[MM Encoder]: Make MMEncoderAttention's scale takes effect properly (#31950)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
2026-01-08 18:33:48 +08:00
parent 5576227bc1
commit 2972a05473
11 changed files with 32 additions and 8 deletions
--- a/vllm/model_executor/models/paddleocr_vl.py
+++ b/vllm/model_executor/models/paddleocr_vl.py
@@ -564,6 +564,7 @@ class SiglipAttention(nn.Module):
        self.attn = MMEncoderAttention(
            num_heads=self.num_attention_heads_per_partition,
            head_size=self.hidden_size_per_attention_head,
+            scale=self.hidden_size_per_attention_head**-0.5,
            multimodal_config=multimodal_config,
            prefix=f"{prefix}.attn",
        )