@@ -191,7 +188,6 @@ steps:
 #   grade: Blocking
 #   commands:
 #     - export GPU_ARCHS=gfx942
-#     - export VLLM_TEST_CLEAN_GPU_MEMORY=1
 #     - export VLLM_IMAGE_FETCH_TIMEOUT=60
 #     - export VLLM_LOGGING_LEVEL=DEBUG
 #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn

.buildkite/test-amd.yaml

@@ -100,7 +100,6 @@ steps:
   commands:
     - export VLLM_LOGGING_LEVEL=DEBUG
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - export VLLM_TEST_CLEAN_GPU_MEMORY="1"
     - export VLLM_ROCM_USE_AITER=0
     - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py
     - pytest -s -v tests/e2e/online_serving/test_qwen3_omni.py
@@ -116,5 +115,4 @@ steps:
     - export GPU_ARCHS=gfx942
     - export VLLM_LOGGING_LEVEL=DEBUG
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - export VLLM_TEST_CLEAN_GPU_MEMORY="1"
     - pytest -s -v tests/entrypoints/test_omni_sleep_mode.py -m "advanced_model and omni and MI325" --run-level "advanced_model"

.buildkite/test-merge.yml

@@ -7,25 +7,9 @@ steps:
   - group: ":card_index_dividers: Simple Test"
     depends_on: upload-merge-pipeline
     steps:
-      - label: "Simple · Diffusion Test"
+      - label: "Simple · Diffusion & Model Executor Test"
         commands:
-          - "timeout 20m pytest -sv tests/diffusion -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
-        agents:
-          queue: "gpu_1_queue"
-        plugins:
-          - docker#v5.2.0:
-              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-              always-pull: true
-              propagate-environment: true
-              environment:
-                - "HF_HOME=/fsx/hf_cache"
-                - "HF_TOKEN"
-              volumes:
-                - "/fsx/hf_cache:/fsx/hf_cache"
-
-      - label: "Simple · Model Executor Test"
-        commands:
-          - "timeout 20m pytest -sv tests/model_executor -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+          - "timeout 20m pytest -sv tests/diffusion tests/model_executor -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
         agents:
           queue: "gpu_1_queue"
         plugins:
@@ -81,7 +65,6 @@ steps:
       - |
         timeout 55m bash -c "
           set -e
-          export VLLM_TEST_CLEAN_GPU_MEMORY=1
           export VLLM_IMAGE_FETCH_TIMEOUT=60
           pytest -s -v tests/distributed -m 'advanced_model and cuda' --run-level 'advanced_model'
         "

.buildkite/test-nightly.yml

@@ -534,7 +534,6 @@ steps:
       - label: ":full_moon: Diffusion X2I(&A&T) · Doc Test"
         timeout_in_minutes: 60
         commands:
-          - export VLLM_TEST_CLEAN_GPU_MEMORY="1"
           - pytest -s -v tests/examples/*/test_text_to_image.py -m "full_model and example and H100" --run-level "full_model"
         agents:
           queue: "mithril-h100-pool"

