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 10, 04:00 UTC
[Model] Avoid token selection in SigLIP pooling head (#32389)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Jan 15, 2026 at 17:01 UTC
cdba4c74b30824b22d9d7cd22b187d5e4f690404
1 file changed
+3
-3
vllm/model_executor/models/siglip.py
+3
-3
@@ -690,9 +690,9 @@ class SiglipMultiheadAttentionPoolingHead(nn.Module):
690
hidden_state = self.mlp(hidden_state)
691
hidden_state += residual
692
693
- pooled = hidden_state[:, 0]
694
-
695
- return pooled.unsqueeze(1)
693
+ # Handled by resolve_visual_encoder_outputs
694
+ # return hidden_state[:, 0]
695
+ return hidden_state
696
697
698
class SiglipVisionTransformer(nn.Module):