[Core] Ensure memory is pinned prior to async h2d copy (#45424)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
@@ -9,7 +9,6 @@ from typing import TYPE_CHECKING
|
||||
from vllm import envs
|
||||
from vllm.plugins import PLATFORM_PLUGINS_GROUP, load_plugins_by_group
|
||||
from vllm.utils.import_utils import resolve_obj_by_qualname
|
||||
from vllm.utils.torch_utils import supports_xccl
|
||||
|
||||
from .interface import CpuArchEnum, Platform, PlatformEnum
|
||||
|
||||
@@ -135,7 +134,7 @@ def xpu_platform_plugin() -> str | None:
|
||||
try:
|
||||
import torch
|
||||
|
||||
if supports_xccl():
|
||||
if torch.distributed.is_xccl_available():
|
||||
dist_backend = "xccl"
|
||||
from vllm.platforms.xpu import XPUPlatform
|
||||
|
||||
|
||||
@@ -23,7 +23,6 @@ import vllm._C_stable_libtorch # noqa
|
||||
import vllm.envs as envs
|
||||
from vllm.logger import init_logger
|
||||
from vllm.utils.import_utils import import_pynvml
|
||||
from vllm.utils.torch_utils import is_quantized_kv_cache
|
||||
from vllm.v1.attention.backends.registry import AttentionBackendEnum
|
||||
|
||||
from .interface import DeviceCapability, Platform, PlatformEnum, in_wsl
|
||||
@@ -88,6 +87,8 @@ def _get_backend_priorities(
|
||||
kv_cache_dtype: CacheDType | None = None,
|
||||
) -> list[AttentionBackendEnum]:
|
||||
"""Get backend priorities with lazy import to avoid circular dependency."""
|
||||
from vllm.utils.torch_utils import is_quantized_kv_cache
|
||||
|
||||
if use_mla:
|
||||
if device_capability.major == 10:
|
||||
# Sparse MLA backend priorities
|
||||
|
||||
@@ -14,7 +14,6 @@ import vllm_xpu_kernels._xpu_C # noqa
|
||||
|
||||
import vllm.envs as envs
|
||||
from vllm.logger import init_logger
|
||||
from vllm.utils.torch_utils import supports_xpu_graph
|
||||
from vllm.v1.attention.backends.registry import AttentionBackendEnum
|
||||
|
||||
from .interface import DeviceCapability, Platform, PlatformEnum
|
||||
@@ -178,8 +177,6 @@ class XPUPlatform(Platform):
|
||||
|
||||
@classmethod
|
||||
def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
|
||||
parallel_config = vllm_config.parallel_config
|
||||
|
||||
# lazy import to avoid circular import
|
||||
from vllm.config import CUDAGraphMode
|
||||
|
||||
@@ -190,6 +187,10 @@ class XPUPlatform(Platform):
|
||||
attention_config = vllm_config.attention_config
|
||||
if attention_config.backend is None:
|
||||
attention_config.backend = AttentionBackendEnum.FLASH_ATTN
|
||||
|
||||
# lazy import to avoid circular import
|
||||
from vllm.utils.torch_utils import supports_xpu_graph
|
||||
|
||||
if not supports_xpu_graph():
|
||||
compilation_config.cudagraph_mode = CUDAGraphMode.NONE
|
||||
logger.warning(
|
||||
@@ -324,9 +325,8 @@ class XPUPlatform(Platform):
|
||||
|
||||
@classmethod
|
||||
def get_device_communicator_cls(cls) -> str:
|
||||
from vllm.utils.torch_utils import supports_xccl
|
||||
|
||||
if not supports_xccl():
|
||||
if not torch.distributed.is_xccl_available():
|
||||
# Supports xccl with PyTorch versions >= 2.8.0.dev for XPU platform
|
||||
logger.warning(
|
||||
"xccl is not enabled in this torch build, communication"
|
||||
" is not available."
|
||||
|
||||
Reference in New Issue
Block a user