[Refactor]A simple device-related refactor (#11163)

Signed-off-by: noemotiovon <noemotiovon@gmail.com>
Co-authored-by: noemotiovon <noemotiovon@gmail.com>
This commit is contained in:
Chenguang Li
2024-12-13 13:39:00 +00:00
committed by GitHub
co-authored by noemotiovon
parent 969da7d70b
commit d1fa714cb1
7 changed files with 51 additions and 31 deletions
+17
View File
@@ -1,6 +1,7 @@
import enum
import platform
import random
from platform import uname
from typing import TYPE_CHECKING, NamedTuple, Optional, Tuple, Union
import numpy as np
@@ -16,6 +17,11 @@ else:
logger = init_logger(__name__)
def in_wsl() -> bool:
# Reference: https://github.com/microsoft/WSL/issues/4071
return "microsoft" in " ".join(uname()).lower()
class _Backend(enum.Enum):
FLASH_ATTN = enum.auto()
FLASH_ATTN_VLLM_V1 = enum.auto()
@@ -221,6 +227,17 @@ class Platform:
return CpuArchEnum.OTHER if machine else CpuArchEnum.UNKNOWN
@classmethod
def is_pin_memory_available(cls) -> bool:
"""Checks whether pin memory is available on the current platform."""
if in_wsl():
# Pinning memory in WSL is not supported.
# https://docs.nvidia.com/cuda/wsl-user-guide/index.html#known-limitations-for-linux-cuda-applications
logger.warning("Using 'pin_memory=False' as WSL is detected. "
"This may slow down the performance.")
return False
return True
class UnspecifiedPlatform(Platform):
_enum = PlatformEnum.UNSPECIFIED