[Attention] Refactor CUDA attention backend selection logic (#24794)

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Signed-off-by: Matthew Bonanni <mbonanni001@gmail.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
This commit is contained in:
Matthew Bonanni
2025-11-11 07:40:44 -05:00
committed by GitHub
co-authored by Luka Govedič
parent 2e78150d24
commit b30dfa03c5
61 changed files with 1338 additions and 1002 deletions
+4 -4
View File
@@ -10,7 +10,7 @@ from typing import Final, Generic, Literal, Protocol, TypeAlias, TypeVar
import torch
from transformers import PretrainedConfig
from vllm.attention.backends.registry import _Backend
from vllm.attention.backends.registry import AttentionBackendEnum
from vllm.config import VllmConfig
from vllm.distributed import (
get_tensor_model_parallel_rank,
@@ -83,8 +83,8 @@ def get_vit_attn_backend(
head_size: int,
dtype: torch.dtype,
*,
attn_backend_override: _Backend | None = None,
) -> _Backend:
attn_backend_override: AttentionBackendEnum | None = None,
) -> AttentionBackendEnum:
"""
Get the available attention backend for Vision Transformer.
"""
@@ -94,7 +94,7 @@ def get_vit_attn_backend(
# Lazy import to avoid circular dependency
from vllm.attention.selector import get_env_variable_attn_backend
selected_backend: _Backend | None = get_env_variable_attn_backend()
selected_backend: AttentionBackendEnum | None = get_env_variable_attn_backend()
if selected_backend is not None:
return selected_backend