[BUG] <title>batch_infer报错：'tuple' object has no attribute 'dtype'

[PineappleWill]

是否已有关于该错误的issue或讨论？ | Is there an existing issue / discussion for this?

• [X] 我已经搜索过已有的issues和讨论 | I have searched the existing issues / discussions

该问题是否在FAQ中有解答？ | Is there an existing answer for this in FAQ?

• [X] 我已经搜索过FAQ | I have searched FAQ

当前行为 | Current Behavior

AttributeError                            Traceback (most recent call last)
Cell In[1], line 242
    239 prompts_batch = split_list_into_batches(descriptions, batch_size)
    240 for i, prompts in tqdm(enumerate(prompts_batch)):
    241     # if i > 1: break
--> 242     batch_response = batch_infer(prompts, instruction_extract)
    243     for idx, tag in enumerate(batch_response):
    244         title_temp = titles[i*batch_size + idx]

Cell In[1], line 68, in batch_infer(all_raw_text, instruction)
     66 batch_input_ids = tokenizer(batch_raw_text, padding='longest')
     67 batch_input_ids = torch.LongTensor(batch_input_ids['input_ids']).to(model.device)
---> 68 batch_out_ids = model.generate(batch_input_ids, return_dict_in_generate=False, generation_config=model.generation_config)
     69 padding_lens = [batch_input_ids[i].eq(tokenizer.pad_token_id).sum().item() for i in range(batch_input_ids.size(0))]
     71 batch_response = [
     72     decode_tokens(
     73         batch_out_ids[i][padding_lens[i]:],
   (...)
     80     ) for i in range(len(all_raw_text))
     81 ]

File ~/.cache/huggingface/modules/transformers_modules/modeling_qwen.py:1259, in QWenLMHeadModel.generate(self, inputs, generation_config, logits_processor, stopping_criteria, prefix_allowed_tokens_fn, synced_gpus, assistant_model, streamer, **kwargs)
   1256     else:
   1257         logits_processor.append(stop_words_logits_processor)
-> 1259 return super().generate(
   1260     inputs,
   1261     generation_config=generation_config,
   1262     logits_processor=logits_processor,
   1263     stopping_criteria=stopping_criteria,
   1264     prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
   1265     synced_gpus=synced_gpus,
   1266     assistant_model=assistant_model,
   1267     streamer=streamer,
   1268     **kwargs,
   1269 )

File /opt/conda/lib/python3.8/site-packages/torch/utils/_contextlib.py:115, in context_decorator.<locals>.decorate_context(*args, **kwargs)
    112 @functools.wraps(func)
    113 def decorate_context(*args, **kwargs):
    114     with ctx_factory():
--> 115         return func(*args, **kwargs)

File /opt/conda/lib/python3.8/site-packages/transformers/generation/utils.py:1642, in GenerationMixin.generate(self, inputs, generation_config, logits_processor, stopping_criteria, prefix_allowed_tokens_fn, synced_gpus, assistant_model, streamer, negative_prompt_ids, negative_prompt_attention_mask, **kwargs)
   1634     input_ids, model_kwargs = self._expand_inputs_for_generation(
   1635         input_ids=input_ids,
   1636         expand_size=generation_config.num_return_sequences,
   1637         is_encoder_decoder=self.config.is_encoder_decoder,
   1638         **model_kwargs,
   1639     )
   1641     # 13. run sample
-> 1642     return self.sample(
   1643         input_ids,
   1644         logits_processor=logits_processor,
   1645         logits_warper=logits_warper,
   1646         stopping_criteria=stopping_criteria,
   1647         pad_token_id=generation_config.pad_token_id,
   1648         eos_token_id=generation_config.eos_token_id,
   1649         output_scores=generation_config.output_scores,
   1650         return_dict_in_generate=generation_config.return_dict_in_generate,
   1651         synced_gpus=synced_gpus,
   1652         streamer=streamer,
   1653         **model_kwargs,
   1654     )
   1656 elif generation_mode == GenerationMode.BEAM_SEARCH:
   1657     # 11. prepare beam search scorer
   1658     beam_scorer = BeamSearchScorer(
   1659         batch_size=batch_size,
   1660         num_beams=generation_config.num_beams,
   (...)
   1665         max_length=generation_config.max_length,
   1666     )

