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
[Models] Use in-place adds in Idefics2Vision (#23932)
Signed-off-by: Lukas Geiger <lukas.geiger94@gmail.com>
Lukas Geiger committed
Aug 29, 2025 at 15:42 UTC
0a2f4c0793988d3cf0d47b5f771fb38231db4b2b
1 file changed
+3
-3
vllm/model_executor/models/idefics2_vision_model.py
+3
-3
@@ -108,7 +108,7 @@ class Idefics2VisionEmbeddings(nn.Module):
108
bucket_coords_w).flatten()
109
position_ids[batch_idx][p_attn_mask.view(-1).cpu()] = pos_ids
110
position_ids = position_ids.to(self.position_embedding.weight.device)
111
- embeddings = embeddings + self.position_embedding(position_ids)
111
+ embeddings += self.position_embedding(position_ids)
112
return embeddings
113
114
@@ -262,11 +262,11 @@ class Idefics2EncoderLayer(nn.Module):
262
residual = hidden_states
263
hidden_states = self.layer_norm1(hidden_states)
264
hidden_states = self.self_attn(hidden_states)
265
- hidden_states = residual + hidden_states
265
+ hidden_states += residual
266
residual = hidden_states
267
hidden_states = self.layer_norm2(hidden_states)
268
hidden_states = self.mlp(hidden_states)
269
- hidden_states = residual + hidden_states
269
+ hidden_states += residual
270
return hidden_states
271
272