[Bugfix] qwen2.5-vl spec decode accept_len low (#13904)

Co-authored-by: Yuan Luo <yuan.luo@hotmail.com>

python/sglang/srt/models/qwen2_5_vl.py

@@ -654,6 +654,7 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
                     hidden_states,
                     self.lm_head,
                     forward_batch,
+                    aux_hidden_states,
                 )
             else:
                 return self.pooler(hidden_states, forward_batch)