.buildkite/test-ready.yml

@@ -7,25 +7,9 @@ steps:
   - group: ":card_index_dividers: Simple Test"
     depends_on: upload-ready-pipeline
     steps:
-      - label: "Simple · Diffusion Test"
+      - label: "Simple · Diffusion & Model Executor Test"
         commands:
-          - "timeout 20m pytest -sv tests/diffusion -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
-        agents:
-          queue: "gpu_1_queue"
-        plugins:
-          - docker#v5.2.0:
-              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-              always-pull: true
-              propagate-environment: true
-              environment:
-                - "HF_HOME=/fsx/hf_cache"
-                - "HF_TOKEN"
-              volumes:
-                - "/fsx/hf_cache:/fsx/hf_cache"
-
-      - label: "Simple · Model Executor Test"
-        commands:
-          - "timeout 20m pytest -sv tests/model_executor -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+          - "timeout 20m pytest -sv tests/diffusion tests/model_executor -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
         agents:
           queue: "gpu_1_queue"
         plugins:
@@ -374,7 +358,6 @@ steps:
         commands:
           - |
             timeout 40m bash -c "
-              export VLLM_TEST_CLEAN_GPU_MEMORY=1
               export VLLM_IMAGE_FETCH_TIMEOUT=60
               pytest -s -v tests/e2e/online_serving/test_bagel.py -m 'core_model' --run-level 'core_model'
             "
@@ -416,7 +399,6 @@ steps:
         commands:
           - |
             timeout 40m bash -c "
-              export VLLM_TEST_CLEAN_GPU_MEMORY=1
               export VLLM_IMAGE_FETCH_TIMEOUT=60
               pytest -s -v tests/e2e/online_serving/test_wan22_t2v.py -m 'core_model' --run-level 'core_model'
             "
@@ -459,7 +441,6 @@ steps:
         commands:
           - |
             timeout 40m bash -c "
-              export VLLM_TEST_CLEAN_GPU_MEMORY=1
               export VLLM_IMAGE_FETCH_TIMEOUT=60
               pytest -s -v tests/e2e/online_serving/test_qwen_image.py -m 'core_model' --run-level 'core_model'
             "
@@ -501,7 +482,6 @@ steps:
         commands:
           - |
             timeout 40m bash -c "
-              export VLLM_TEST_CLEAN_GPU_MEMORY=1
               export VLLM_IMAGE_FETCH_TIMEOUT=60
               pytest -s -v tests/e2e/online_serving/test_qwen_image_layered.py -m 'core_model' --run-level 'core_model'
             "

docs/contributing/ci/tests_style.md

@@ -213,7 +213,6 @@ E2E Online tests for Qwen3-Omni model with mix input and audio+text output.
 import os
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 import threading
 from pathlib import Path

recipes/Tencent-Hunyuan/HunyuanImage-3.0-Instruct.md

@@ -186,9 +186,9 @@ Validated score summary:
 
 The CI assertion threshold is `overall_mean >= 0.45`, so the smoke result is
 comfortably above the gate. The generate server and judge server run
-sequentially through the `OmniServer` fixture, and
-`VLLM_TEST_CLEAN_GPU_MEMORY=1` is used to wait for memory cleanup between
-server lifetimes.
+sequentially through the `OmniServer` fixture, with GPU memory cleanup
+between server lifetimes (for example via the `clean_gpu_memory_between_tests`
+pytest fixture in the smoke path).
 
 The lower-cost 2-GPU Instruct setup was tried for this smoke path but did not
 fit in memory. A previous 2-GPU experiment used the base HunyuanImage-3.0

tests/dfx/perf/scripts/run_benchmark.py

@@ -21,7 +21,6 @@ pytestmark = [pytest.mark.full_model]
 
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 
 def _get_config_file_from_argv() -> str | None:

tests/dfx/perf/scripts/run_diffusion_benchmark.py

@@ -36,7 +36,6 @@ import pytest
 pytestmark = [pytest.mark.diffusion, pytest.mark.full_model]
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 os.environ.setdefault("DIFFUSION_ATTENTION_BACKEND", "FLASH_ATTN")
 
 # ---------------------------------------------------------------------------

tests/diffusion/distributed/test_vae_decode_parallelism.py

@@ -21,7 +21,6 @@ import time
 from tests.helpers.runtime import OmniRunner
 from vllm_omni.platforms import current_omni_platform
 
-# os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "1"
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
 
 all_model_configs = [

tests/distributed/omni_connectors/test_bagel_mooncake_connector.py

@@ -26,6 +26,8 @@ from tests.helpers.runtime import OmniRunner
 from tests.helpers.stage_config import get_deploy_config_path, modify_stage_config
 from vllm_omni.entrypoints.omni import Omni
 
+pytestmark = [pytest.mark.usefixtures("clean_gpu_memory_between_tests")]
+
 BAGEL_MOONCAKE_CI_DEPLOY = get_deploy_config_path("ci/bagel_mooncake.yaml")
 
 # Reference pixel data extracted from the known-good output image

tests/distributed/omni_connectors/test_bagel_shared_memory_connector.py

@@ -26,6 +26,8 @@ from tests.helpers.stage_config import get_deploy_config_path, modify_stage_conf
 from vllm_omni.entrypoints.omni import Omni
 from vllm_omni.platforms import current_omni_platform
 
