[ROCm] Enable pinned memory on supported WSL2 kernels (#50126)
Signed-off-by: Flora Cui <flora.cui@amd.com>
This commit is contained in:
@@ -8,6 +8,6 @@ ARCH_TAG="${BUILDKITE_COMMIT}-$(uname -m)-xpu"
|
|||||||
PLATFORM_TAG="${BUILDKITE_COMMIT}-xpu"
|
PLATFORM_TAG="${BUILDKITE_COMMIT}-xpu"
|
||||||
|
|
||||||
aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin ${REGISTRY}
|
aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin ${REGISTRY}
|
||||||
docker manifest rm ${REGISTRY}/${REPO}:${PLATFORM_TAG} || true
|
docker manifest rm "${REGISTRY}/${REPO}:${PLATFORM_TAG}" || true
|
||||||
docker manifest create ${REGISTRY}/${REPO}:${PLATFORM_TAG} ${REGISTRY}/${REPO}:${ARCH_TAG} --amend
|
docker manifest create "${REGISTRY}/${REPO}:${PLATFORM_TAG}" "${REGISTRY}/${REPO}:${ARCH_TAG}" --amend
|
||||||
docker manifest push ${REGISTRY}/${REPO}:${PLATFORM_TAG}
|
docker manifest push ${REGISTRY}/${REPO}:${PLATFORM_TAG}
|
||||||
+27
-1
@@ -2,6 +2,7 @@
|
|||||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||||
|
|
||||||
import os
|
import os
|
||||||
|
import platform
|
||||||
from datetime import timedelta
|
from datetime import timedelta
|
||||||
from functools import cache, lru_cache, wraps
|
from functools import cache, lru_cache, wraps
|
||||||
from typing import TYPE_CHECKING
|
from typing import TYPE_CHECKING
|
||||||
@@ -15,7 +16,7 @@ import vllm.envs as envs
|
|||||||
from vllm.logger import init_logger
|
from vllm.logger import init_logger
|
||||||
from vllm.v1.attention.backends.registry import AttentionBackendEnum
|
from vllm.v1.attention.backends.registry import AttentionBackendEnum
|
||||||
|
|
||||||
from .interface import DeviceCapability, Platform, PlatformEnum
|
from .interface import DeviceCapability, Platform, PlatformEnum, in_wsl
|
||||||
|
|
||||||
if TYPE_CHECKING:
|
if TYPE_CHECKING:
|
||||||
from vllm.config import VllmConfig
|
from vllm.config import VllmConfig
|
||||||
@@ -113,6 +114,16 @@ def _rocm_device_count_stateless(cuda_visible_devices: str | None = None) -> int
|
|||||||
return r
|
return r
|
||||||
|
|
||||||
|
|
||||||
|
@cache
|
||||||
|
def _get_wsl_kernel_version() -> tuple[int, ...] | None:
|
||||||
|
try:
|
||||||
|
release = platform.uname().release
|
||||||
|
parts = release.split("-")[0].split(".")
|
||||||
|
return tuple(int(part) for part in parts[:3])
|
||||||
|
except (TypeError, ValueError):
|
||||||
|
return None
|
||||||
|
|
||||||
|
|
||||||
def _sync_hip_cuda_env_vars():
|
def _sync_hip_cuda_env_vars():
|
||||||
"""Ensure HIP_VISIBLE_DEVICES and CUDA_VISIBLE_DEVICES are consistent.
|
"""Ensure HIP_VISIBLE_DEVICES and CUDA_VISIBLE_DEVICES are consistent.
|
||||||
Treats empty string as unset. Raises on genuine conflicts."""
|
Treats empty string as unset. Raises on genuine conflicts."""
|
||||||
@@ -536,6 +547,21 @@ class RocmPlatform(Platform):
|
|||||||
with contextlib.suppress(ImportError):
|
with contextlib.suppress(ImportError):
|
||||||
import vllm._rocm_C # noqa: F401
|
import vllm._rocm_C # noqa: F401
|
||||||
|
|
||||||
|
@classmethod
|
||||||
|
def is_pin_memory_available(cls) -> bool:
|
||||||
|
if in_wsl():
|
||||||
|
version = _get_wsl_kernel_version()
|
||||||
|
if version is None or version < (4, 19, 121):
|
||||||
|
# warning_once() causes a circular import on WSL, see #48397.
|
||||||
|
logger.warning(
|
||||||
|
"Using 'pin_memory=False' as WSL is detected and the "
|
||||||
|
"WSL2 kernel version is below 4.19.121. This may slow "
|
||||||
|
"down performance. Please run `wsl --update`."
|
||||||
|
)
|
||||||
|
return False
|
||||||
|
|
||||||
|
return True
|
||||||
|
|
||||||
@classmethod
|
@classmethod
|
||||||
def get_valid_backends(
|
def get_valid_backends(
|
||||||
cls,
|
cls,
|
||||||
|
|||||||
Reference in New Issue
Block a user