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 14, 05:30 UTC
[Bugfix][Qwen3VL] fix deepstack in qwen3vl (#26626)
Signed-off-by: liuye.hj <liuye.hj@alibaba-inc.com> Signed-off-by: JJJYmmm <92386084+JJJYmmm@users.noreply.github.com> Co-authored-by: liuye.hj <liuye.hj@alibaba-inc.com>
JJJYmmm committed
Oct 11, 2025 at 20:58 UTC
9d6cff3edeb2421699671881592fd7558946695e
1 file changed
-6
vllm/model_executor/models/qwen3_vl.py
-6
@@ -1702,12 +1702,6 @@ class Qwen3VLForConditionalGeneration(
1702
)
1703
1704
if deepstack_input_embeds is not None:
1705
- deepstack_input_embeds = (
1706
- torch.zeros_like(inputs_embeds)
1707
- .unsqueeze(0)
1708
- .repeat(self.deepstack_num_level, 1, 1)
1709
- .contiguous()
1710
- )
1705
self._set_deepstack_input_embeds(deepstack_input_embeds)
1706
1707
return inputs_embeds