+pytestmark = [pytest.mark.usefixtures("clean_gpu_memory_between_tests")]
+
 BAGEL_CI_DEPLOY = get_deploy_config_path("ci/bagel.yaml")
 
 # Reference pixel data extracted from the known-good output image
@@ -234,24 +236,6 @@ def _generate_bagel_text2img(omni: Omni, prompt: str = TEXT2IMG_DEFAULT_PROMPT)
     return generated_image
 
 
-def _generate_bagel_text2img(omni: Omni, prompt: str = TEXT2IMG_DEFAULT_PROMPT) -> Image.Image:
-    """Generate an image using Bagel text2img with configured parameters."""
-    params_list = _configure_sampling_params(omni)
-
-    omni_outputs = list(
-        omni.generate(
-            prompts=[{"prompt": prompt, "modalities": ["image"]}],
-            sampling_params_list=params_list,
-        )
-    )
-
-    generated_image = _extract_generated_image(omni_outputs)
-    assert generated_image is not None, "No images generated"
-    assert generated_image.size == (1024, 1024), f"Expected 1024x1024, got {generated_image.size}"
-
-    return generated_image
-
-
 def _resolve_deploy_config(config_path: str, run_level: str) -> str:
     """Resolve deploy config based on run level.
 

tests/e2e/accuracy/qwen3_omni/test_qwen3_omni.py

@@ -111,7 +111,6 @@ def _daily_omni_hub_inline_media(monkeypatch: pytest.MonkeyPatch) -> None:
 
     monkeypatch.setattr(_acc_bench, "daily_omni_bench_argv", _wrapped)
     monkeypatch.setenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
-    monkeypatch.setenv("VLLM_TEST_CLEAN_GPU_MEMORY", "0")
 
 
 @hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)

tests/e2e/accuracy/test_diffusers_backend_similarity.py

@@ -119,7 +119,7 @@ def _run_diffusers_wan22_i2v(*, model: str, output_path: Path, conditioning_imag
     from diffusers import WanImageToVideoPipeline  # pyright: ignore[reportPrivateImportUsage]
     from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler
 
-    run_pre_test_cleanup(enable_force=True)
+    run_pre_test_cleanup()
     apply_ftfy_mock()
     pipe: WanImageToVideoPipeline | None = None
     try:
@@ -161,7 +161,7 @@ def _run_diffusers_wan22_i2v(*, model: str, output_path: Path, conditioning_imag
         gc.collect()
         if torch.cuda.is_available():
             torch.accelerator.empty_cache()
-        run_post_test_cleanup(enable_force=True)
+        run_post_test_cleanup()
 
 
 def _run_vllm_omni_qwen_image(*, model: str, output_path: Path) -> tuple[Image.Image, float]:
@@ -206,7 +206,7 @@ def _run_vllm_omni_qwen_image(*, model: str, output_path: Path) -> tuple[Image.I
 
 
 def _run_diffusers_qwen_image(*, model: str, output_path: Path) -> tuple[Image.Image, float]:
-    run_pre_test_cleanup(enable_force=True)
+    run_pre_test_cleanup()
     pipe: DiffusionPipeline | None = None
     try:
         pipe = DiffusionPipeline.from_pretrained(
@@ -241,7 +241,7 @@ def _run_diffusers_qwen_image(*, model: str, output_path: Path) -> tuple[Image.I
         gc.collect()
         if torch.cuda.is_available():
             torch.accelerator.empty_cache()
-        run_post_test_cleanup(enable_force=True)
+        run_post_test_cleanup()
 
 
 @pytest.mark.benchmark

tests/e2e/accuracy/test_ltx2_3_video_similarity.py

@@ -172,7 +172,7 @@ def _run_diffusers_baseline(model: str, output_dir: Path) -> list[Image.Image]:
     """Generate video using stock diffusers LTX2Pipeline."""
     from diffusers import LTX2Pipeline
 
