Construction Accident Prevention and Response Generation: Hansol Deco Season 3 Generative AI Competition

베이스라인 코드에서 이런 오류가 발생합니다.

transformers, accelerate, bitsandbytes 최신 업데이트도 다 해보고

python version 3.12 뿐만 아니라 3.11, 3.8 에서도 전부 똑같은 에러가 발생합니다.

이것을 어떻게 해결해야 하는지 전문가님들의 해결책을 구합니다.

감사합니다.

bnb_config = BitsAndBytesConfig(

  load_in_4bit=True,

  bnb_4bit_use_double_quant=True,

  bnb_4bit_quant_type="nf4",

  bnb_4bit_compute_dtype=torch.bfloat16

)

# model_id = "ybelkada/falcon-7b-sharded-bf16"

model_id = "NCSOFT/Llama-VARCO-8B-Instruct"

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(model_id, 

                       quantization_config=bnb_config, 

                       device_map='auto',

                      torch_dtype=torch.bfloat16,

                      trust_remote_code=True)

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Cell In[7], line 6
      3 # model_id = 'Bllossom/llama-3.2-Korean-Bllossom-3B'
      5 tokenizer = AutoTokenizer.from_pretrained(model_id)
----> 6 model = AutoModelForCausalLM.from_pretrained(model_id, 
      7                                              quantization_config=bnb_config, 
      8                                              device_map='auto',
      9                                             torch_dtype=torch.bfloat16,
     10                                             trust_remote_code=True)
     12 model.to('cpu')
     13 if torch.cuda.is_available():

File ~\AppData\Local\Programs\Python\Python312\Lib\site-packages\transformers\models\auto\auto_factory.py:563, in _BaseAutoModelClass.from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs)
    561 elif type(config) in cls._model_mapping.keys():
    562     model_class = _get_model_class(config, cls._model_mapping)
--> 563     return model_class.from_pretrained(
    564         pretrained_model_name_or_path, *model_args, config=config, **hub_kwargs, **kwargs
    565     )
    566 raise ValueError(
    567     f"Unrecognized configuration class {config.__class__} for this kind of AutoModel: {cls.__name__}.\n"
    568     f"Model type should be one of {', '.join(c.__name__ for c in cls._model_mapping.keys())}."
    569 )

File ~\AppData\Local\Programs\Python\Python312\Lib\site-packages\transformers\modeling_utils.py:3820, in PreTrainedModel.from_pretrained(cls, pretrained_model_name_or_path, config, cache_dir, ignore_mismatched_sizes, force_download, local_files_only, token, revision, use_safetensors, *model_args, **kwargs)
   3818         device_map_kwargs["force_hooks"] = True
   3819     if not is_fsdp_enabled() and not is_deepspeed_zero3_enabled():
-> 3820         dispatch_model(model, **device_map_kwargs)
   3822 if hf_quantizer is not None:
   3823     hf_quantizer.postprocess_model(model)

File ~\AppData\Local\Programs\Python\Python312\Lib\site-packages\accelerate\big_modeling.py:496, in dispatch_model(model, device_map, main_device, state_dict, offload_dir, offload_index, offload_buffers, skip_keys, preload_module_classes, force_hooks)
    494     device = f"xpu:{device}"
    495 if device != "disk":
--> 496     model.to(device)
    497 else:
    498     raise ValueError(
    499         "You are trying to offload the whole model to the disk. Please use the `disk_offload` function instead."
    500     )

File ~\AppData\Local\Programs\Python\Python312\Lib\site-packages\transformers\modeling_utils.py:2702, in PreTrainedModel.to(self, *args, **kwargs)
   2700 # Checks if the model has been loaded in 8-bit
   2701 if getattr(self, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES:
-> 2702     raise ValueError(
   2703         "`.to` is not supported for `4-bit` or `8-bit` bitsandbytes models. Please use the model as it is, since the"
   2704         " model has already been set to the correct devices and casted to the correct `dtype`."
   2705     )
   2706 elif getattr(self, "quantization_method", None) == QuantizationMethod.GPTQ:
   2707     # For GPTQ models, we prevent users from casting the model to another dytpe to restrict unwanted behaviours.
   2708     # the correct API should be to load the model with the desired dtype directly through `from_pretrained`.
   2709     dtype_present_in_args = False

ValueError: `.to` is not supported for `4-bit` or `8-bit` bitsandbytes models. Please use the model as it is, since the model has already been set to the correct devices and casted to the correct `dtype`.