[ROCm][Bugfix] Fix use_v2_model_runner inside Ray driver thread (#45998)

Signed-off-by: Micah Williamson <micah.williamson@amd.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Micah Williamson
2026-06-24 08:41:56 +08:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent bcbeaac786
commit e2bdc24612
2 changed files with 23 additions and 4 deletions
+10 -3
View File
@@ -7,6 +7,7 @@ from importlib.metadata import version
from importlib.util import find_spec
from vllm.logger import init_logger
from vllm.platforms import current_platform
from vllm.utils.math_utils import cdiv
logger = init_logger(__name__)
@@ -27,10 +28,16 @@ if HAS_TRITON:
]
# Check if we're in a distributed environment where CUDA_VISIBLE_DEVICES
# might be temporarily empty (e.g., Ray sets it to "" during actor init)
cuda_visible_devices = os.environ.get("CUDA_VISIBLE_DEVICES")
# or HIP_VISIBLE_DEVICES might be temporarily empty (e.g., Ray sets it to ""
# during actor init)
visible_devices_env = (
"HIP_VISIBLE_DEVICES"
if current_platform.is_rocm()
else "CUDA_VISIBLE_DEVICES"
)
visible_devices = os.environ.get(visible_devices_env)
is_distributed_env = (
cuda_visible_devices is not None and len(cuda_visible_devices.strip()) == 0
visible_devices is not None and len(visible_devices.strip()) == 0
)
# Apply lenient driver check for distributed environments