[diffusion] use fa3 by default when device supports it (#783)

Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
This commit is contained in:
Jiangyun Zhu authored and GitHub committed 2026-01-22 15:00:29 +08:00
1 parent f266b04534
commit 0df8e80574
4 files changed
+22 -14

No files matched your search

@@ -2,6 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import torch
from fa3_fwd_interface import flash_attn_func, flash_attn_varlen_func
from vllm.logger import init_logger
from vllm_omni.diffusion.attention.backends.abstract import (
@@ -13,18 +14,6 @@ from vllm_omni.diffusion.attention.backends.utils.fa import _pad_input, _unpad_i
logger = init_logger(__name__)
try:
# only tested with flash_attn v3
# from flash_attn_interface import flash_attn_func as flash_attn_3_func # not available in flash-attn 2.8.1
from flash_attn import flash_attn_func, flash_attn_varlen_func # can be FA2 or FA3
except ImportError:
logger.warning(
"FlashAttentionBackend is not available. You may install flash-attn "
"by running `uv pip install flash-attn==2.8.1 --no-build-isolation`"
" or install pre-built flash-attn from https://github.com/Dao-AILab/flash-attention/releases"
)
raise ImportError
class FlashAttentionBackend(AttentionBackend):
accept_output_buffer: bool = True