fix batch infer

modeling_baichuan.py

@@ -358,9 +358,12 @@ class BaichuanModel(BaichuanPreTrainedModel):
             expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
                 inputs_embeds.device
             )
-            combined_attention_mask = (
-                expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
-            )
+            if combined_attention_mask is None:
+                combined_attention_mask = expanded_attn_mask
+            else:
+                expanded_attn_mask = torch.where(expanded_attn_mask == torch.finfo(inputs_embeds.dtype).min, torch.finfo(inputs_embeds.dtype).min / 2, expanded_attn_mask)
+                combined_attention_mask = torch.where(combined_attention_mask == torch.finfo(inputs_embeds.dtype).min, torch.finfo(inputs_embeds.dtype).min / 2, combined_attention_mask)
+                combined_attention_mask = expanded_attn_mask + combined_attention_mask
 
         return combined_attention_mask