A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 6, 17:00 UTC
[Models][Qwen3 ViT] Keep `max_seqlen` on CPU to prevent D2H sync (#37139)
Signed-off-by: Lukas Geiger <lukas.geiger94@gmail.com> Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Lukas Geiger committed
Mar 16, 2026 at 12:11 UTC
f9e6db30349d7ec70410981b1f634a1e661e61e1
1 file changed
-1
vllm/model_executor/models/qwen3_vl.py
-1
@@ -557,7 +557,6 @@ class Qwen3_VisionTransformer(nn.Module):
557
max_seqlen = torch.tensor(
558
MMEncoderAttention.compute_max_seqlen(self.attn_backend, cu_seqlens),
559
dtype=torch.int32,
560
- device=self.device,
560
)
561
cu_seqlens = MMEncoderAttention.maybe_recompute_cu_seqlens(
562
self.attn_backend,