A high-throughput and memory-efficient inference and serving engine for LLMs
[BUGFIX][XPU] fix xpu `flash_attn_varlen_func` interface (#45110)
Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
Kunshang Ji committed
Jun 10, 2026 at 17:07 UTC
fe1d923afccb2eba49e7525389ddf1e81ef6b7dd
1 file changed
+3
vllm/_xpu_ops.py
+3
@@ -1,6 +1,7 @@
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
+from collections.abc import Callable
5
from typing import TYPE_CHECKING
6
7
import torch
@@ -783,6 +784,8 @@ class xpu_ops:
784
return_softmax_lse: bool | None = False,
785
s_aux: torch.Tensor | None = None,
786
return_attn_probs: bool | None = False,
787
+ mask_mod: Callable | None = None,
788
+ aux_tensors: list | None = None,
789
):
790
assert cu_seqlens_k is not None or seqused_k is not None, (
791
"cu_seqlens_k or seqused_k must be provided"