From c3bb9f23315f8cdc5007349717fd58d16a833ff6 Mon Sep 17 00:00:00 2001 From: Lukas Geiger Date: Fri, 30 May 2025 10:12:59 +0100 Subject: [PATCH] [Model] Use in-place adds in SigLIP (#18922) Signed-off-by: Lukas Geiger --- vllm/model_executor/models/siglip.py | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/vllm/model_executor/models/siglip.py b/vllm/model_executor/models/siglip.py index 3b5334afa7af8..4803da2956ef1 100644 --- a/vllm/model_executor/models/siglip.py +++ b/vllm/model_executor/models/siglip.py @@ -130,11 +130,10 @@ class SiglipVisionEmbeddings(nn.Module): embeddings = patch_embeds.flatten(2).transpose(1, 2) if interpolate_pos_encoding: - embeddings = embeddings + self.interpolate_pos_encoding( + embeddings += self.interpolate_pos_encoding( embeddings, height, width) else: - embeddings = embeddings + self.position_embedding( - self.position_ids) + embeddings += self.position_embedding(self.position_ids) return embeddings @@ -271,12 +270,12 @@ class SiglipEncoderLayer(nn.Module): hidden_states = self.layer_norm1(hidden_states) hidden_states, _ = self.self_attn(hidden_states=hidden_states) - hidden_states = residual + hidden_states + hidden_states += residual residual = hidden_states hidden_states = self.layer_norm2(hidden_states) hidden_states = self.mlp(hidden_states) - hidden_states = residual + hidden_states + hidden_states += residual return hidden_states, None @@ -354,7 +353,8 @@ class SiglipMultiheadAttentionPoolingHead(nn.Module): residual = hidden_state hidden_state = self.layernorm(hidden_state) - hidden_state = residual + self.mlp(hidden_state) + hidden_state = self.mlp(hidden_state) + hidden_state += residual return hidden_state[:, 0]