File /opt/conda/lib/python3.8/site-packages/transformers/generation/utils.py:2724, in GenerationMixin.sample(self, input_ids, logits_processor, stopping_criteria, logits_warper, max_length, pad_token_id, eos_token_id, output_attentions, output_hidden_states, output_scores, return_dict_in_generate, synced_gpus, streamer, **model_kwargs)
   2721 model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)
   2723 # forward pass to get next token
-> 2724 outputs = self(
   2725     **model_inputs,
   2726     return_dict=True,
   2727     output_attentions=output_attentions,
   2728     output_hidden_states=output_hidden_states,
   2729 )
   2731 if synced_gpus and this_peer_finished:
   2732     continue  # don't waste resources running the code we don't need

File /opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py:1501, in Module._call_impl(self, *args, **kwargs)
   1496 # If we don't have any hooks, we want to skip the rest of the logic in
   1497 # this function, and just call forward.
   1498 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
   1499         or _global_backward_pre_hooks or _global_backward_hooks
   1500         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1501     return forward_call(*args, **kwargs)
   1502 # Do not call functions when jit is used
   1503 full_backward_hooks, non_full_backward_hooks = [], []

File ~/.cache/huggingface/modules/transformers_modules/modeling_qwen.py:1043, in QWenLMHeadModel.forward(self, input_ids, past_key_values, attention_mask, token_type_ids, position_ids, head_mask, inputs_embeds, encoder_hidden_states, encoder_attention_mask, labels, use_cache, output_attentions, output_hidden_states, return_dict)
   1021 def forward(
   1022     self,
   1023     input_ids: Optional[torch.LongTensor] = None,
   (...)
   1036     return_dict: Optional[bool] = None,
   1037 ) -> Union[Tuple, CausalLMOutputWithPast]:
   1039     return_dict = (
   1040         return_dict if return_dict is not None else self.config.use_return_dict
   1041     )
-> 1043     transformer_outputs = self.transformer(
   1044         input_ids,
   1045         past_key_values=past_key_values,
   1046         attention_mask=attention_mask,
   1047         token_type_ids=token_type_ids,
   1048         position_ids=position_ids,
   1049         head_mask=head_mask,
   1050         inputs_embeds=inputs_embeds,
   1051         encoder_hidden_states=encoder_hidden_states,
   1052         encoder_attention_mask=encoder_attention_mask,
   1053         use_cache=use_cache,
   1054         output_attentions=output_attentions,
   1055         output_hidden_states=output_hidden_states,
   1056         return_dict=return_dict,
   1057     )
   1058     hidden_states = transformer_outputs[0]
   1060     lm_logits = self.lm_head(hidden_states)

File /opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py:1501, in Module._call_impl(self, *args, **kwargs)
   1496 # If we don't have any hooks, we want to skip the rest of the logic in
   1497 # this function, and just call forward.
   1498 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
   1499         or _global_backward_pre_hooks or _global_backward_hooks
   1500         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1501     return forward_call(*args, **kwargs)
   1502 # Do not call functions when jit is used
   1503 full_backward_hooks, non_full_backward_hooks = [], []

File ~/.cache/huggingface/modules/transformers_modules/modeling_qwen.py:891, in QWenModel.forward(self, input_ids, past_key_values, attention_mask, token_type_ids, position_ids, head_mask, inputs_embeds, encoder_hidden_states, encoder_attention_mask, use_cache, output_attentions, output_hidden_states, return_dict)
    880     outputs = torch.utils.checkpoint.checkpoint(
    881         create_custom_forward(block),
    882         hidden_states,
   (...)
    888         encoder_attention_mask,
    889     )
    890 else:
--> 891     outputs = block(
    892         hidden_states,
    893         layer_past=layer_past,
    894         rotary_pos_emb_list=rotary_pos_emb_list,
    895         attention_mask=attention_mask,
    896         head_mask=head_mask[i],
    897         encoder_hidden_states=encoder_hidden_states,
    898         encoder_attention_mask=encoder_attention_mask,
    899         use_cache=use_cache,
    900         output_attentions=output_attentions,
    901     )
    903 hidden_states = outputs[0]
    904 if use_cache is True:

File /opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py:1501, in Module._call_impl(self, *args, **kwargs)
   1496 # If we don't have any hooks, we want to skip the rest of the logic in
   1497 # this function, and just call forward.
   1498 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
   1499         or _global_backward_pre_hooks or _global_backward_hooks
   1500         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1501     return forward_call(*args, **kwargs)
   1502 # Do not call functions when jit is used
   1503 full_backward_hooks, non_full_backward_hooks = [], []

