biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Multimodal][XPU]Enable vision attn backend for xpu platform (#27525)

Browse Source 
Signed-off-by: Yan Ma <yan.ma@intel.com>
Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
Co-authored-by: Yejing Lai <yejing.lai@intel.com>
Co-authored-by: Guancheng Fu <110874468+gc-fu@users.noreply.github.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
This commit is contained in:
Yan Ma
2025-11-01 12:45:02 +08:00
committed by GitHub
parent 3a5de7d2d6
commit 7e2729b57e
6 changed files with 88 additions and 51 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
vllm/model_executor/models/qwen2_vl.py
View File

@@ -789,10 +789,7 @@ class Qwen2VisionTransformer(nn.Module):
self, cu_seqlens: torch.Tensor
) -> tuple[int | None, list[int] | None]:
max_seqlen, seqlens = None, None
if (
self.attn_backend == _Backend.FLASH_ATTN
or self.attn_backend == _Backend.ROCM_AITER_FA
):
if self.attn_backend in {_Backend.FLASH_ATTN, _Backend.ROCM_AITER_FA}:
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
elif self.attn_backend == _Backend.XFORMERS:
seqlens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist()