From b21390f8f396f734a0d516009fa2c87f2b5e645b Mon Sep 17 00:00:00 2001 From: pansicheng Date: Tue, 17 Feb 2026 10:20:41 +0800 Subject: [PATCH] Adapt the Qwen2Model._update_causal_mask for transformers==4.57.1 (#18774) --- python/sglang/srt/models/deepseek_ocr.py | 11 ++++++++++- 1 file changed, 10 insertions(+), 1 deletion(-) diff --git a/python/sglang/srt/models/deepseek_ocr.py b/python/sglang/srt/models/deepseek_ocr.py index b143ddfa4..fa334af7b 100644 --- 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,