File ~/.cache/huggingface/modules/transformers_modules/modeling_qwen.py:610, in QWenBlock.forward(self, hidden_states, rotary_pos_emb_list, layer_past, attention_mask, head_mask, encoder_hidden_states, encoder_attention_mask, use_cache, output_attentions)
    596 def forward(
    597     self,
    598     hidden_states: Optional[Tuple[torch.FloatTensor]],
   (...)
    606     output_attentions: Optional[bool] = False,
    607 ):
    608     layernorm_output = self.ln_1(hidden_states)
--> 610     attn_outputs = self.attn(
    611         layernorm_output,
    612         rotary_pos_emb_list,
    613         layer_past=layer_past,
    614         attention_mask=attention_mask,
    615         head_mask=head_mask,
    616         use_cache=use_cache,
    617         output_attentions=output_attentions,
    618     )
    619     attn_output = attn_outputs[0]
    621     outputs = attn_outputs[1:]

File /opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py:1501, in Module._call_impl(self, *args, **kwargs)
   1496 # If we don't have any hooks, we want to skip the rest of the logic in
   1497 # this function, and just call forward.
   1498 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
   1499         or _global_backward_pre_hooks or _global_backward_hooks
   1500         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1501     return forward_call(*args, **kwargs)
   1502 # Do not call functions when jit is used
   1503 full_backward_hooks, non_full_backward_hooks = [], []

File ~/.cache/huggingface/modules/transformers_modules/modeling_qwen.py:499, in QWenAttention.forward(self, hidden_states, rotary_pos_emb_list, layer_past, attention_mask, head_mask, encoder_hidden_states, encoder_attention_mask, output_attentions, use_cache)
    492 if (
    493     self.use_flash_attn
    494     and flash_attn_unpadded_func is not None
    495     and not self.is_fp32
    496     and query.is_cuda
    497 ):
    498     q, k, v = query, key, value
--> 499     attn_output = self.core_attention_flash(q, k, v, attention_mask=attention_mask)
    500 else:
    501     key_size = key[0].size(2) if self.use_cache_quantization else key.size(1)

File /opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py:1501, in Module._call_impl(self, *args, **kwargs)
   1496 # If we don't have any hooks, we want to skip the rest of the logic in
   1497 # this function, and just call forward.
   1498 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
   1499         or _global_backward_pre_hooks or _global_backward_hooks
   1500         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1501     return forward_call(*args, **kwargs)
   1502 # Do not call functions when jit is used
   1503 full_backward_hooks, non_full_backward_hooks = [], []

File ~/.cache/huggingface/modules/transformers_modules/modeling_qwen.py:183, in FlashSelfAttention.forward(self, q, k, v, attention_mask)
    182 def forward(self, q, k, v, attention_mask=None):
--> 183     assert all((i.dtype in [torch.float16, torch.bfloat16] for i in (q, k, v)))
    184     assert all((i.is_cuda for i in (q, k, v)))
    185     batch_size, seqlen_q = q.shape[0], q.shape[1]

File ~/.cache/huggingface/modules/transformers_modules/modeling_qwen.py:183, in <genexpr>(.0)
    182 def forward(self, q, k, v, attention_mask=None):
--> 183     assert all((i.dtype in [torch.float16, torch.bfloat16] for i in (q, k, v)))
    184     assert all((i.is_cuda for i in (q, k, v)))
    185     batch_size, seqlen_q = q.shape[0], q.shape[1]

AttributeError: 'tuple' object has no attribute 'dtype'

期望行为 | Expected Behavior

No response

复现方法 | Steps To Reproduce

No response

运行环境 | Environment

- OS:
- Python:
- Transformers:
- PyTorch:
- CUDA (`python -c 'import torch; print(torch.version.cuda)'`):

备注 | Anything else?

No response

[PineappleWill]

没有用量化之类的，看起来用的是最普通的generate类时报的错

[PineappleWill]

报错日志可能有点乱，截图更直观一点

[github-actions[bot]]

This issue has been automatically marked as inactive due to lack of recent activity. Should you believe it remains unresolved and warrants attention, kindly leave a comment on this thread. 此问题由于长期未有新进展而被系统自动标记为不活跃。如果您认为它仍有待解决，请在此帖下方留言以补充信息。