[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
+6 -2
View File
@@ -150,11 +150,15 @@ class EagleProposer:
)
# Determine allowed attention backends once during initialization.
from vllm.attention.backends.registry import AttentionBackendEnum
self.allowed_attn_types: tuple | None = None
if current_platform.is_rocm():
rocm_types = [TritonAttentionMetadata, FlashAttentionMetadata]
# vllm.v1.attention.backends.rocm_aiter_fa is an optional backend
if find_spec("vllm.v1.attention.backends.rocm_aiter_fa"):
# ROCM_AITER_FA is an optional backend
if find_spec(
AttentionBackendEnum.ROCM_AITER_FA.get_path(include_classname=False)
):
from vllm.v1.attention.backends.rocm_aiter_fa import (
AiterFlashAttentionMetadata,
)