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

[Models][Qwen3 ViT] Keep max_seqlen on CPU to prevent D2H sync (#37139)

Browse Source 
Signed-off-by: Lukas Geiger <lukas.geiger94@gmail.com>
Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Lukas Geiger
2026-03-16 12:11:59 +00:00
committed by GitHub
parent d61d2b08e9
commit f9e6db3034
1 changed files with 0 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
vllm/model_executor/models/qwen3_vl.py
View File

@@ -557,7 +557,6 @@ class Qwen3_VisionTransformer(nn.Module):
max_seqlen = torch.tensor( max_seqlen = torch.tensor(
MMEncoderAttention.compute_max_seqlen(self.attn_backend, cu_seqlens), MMEncoderAttention.compute_max_seqlen(self.attn_backend, cu_seqlens),
dtype=torch.int32, dtype=torch.int32,
device=self.device,
) )
cu_seqlens = MMEncoderAttention.maybe_recompute_cu_seqlens( cu_seqlens = MMEncoderAttention.maybe_recompute_cu_seqlens(
self.attn_backend, self.attn_backend,