求助：Torch无float8_e4m3fn属性报错及CUDA版本冲突问题

在Jupyter Lab环境（Python 3.10.12）运行脚本时，导入Keras出现如下错误：

Traceback (most recent call last):
  File "/home/jupyter/work/resources/evalatin2024-latinpipe/latinpipe_evalatin24.py", line 25, in <module>
    import keras
  File "/home/jupyter/.local/lib/python3.10/site-packages/keras/__init__.py", line 4, in <module>
    from keras.api import DTypePolicy
  File "/home/jupyter/.local/lib/python3.10/site-packages/keras/api/__init__.py", line 8, in <module>
    from keras.api import activations
  File "/home/jupyter/.local/lib/python3.10/site-packages/keras/api/activations/__init__.py", line 7, in <module>
    from keras.src.activations import deserialize
  File "/home/jupyter/.local/lib/python3.10/site-packages/keras/src/__init__.py", line 1, in <module>
    from keras.src import activations
  File "/home/jupyter/.local/lib/python3.10/site-packages/keras/src/activations/__init__.py", line 3, in <module>
    from keras.src.activations.activations import elu
  File "/home/jupyter/.local/lib/python3.10/site-packages/keras/src/activations/activations.py", line 1, in <module>
    from keras.src import backend
  File "/home/jupyter/.local/lib/python3.10/site-packages/keras/src/backend/__init__.py", line 39, in <module>
    from keras.src.backend.torch import *  # noqa: F403
  File "/home/jupyter/.local/lib/python3.10/site-packages/keras/src/backend/torch/__init__.py", line 17, in <module>
    from keras.src.backend.torch import core
  File "/home/jupyter/.local/lib/python3.10/site-packages/keras/src/backend/torch/core.py", line 47, in <module>
    "float8_e4m3fn": torch.float8_e4m3fn,
AttributeError: module 'torch' has no attribute 'float8_e4m3fn'

环境信息 #

• 系统CUDA版本：12.2（通过!nvidia-smi查看）
• 关键已安装包：

keras                            3.5.0
tensorflow                       2.17.0
torch                            2.0.1+cu118
torchaudio                       2.0.2+cu118
torchvision                      0.15.2+cu118

• torch.version.cuda返回11.8（符合脚本要求）

解决方案 #

核心原因 #

Keras 3.5.0依赖的Torch版本需要支持float8_e4m3fn属性，该属性在Torch 2.1及以上版本才引入，而当前安装的Torch 2.0.1不具备此属性。同时脚本要求Torch对应CUDA 11.8，因此需要安装Torch 2.1.0+cu118版本，该版本既满足CUDA 11.8要求，又包含所需属性。

操作步骤 #

1. 卸载当前Torch相关包：

pip uninstall -y torch torchaudio torchvision

2. 安装兼容版本：

pip install torch==2.1.0+cu118 torchaudio==2.1.0+cu118 torchvision==0.16.0+cu118 --index-url https://download.pytorch.org/whl/cu118

3. 验证安装结果：

import torch
print(torch.version.cuda)  # 预期输出：11.8
print(hasattr(torch, 'float8_e4m3fn'))  # 预期输出：True

额外说明 #

系统CUDA 12.2可向下兼容CUDA 11.8的Torch包，无需降级系统CUDA。PyTorch自带独立的CUDA runtime，只要系统CUDA版本≥Torch所需版本即可正常运行。

内容的提问来源于stack exchange，提问作者LocusClassicus