-    run_pre_test_cleanup(enable_force=True)
+    run_pre_test_cleanup()
     pipe = None
     try:
         pipe = LTX2Pipeline.from_pretrained(
@@ -200,7 +200,7 @@ def _run_diffusers_baseline(model: str, output_dir: Path) -> list[Image.Image]:
         gc.collect()
         if torch.cuda.is_available():
             torch.accelerator.empty_cache()
-        run_post_test_cleanup(enable_force=True)
+        run_post_test_cleanup()
 
 
 # ---------------------------------------------------------------------------

tests/e2e/accuracy/test_qwen_image.py

@@ -70,7 +70,7 @@ def _run_vllm_omni_qwen_image(*, model: str, output_path: Path) -> Image.Image:
 
 
 def _run_diffusers_qwen_image(*, model: str, output_path: Path) -> Image.Image:
-    run_pre_test_cleanup(enable_force=True)
+    run_pre_test_cleanup()
     pipe: DiffusionPipeline | None = None
     try:
         pipe = DiffusionPipeline.from_pretrained(
@@ -99,7 +99,7 @@ def _run_diffusers_qwen_image(*, model: str, output_path: Path) -> Image.Image:
         gc.collect()
         if torch.cuda.is_available():
             torch.accelerator.empty_cache()
-        run_post_test_cleanup(enable_force=True)
+        run_post_test_cleanup()
 
 
 @pytest.mark.benchmark

tests/e2e/accuracy/test_qwen_image_edit.py

@@ -77,7 +77,7 @@ def _run_diffusers_image_edit(
     input_images: list[Image.Image],
     output_path: Path,
 ) -> Image.Image:
-    run_pre_test_cleanup(enable_force=True)
+    run_pre_test_cleanup()
     pipe: QwenImageEditPipeline | QwenImageEditPlusPipeline | None = None
     device = torch.device("cuda:0")
     torch.cuda.set_device(device)
@@ -110,7 +110,7 @@ def _run_diffusers_image_edit(
         gc.collect()
         if torch.cuda.is_available():
             torch.accelerator.empty_cache()
-        run_post_test_cleanup(enable_force=True)
+        run_post_test_cleanup()
 
 
 def _vllm_omni_output_single_image(

tests/e2e/accuracy/test_qwen_image_layered.py

@@ -93,7 +93,7 @@ def _run_vllm_omni_qwen_image_layered(*, model: str, input_image: Image.Image, o
 
 
 def _run_diffusers_qwen_image_layered(*, model: str, input_image: Image.Image, output_dir: Path) -> list[Image.Image]:
-    run_pre_test_cleanup(enable_force=True)
+    run_pre_test_cleanup()
     pipe: DiffusionPipeline | None = None
     try:
         pipe = DiffusionPipeline.from_pretrained(
@@ -126,7 +126,7 @@ def _run_diffusers_qwen_image_layered(*, model: str, input_image: Image.Image, o
         gc.collect()
         if torch.cuda.is_available():
             torch.accelerator.empty_cache()
-        run_post_test_cleanup(enable_force=True)
+        run_post_test_cleanup()
 
 
 @pytest.mark.benchmark

tests/e2e/offline_inference/test_dynin_omni_expansion.py

@@ -22,7 +22,6 @@ from tests.helpers.mark import hardware_test
 from tests.helpers.stage_config import get_deploy_config_path
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 _REPO_ROOT = Path(__file__).resolve().parents[3]
 _DEFAULT_DYNIN_CONFIG_PATH: Path | None = None

tests/e2e/offline_inference/test_ming_flash_omni_expansion.py

@@ -4,7 +4,6 @@
 import os
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 import pytest
 

tests/e2e/offline_inference/test_omnivoice_expansion.py

@@ -9,7 +9,6 @@ Uses GPUGenerationWorker for both stages (iterative unmasking + DAC decoder).
 import os
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 import numpy as np
 import pytest

tests/e2e/offline_inference/test_qwen3_omni.py

@@ -5,7 +5,6 @@ E2E offline tests for Omni model with video input and audio output.
 import os
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 import pytest
 

tests/e2e/offline_inference/test_qwen3_omni_autoround_w4a16_expansion.py

@@ -50,7 +50,6 @@ def _qwen3_omni_env():
     """
     with pytest.MonkeyPatch.context() as mp:
         mp.setenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
-        mp.setenv("VLLM_TEST_CLEAN_GPU_MEMORY", "0")
         yield
 
