[Attention] Update attention imports (#29540)

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
Matthew Bonanni
2025-11-27 11:19:09 -05:00
committed by GitHub
parent cd007a53b4
commit fc1d8be3dc
38 changed files with 63 additions and 126 deletions
+2 -8
View File
@@ -15,6 +15,8 @@ from typing_extensions import ParamSpec
# import custom ops, trigger op registration
import vllm._C # noqa
import vllm.envs as envs
from vllm.attention.backends.abstract import AttentionType
from vllm.attention.backends.registry import AttentionBackendEnum
from vllm.logger import init_logger
from vllm.utils.import_utils import import_pynvml
from vllm.utils.torch_utils import cuda_device_count_stateless
@@ -22,11 +24,9 @@ from vllm.utils.torch_utils import cuda_device_count_stateless
from .interface import DeviceCapability, Platform, PlatformEnum
if TYPE_CHECKING:
from vllm.attention.backends.registry import AttentionBackendEnum
from vllm.config import VllmConfig
from vllm.config.cache import CacheDType
else:
AttentionBackendEnum = None
VllmConfig = None
CacheDType = None
@@ -48,8 +48,6 @@ def _get_backend_priorities(
device_capability: DeviceCapability,
) -> list[AttentionBackendEnum]:
"""Get backend priorities with lazy import to avoid circular dependency."""
from vllm.attention.backends.registry import AttentionBackendEnum
if use_mla:
if device_capability.major == 10:
return [
@@ -265,8 +263,6 @@ class CudaPlatformBase(Platform):
def get_vit_attn_backend(
cls, head_size: int, dtype: torch.dtype
) -> "AttentionBackendEnum":
from vllm.attention.backends.registry import AttentionBackendEnum
# Try FlashAttention first
try:
backend_class = AttentionBackendEnum.FLASH_ATTN.get_class()
@@ -335,8 +331,6 @@ class CudaPlatformBase(Platform):
use_sparse: bool,
attn_type: str | None = None,
) -> str:
from vllm.attention.backends.abstract import AttentionType
if attn_type is None:
attn_type = AttentionType.DECODER