From cdba4c74b30824b22d9d7cd22b187d5e4f690404 Mon Sep 17 00:00:00 2001 From: Cyrus Leung Date: Thu, 15 Jan 2026 17:01:59 +0800 Subject: [PATCH] [Model] Avoid token selection in SigLIP pooling head (#32389) Signed-off-by: DarkLight1337 --- vllm/model_executor/models/siglip.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/vllm/model_executor/models/siglip.py b/vllm/model_executor/models/siglip.py index 37bb568dd..1e06ace63 100644 --- a/vllm/model_executor/models/siglip.py +++ b/vllm/model_executor/models/siglip.py @@ -690,9 +690,9 @@ class SiglipMultiheadAttentionPoolingHead(nn.Module): hidden_state = self.mlp(hidden_state) hidden_state += residual - pooled = hidden_state[:, 0] - - return pooled.unsqueeze(1) + # Handled by resolve_visual_encoder_outputs + # return hidden_state[:, 0] + return hidden_state class SiglipVisionTransformer(nn.Module):