 

tests/e2e/offline_inference/test_qwen3_tts_base.py

@@ -11,7 +11,6 @@ Same structure as test_qwen3_omni (models, stage_configs, test_params, parametri
 import os
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 import pytest
 

tests/e2e/offline_inference/test_qwen3_tts_customvoice.py

@@ -11,7 +11,6 @@ Same structure as test_qwen3_omni (models, stage_configs, test_params, parametri
 import os
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 import pytest
 

tests/e2e/online_serving/test_bagel.py

@@ -32,7 +32,6 @@ from tests.helpers.runtime import OmniServerParams
 from tests.helpers.stage_config import get_deploy_config_path
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 MODEL = "ByteDance-Seed/BAGEL-7B-MoT"
 STAGE_CONFIGS_PATH = get_deploy_config_path("ci/bagel.yaml")

tests/e2e/online_serving/test_bagel_multi_replicas.py

@@ -12,7 +12,6 @@ from tests.helpers.runtime import OmniServerParams, dummy_messages_from_mix_data
 from tests.helpers.stage_config import get_deploy_config_path
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 MODEL = "ByteDance-Seed/BAGEL-7B-MoT"
 MULTI_REPLICA_DEPLOY = get_deploy_config_path("ci/bagel_multi_replicas_4gpu.yaml")

tests/e2e/online_serving/test_cosyvoice3_tts_expansion.py

@@ -10,7 +10,6 @@ the CosyVoice3 model, which requires reference audio for voice cloning.
 import os
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 import pytest
 

tests/e2e/online_serving/test_dynin_omni_expansion.py

@@ -18,7 +18,6 @@ from tests.helpers.stage_config import get_deploy_config_path
 pytestmark = [pytest.mark.full_model, pytest.mark.omni]
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 MODEL = "snu-aidas/Dynin-Omni"
 STAGE_CONFIG = get_deploy_config_path("dynin_omni_ci.yaml")

tests/e2e/online_serving/test_ming_flash_omni_expansion.py

@@ -19,7 +19,6 @@ from tests.helpers.runtime import OmniServerParams, dummy_messages_from_mix_data
 from tests.helpers.stage_config import get_deploy_config_path
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 pytestmark = [pytest.mark.omni, pytest.mark.full_model]
 

tests/e2e/online_serving/test_moss_tts_nano_expansion.py

@@ -17,7 +17,6 @@ import os
 import urllib.request
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 import pytest
 

tests/e2e/online_serving/test_omnivoice_expansion.py

@@ -10,7 +10,6 @@ accessed through the standard OpenAI-compatible speech API.
 import os
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 import pytest
 

tests/e2e/online_serving/test_qwen2_5_omni_expansion.py

@@ -12,7 +12,6 @@ from tests.helpers.runtime import OmniServerParams, dummy_messages_from_mix_data
 from tests.helpers.stage_config import get_deploy_config_path, modify_stage_config
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 
 models = ["Qwen/Qwen2.5-Omni-7B"]

tests/e2e/online_serving/test_qwen3_omni.py

@@ -13,7 +13,6 @@ from tests.helpers.stage_config import get_deploy_config_path, modify_stage_conf
 from vllm_omni.platforms import current_omni_platform
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 
 models = ["Qwen/Qwen3-Omni-30B-A3B-Instruct"]

tests/e2e/online_serving/test_qwen3_omni_multi_replicas.py

@@ -12,7 +12,6 @@ from tests.helpers.runtime import OmniResponse, OmniServerParams, dummy_messages
 from tests.helpers.stage_config import get_deploy_config_path
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 MODEL = "Qwen/Qwen3-Omni-30B-A3B-Instruct"
 MULTI_REPLICA_DEPLOY = get_deploy_config_path("ci/qwen3_omni_moe_multi_replicas_4gpu.yaml")

