PyTorch降级至1.13.0后出现AttributeError: module 'torch' has no attribute 'version'求助
2026-6-26
问题：PyTorch 1.13.0加载Falcon-7B时出现
AttributeError: module 'torch' has no attribute 'version' 背景
因CUDA 11.3与PyTorch 2.2.2不兼容，将PyTorch降级至1.13.0以适配transformers库，但运行加载Falcon-7B模型的代码时触发上述错误。
环境信息
OS: Ubuntu 18.04 LTS CUDA: 11.3 GPU: NVIDIA P5000 Quadro IDE: Jupyter Notebook Environment: VirtualEnv (venv)
运行代码
# 导入所需库 import torch from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig # 指定Falcon模型名称 model_name = "ybelkada/falcon-7b-sharded-bf16" # 配置BitsAndBytes量化参数 bnb_config = BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.float16, ) # 加载带量化配置的Falcon模型 model = AutoModelForCausalLM.from_pretrained( model_name, quantization_config=bnb_config, trust_remote_code=True ) # 禁用模型缓存 model.config.use_cache = False
错误信息
--------------------------------------------------------------------------- AttributeError Traceback (most recent call last) File ~/FYP_Chatbot/test02/src/myenv-test002-02/lib/python3.9/site-packages/torch/cuda/__init__.py:242, in _lazy_init() 241 try: --> 242 queued_call() 243 except Exception as e: File ~/FYP_Chatbot/test02/src/myenv-test002-02/lib/python3.9/site-packages/torch/cuda/__init__.py:122, in _check_capability() 116 old_gpu_warn = """ 117 Found GPU%d %s which is of cuda capability %d.%d. 118 PyTorch no longer supports this GPU because it is too old. 119 The minimum cuda capability supported by this library is %d.%d. 120 """ --> 122 if torch.version.cuda is not None: # on ROCm we don't want this check 123 CUDA_VERSION = torch._C._cuda_getCompiledVersion() AttributeError: module 'torch' has no attribute 'version' The above exception was the direct cause of the following exception: DeferredCudaCallError Traceback (most recent call last) Cell In[10], line 17 10 bnb_config = BitsAndBytesConfig( 11 load_in_4bit=True, 12 bnb_4bit_quant_type="nf4", 13 bnb_4bit_compute_dtype=torch.float16, 14 ) 16 # Loading the Falcon model with quantization configuration ---> 17 model = AutoModelForCausalLM.from_pretrained( 18 model_name, 19 quantization_config=bnb_config, 20 trust_remote_code=True 21 ) 23 # Disabling cache usage in the model configuration 24 model.config.use_cache = False File ~/FYP_Chatbot/test02/src/myenv-test002-02/lib/python3.9/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 ~/FYP_Chatbot/test02/src/myenv-test002-02/lib/python3.9/site-packages/transformers/modeling_utils.py:3053, 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) 3049 hf_quantizer.validate_environment( 3050 torch_dtype=torch_dtype, from_tf=from_tf, from_flax=from_flax, device_map=device_map 3051 ) 3052 torch_dtype = hf_quantizer.update_torch_dtype(torch_dtype) -> 3053 device_map = hf_quantizer.update_device_map(device_map) 3055 # Force-set to `True` for more mem efficiency 3056 if low_cpu_mem_usage is None: File ~/FYP_Chatbot/test02/src/myenv-test002-02/lib/python3.9/site-packages/transformers/quantizers/quantizer_bnb_4bit.py:246, in Bnb4BitHfQuantizer.update_device_map(self, device_map) 244 def update_device_map(self, device_map): 245 if device_map is None: --> 246 device_map = {"": torch.cuda.current_device()} 247 logger.info( 248 "The device_map was not initialized. " 249 "Setting device_map to {'':torch.cuda.current_device()}. " 250 "If you want to