[Chore] Separate out vllm.utils.platform_utils.py (#27374)

Signed-off-by: Jonathan <chenleejonathan@gmail.com>
This commit is contained in:
Jonathan Chen
2025-10-23 19:08:06 +00:00
committed by GitHub
parent a9f55dc588
commit ca76486a16
21 changed files with 79 additions and 70 deletions
+3 -50
View File
@@ -27,12 +27,8 @@ from argparse import (
_ArgumentGroup,
)
from collections import defaultdict
from collections.abc import (
Callable,
Sequence,
)
from concurrent.futures.process import ProcessPoolExecutor
from functools import cache, partial, wraps
from collections.abc import Callable
from functools import partial, wraps
from typing import TYPE_CHECKING, Any, TypeVar
import cloudpickle
@@ -44,6 +40,7 @@ import yaml
import vllm.envs as envs
from vllm.logger import enable_trace_function_call, init_logger
from vllm.ray.lazy_utils import is_in_ray_actor
from vllm.utils.platform_utils import cuda_is_initialized, xpu_is_initialized
_DEPRECATED_MAPPINGS = {
"cprofile": "profiling",
@@ -169,21 +166,6 @@ def round_down(x: int, y: int) -> int:
return (x // y) * y
@cache
def is_pin_memory_available() -> bool:
from vllm.platforms import current_platform
return current_platform.is_pin_memory_available()
@cache
def is_uva_available() -> bool:
"""Check if Unified Virtual Addressing (UVA) is available."""
# UVA requires pinned memory.
# TODO: Add more requirements for UVA if needed.
return is_pin_memory_available()
# TODO: This function can be removed if transformer_modules classes are
# serialized by value when communicating between processes
def init_cached_hf_modules() -> None:
@@ -216,35 +198,6 @@ def enable_trace_function_call_for_thread(vllm_config: VllmConfig) -> None:
enable_trace_function_call(log_path)
def cuda_is_initialized() -> bool:
"""Check if CUDA is initialized."""
if not torch.cuda._is_compiled():
return False
return torch.cuda.is_initialized()
def xpu_is_initialized() -> bool:
"""Check if XPU is initialized."""
if not torch.xpu._is_compiled():
return False
return torch.xpu.is_initialized()
def cuda_get_device_properties(
device, names: Sequence[str], init_cuda=False
) -> tuple[Any, ...]:
"""Get specified CUDA device property values without initializing CUDA in
the current process."""
if init_cuda or cuda_is_initialized():
props = torch.cuda.get_device_properties(device)
return tuple(getattr(props, name) for name in names)
# Run in subprocess to avoid initializing CUDA as a side effect.
mp_ctx = multiprocessing.get_context("fork")
with ProcessPoolExecutor(max_workers=1, mp_context=mp_ctx) as executor:
return executor.submit(cuda_get_device_properties, device, names, True).result()
def weak_bind(
bound_method: Callable[..., Any],
) -> Callable[..., None]:
+54
View File
@@ -0,0 +1,54 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import multiprocessing
from collections.abc import Sequence
from concurrent.futures.process import ProcessPoolExecutor
from functools import cache
from typing import Any
import torch
def cuda_is_initialized() -> bool:
"""Check if CUDA is initialized."""
if not torch.cuda._is_compiled():
return False
return torch.cuda.is_initialized()
def xpu_is_initialized() -> bool:
"""Check if XPU is initialized."""
if not torch.xpu._is_compiled():
return False
return torch.xpu.is_initialized()
def cuda_get_device_properties(
device, names: Sequence[str], init_cuda=False
) -> tuple[Any, ...]:
"""Get specified CUDA device property values without initializing CUDA in
the current process."""
if init_cuda or cuda_is_initialized():
props = torch.cuda.get_device_properties(device)
return tuple(getattr(props, name) for name in names)
# Run in subprocess to avoid initializing CUDA as a side effect.
mp_ctx = multiprocessing.get_context("fork")
with ProcessPoolExecutor(max_workers=1, mp_context=mp_ctx) as executor:
return executor.submit(cuda_get_device_properties, device, names, True).result()
@cache
def is_pin_memory_available() -> bool:
from vllm.platforms import current_platform
return current_platform.is_pin_memory_available()
@cache
def is_uva_available() -> bool:
"""Check if Unified Virtual Addressing (UVA) is available."""
# UVA requires pinned memory.
# TODO: Add more requirements for UVA if needed.
return is_pin_memory_available()