tests/e2e/online_serving/test_qwen3_tts_base.py

@@ -10,7 +10,6 @@ actual model inference, not mocks.
 import os
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 import pytest
 

tests/e2e/online_serving/test_qwen3_tts_base_expansion.py

@@ -19,7 +19,6 @@ from tests.helpers.stage_config import get_deploy_config_path
 pytestmark = [pytest.mark.full_model, pytest.mark.tts]
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 MODEL = "Qwen/Qwen3-TTS-12Hz-0.6B-Base"
 

tests/e2e/online_serving/test_qwen3_tts_customvoice.py

@@ -10,7 +10,6 @@ actual model inference, not mocks.
 import os
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 import pytest
 

tests/e2e/online_serving/test_qwen3_tts_customvoice_expansion.py

@@ -18,7 +18,6 @@ from tests.helpers.stage_config import get_deploy_config_path
 pytestmark = [pytest.mark.full_model, pytest.mark.tts]
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 MODEL = "Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice"
 

tests/e2e/online_serving/test_qwen_image.py

@@ -27,7 +27,6 @@ from tests.helpers.mark import hardware_marks
 from tests.helpers.runtime import OmniServer, OmniServerParams, OpenAIClientHandler, dummy_messages_from_mix_data
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 MODEL = "Qwen/Qwen-Image"
 T2I_PROMPT = "A photo of a cat sitting on a laptop keyboard, digital art style."

tests/e2e/online_serving/test_voxtral_tts.py

@@ -10,7 +10,6 @@ actual model inference, not mocks.
 import os
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 import pytest
 

tests/e2e/online_serving/test_wan22_t2v.py

@@ -21,7 +21,6 @@ from tests.helpers.mark import hardware_marks
 from tests.helpers.runtime import OmniServer, OmniServerParams, OpenAIClientHandler
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 MODEL = "Wan-AI/Wan2.2-T2V-A14B-Diffusers"
 PROMPT = "Two anthropomorphic cats in boxing gear on a spotlighted stage."

tests/entrypoints/openai_api/test_openai_audio_list_endpoints.py

@@ -9,7 +9,6 @@ Migrated from ``tests/e2e/online_serving/test_voxtral_tts.py`` to colocate with
 import os
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 import httpx
 import pytest

tests/entrypoints/openai_api/test_qwen3_tts_speaker_embedding.py

@@ -10,7 +10,6 @@ via pre-computed ECAPA-TDNN embeddings, bypassing ref_audio extraction.
 import os
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 import struct
 

tests/entrypoints/openai_api/test_qwen3_tts_websocket.py

@@ -16,7 +16,6 @@ from tests.helpers.runtime import OmniServerParams
 from tests.helpers.stage_config import get_deploy_config_path
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 MODEL = "Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice"
 

tests/entrypoints/openai_api/test_serving_speech_batch.py

@@ -9,7 +9,6 @@ Validates bulk synthesis via the batch API with actual model inference.
 import os
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 import base64
 import struct

tests/entrypoints/test_omni_sleep_mode.py

@@ -18,7 +18,10 @@ from vllm_omni.platforms import current_omni_platform
 
 logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger("OmniTest")
-pytestmark = [pytest.mark.advanced_model]
+pytestmark = [
+    pytest.mark.advanced_model,
+    pytest.mark.usefixtures("clean_gpu_memory_between_tests"),
+]
 
 
 def clean_gpu_envs():

tests/entrypoints/test_stage_utils.py

@@ -55,7 +55,6 @@ def _make_mock_platform(mocker, device_type: str = "cuda", env_var: str = "CUDA_
 
 @pytest.mark.core_model
 @pytest.mark.cpu
-@pytest.mark.usefixtures("clean_gpu_memory_between_tests")
 def test_set_stage_devices_respects_logical_ids(mocker: MockerFixture, monkeypatch: pytest.MonkeyPatch):
     # Preserve an existing logical mapping and ensure devices "0,1" map through it.
     monkeypatch.setenv("CUDA_VISIBLE_DEVICES", "6,7")
@@ -77,7 +76,6 @@ def test_set_stage_devices_respects_logical_ids(mocker: MockerFixture, monkeypat
 
