Update modeling_quiet.py

modeling_quiet.py

@@ -1046,7 +1046,7 @@ class QuietModel(QuietPreTrainedModel):
             inputs_embeds = self.embed_tokens(input_ids)
 
         if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache:
-            is_padding_right = attention_mask[:, -1].sum().item() != batch_size
+            is_padding_right = (attention_mask[:, -1] == 0).any().item()
             if is_padding_right:
                 raise ValueError(
                     "You are attempting to perform batched generation with padding_side='right'"