use the model for inference, please set device_map ='auto' " 251 ) 252 return device_map File ~/FYP_Chatbot/test02/src/myenv-test002-02/lib/python3.9/site-packages/torch/cuda/__init__.py:552, in current_device() 550 def current_device() -> int: 551 r"""Returns the index of a currently selected device.""" -> 552 _lazy_init() 553 return torch._C._cuda_getDevice() File ~/FYP_Chatbot/test02/src/myenv-test002-02/lib/python3.9/site-packages/torch/cuda/__init__.py:246, in _lazy_init() 243 except Exception as e: 244 msg = (f"CUDA call failed lazily at initialization with error: {str(e)}\n\n" 245 f"CUDA call was originally invoked at:\n\n{orig_traceback}") --> 246 raise DeferredCudaCallError(msg) from e 247 finally: 248 delattr(_tls, 'is_initializing') DeferredCudaCallError: CUDA call failed lazily at initialization with error: module 'torch' has no attribute 'version'
环境依赖包
accelerate==0.29.1 bitsandbytes==0.43.0 datasets==2.18.0 einops==0.7.0 fsspec==2023.10.0 peft @ git+https://github.com/huggingface/peft.git@26726bf1ddee6ca75ed4e1bfd292094526707a78 torch==1.13.0 transformers==4.39.3 trl==0.8.1 wandb==0.16.6
NVIDIA显卡信息(nvidia-smi)
Sat Apr 6 22:40:45 2024 +-----------------------------------------------------------------------------+ | NVIDIA-SMI 470.182.03 Driver Version: 470.182.03 CUDA Version: 11.4 | |-------------------------------+----------------------+----------------------+ | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. | | | | MIG M. | |===============================+======================+======================| | 0 Quadro P5000 Off | 00000000:01:00.0 On | Off | | 27% 44C P8 6W / 180W | 295MiB / 16275MiB | 3% Default | | | | N/A | +-------------------------------+----------------------+----------------------+ +-----------------------------------------------------------------------------+ | Processes: | | GPU GI CI PID Type Process name GPU Memory | | ID ID Usage | |=============================================================================| | 0 N/A N/A 938 G /usr/lib/xorg/Xorg 103MiB | | 0 N/A N/A 1150 G /usr/bin/gnome-shell 37MiB | | 0 N/A N/A 1986 G /usr/lib/firefox/firefox 150MiB | +-----------------------------------------------------------------------------+
nvcc版本
nvcc: NVIDIA (R) Cuda compiler driver Copyright (c) 2005-2021 NVIDIA Corporation Built on Sun_Mar_21_19:15:46_PDT_2021 Cuda compilation tools, release 11.3, V11.3.58 Build cuda_11.3.r11.3/compiler.29745058_0
解决方法
1. 修复PyTorch安装
该错误多因PyTorch安装不完整、CUDA组件缺失导致。先验证安装状态：
import torch print(torch.__version__) print(torch.cuda.is_available())
若输出异常，卸载后重新安装适配CUDA 11.3的PyTorch 1.13.0：
pip uninstall torch -y pip install torch==1.13.0+cu113 torchvision==0.14.0+cu113 torchaudio==0.13.0 --extra-index-url https://download.pytorch.org/whl/cu113
2. 降级transformers版本
当前transformers 4.39.3与PyTorch 1.13.0兼容性差，降级至适配版本：
pip install transformers==4.26.1
3. 调整量化配置参数
PyTorch 1.13.0对nf4量化支持有限，修改BitsAndBytes配置：
bnb_config = BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_quant_type="fp4", bnb_4bit_compute_dtype=torch.float32, )
4. 显式指定设备映射
加载模型时设置device_map='auto'，避免自动获取CUDA设备时触发错误：
model = AutoModelForCausalLM.from_pretrained( model_name, quantization_config=bnb_config, trust_remote_code=True, device_map='auto' )
内容的提问来源于stack exchange，提问作者Muhammad Omar Farooq
相关产品推荐
相关产品推荐