 @pytest.mark.core_model
 @pytest.mark.cpu
-@pytest.mark.usefixtures("clean_gpu_memory_between_tests")
 def test_set_stage_devices_handles_not_enough_devices(mocker: MockerFixture, monkeypatch: pytest.MonkeyPatch):
     # Preserve an existing logical mapping and ensure devices "0,1" map through it.
     monkeypatch.setenv("CUDA_VISIBLE_DEVICES", "6,7")
@@ -98,7 +96,6 @@ def test_set_stage_devices_handles_not_enough_devices(mocker: MockerFixture, mon
     assert os.environ["CUDA_VISIBLE_DEVICES"] == "6,7"
 
 
-@pytest.mark.usefixtures("clean_gpu_memory_between_tests")
 def test_set_stage_devices_npu_platform(mocker: MockerFixture, monkeypatch: pytest.MonkeyPatch):
     """Test that set_stage_devices works correctly for NPU platform."""
     monkeypatch.setenv("ASCEND_RT_VISIBLE_DEVICES", "4,5")

tests/examples/offline_inference/test_internvla_a1.py

@@ -16,9 +16,9 @@ import pytest
 
 from tests.helpers.mark import hardware_test
 
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "1"
-
 REPO_ROOT = Path(__file__).resolve().parents[3]
+
+
 EXAMPLE_SCRIPT = REPO_ROOT / "examples" / "offline_inference" / "internvla_a1" / "end2end.py"
 
 
@@ -33,6 +33,7 @@ def _required_env(name: str) -> str:
 @pytest.mark.advanced_model
 @pytest.mark.diffusion
 @hardware_test(res={"cuda": "H100", "rocm": "MI325"})
+@pytest.mark.usefixtures("clean_gpu_memory_between_tests")
 def test_internvla_a1_offline_open_loop(run_level: str) -> None:
     if run_level != "advanced_model":
         pytest.skip("InternVLA-A1 offline evaluation requires real local checkpoints and dataset.")

tests/examples/offline_inference/test_text_to_image.py

@@ -11,7 +11,12 @@ from tests.examples.helpers import EXAMPLES, ExampleRunner, ReadmeSnippet
 from tests.helpers.assertions import assert_image_valid
 from tests.helpers.mark import hardware_marks
 
-pytestmark = [pytest.mark.full_model, pytest.mark.example, *hardware_marks(res={"cuda": "H100"})]
+pytestmark = [
+    pytest.mark.usefixtures("clean_gpu_memory_between_tests"),
+    pytest.mark.full_model,
+    pytest.mark.example,
+    *hardware_marks(res={"cuda": "H100"}),
+]
 
 T2I_SCRIPT = EXAMPLES / "offline_inference" / "text_to_image" / "text_to_image.py"
 README_PATH = T2I_SCRIPT.with_name("README.md")

tests/helpers/env.py

@@ -16,11 +16,6 @@ from contextlib import contextmanager
 import torch
 
 
-def run_forced_gpu_cleanup_round() -> None:
-    run_pre_test_cleanup(enable_force=True)
-    run_post_test_cleanup(enable_force=True)
-
-
 def get_physical_device_indices(devices):
     visible_devices = os.environ.get("CUDA_VISIBLE_DEVICES")
     if visible_devices is None:
@@ -232,16 +227,7 @@ def _print_gpu_processes() -> None:
     print("=" * 80)
 
 
-_SKIPPED_GPU_CLEANUP_MSG = (
-    "\nSkipping GPU memory cleanup check (typically: instance already up; no check needed between tests)\n"
-)
-
-
-def run_pre_test_cleanup(enable_force: bool = False) -> None:
-    if os.getenv("VLLM_TEST_CLEAN_GPU_MEMORY", "0") != "1" and not enable_force:
-        print(_SKIPPED_GPU_CLEANUP_MSG)
-        return
-
+def run_pre_test_cleanup() -> None:
     print("Pre-test GPU status:")
 
     num_gpus = torch.accelerator.device_count()
@@ -255,11 +241,7 @@ def run_pre_test_cleanup(enable_force: bool = False) -> None:
             print(f"Pre-test cleanup note: {e}")
 
