Adapt the Qwen2Model._update_causal_mask for transformers==4.57.1 (#18774)

2026-02-17 10:20:41 +08:00
parent 50ca24aebb
commit b21390f8f3
1 changed files with 10 additions and 1 deletions
--- a/python/sglang/srt/models/deepseek_ocr.py
+++ b/python/sglang/srt/models/deepseek_ocr.py
@@ -1217,9 +1217,18 @@ class CustomQwen2Decoder(nn.Module):
                cache_position=None,
            ):
                self._current_token_type_ids = token_type_ids
+                causal_mask_mapping = {
+                    "full_attention": self._update_causal_mask(
+                        attention_mask,
+                        inputs_embeds,
+                        cache_position,
+                        past_key_values,
+                        output_attentions,
+                    )
+                }
                return super().forward(
                    input_ids=input_ids,
-                    attention_mask=attention_mask,
+                    attention_mask=causal_mask_mapping,
                    position_ids=position_ids,
                    past_key_values=past_key_values,
                    inputs_embeds=inputs_embeds,