 
-def run_post_test_cleanup(enable_force: bool = False) -> None:
-    if os.getenv("VLLM_TEST_CLEAN_GPU_MEMORY", "0") != "1" and not enable_force:
-        print(_SKIPPED_GPU_CLEANUP_MSG)
-        return
-
+def run_post_test_cleanup() -> None:
     if torch.cuda.is_available():
         gc.collect()
         torch.accelerator.empty_cache()
@@ -318,6 +300,5 @@ __all__ = [
     "get_physical_device_indices",
     "run_post_test_cleanup",
     "run_pre_test_cleanup",
-    "run_forced_gpu_cleanup_round",
     "wait_for_gpu_memory_to_clear",
 ]

tests/helpers/fixtures/env.py

@@ -27,10 +27,12 @@ def model_prefix() -> str:
     return f"{prefix.rstrip('/')}/" if prefix else ""
 
 
-@pytest.fixture(autouse=True)
+@pytest.fixture
 def clean_gpu_memory_between_tests():
-    # Import here so ``tests.helpers.env`` (and vLLM platform modules) load only
-    # after session autouse fixtures like ``default_env`` have run (RFC #2299).
+    """Opt-in GPU pre/post hooks for a test (no environment-variable gate).
+
+    Use as a test parameter or ``@pytest.mark.usefixtures("clean_gpu_memory_between_tests")``.
+    """
     from tests.helpers.env import run_post_test_cleanup, run_pre_test_cleanup
 
     print("\n=== PRE-TEST GPU CLEANUP ===")

tests/helpers/runtime.py

@@ -35,7 +35,7 @@ from tests.helpers.assertions import (
     assert_diffusion_response,
     assert_omni_response,
 )
-from tests.helpers.env import run_forced_gpu_cleanup_round
+from tests.helpers.env import run_post_test_cleanup, run_pre_test_cleanup
 from tests.helpers.media import (
     _merge_base64_audio_to_segment,
     convert_audio_bytes_to_text,
@@ -201,7 +201,8 @@ class OmniServer:
         env_dict: dict[str, str] | None = None,
         use_omni: bool = True,
     ) -> None:
-        run_forced_gpu_cleanup_round()
+        run_pre_test_cleanup()
+        run_post_test_cleanup()
         cleanup_dist_env_and_memory()
         self.model = model
         self.serve_args = serve_args
@@ -306,7 +307,8 @@ class OmniServer:
     def __exit__(self, exc_type, exc_val, exc_tb):
         if self.proc:
             self._kill_process_tree(self.proc.pid)
-        run_forced_gpu_cleanup_round()
+        run_pre_test_cleanup()
+        run_post_test_cleanup()
         cleanup_dist_env_and_memory()
 
 
@@ -593,7 +595,8 @@ class OmniServerStageCli(OmniServer):
             proc = self.stage_procs[stage_key]
             if proc.poll() is None:
                 self._kill_process_tree(proc.pid)
-        run_forced_gpu_cleanup_round()
+        run_pre_test_cleanup()
+        run_post_test_cleanup()
         cleanup_dist_env_and_memory()
 
 
@@ -1272,7 +1275,8 @@ class OmniRunner:
         **kwargs,
     ) -> None:
         cleanup_dist_env_and_memory()
-        run_forced_gpu_cleanup_round()
+        run_pre_test_cleanup()
+        run_post_test_cleanup()
         self.model_name = model_name
         self.seed = seed
         self._prompt_len_estimate_cache: dict[str, Any] = {}
@@ -1538,7 +1542,8 @@ class OmniRunner:
         if hasattr(self.omni, "close"):
             self.omni.close()
         self._cleanup_process()
-        run_forced_gpu_cleanup_round()
+        run_pre_test_cleanup()
+        run_post_test_cleanup()
         cleanup_dist_env_and_memory()
 
 
@@ -1876,6 +1881,5 @@ __all__ = [
     "OmniServerStageCli",
     "OpenAIClientHandler",
     "get_open_port",
-    "run_forced_gpu_cleanup_round",
     "dummy_messages_from_mix_data",
 ]