Apex: Segmentation fault

[lemonhu]

Hi, the problem Segmentation fault occurs when I use apex for mixed precision training, the test script is as follows:

import torch
import apex

input = torch.rand(3, 10).cuda()
fln = apex.normalization.FusedLayerNorm(10).cuda()
fln(input)

Then I refer to the answer to issue #156 and try to solve this problem. When I execute the following commands in turn. Environment: CUDA 10.0.30, torch-1.0.1.post2-cp37

$ pip uninstall apex
$ cd apex_repo_dir
$ rm-rf build
$ python setup.py install --cuda_ext --cpp_ext

It does not work for me, looking forward to your reply.

[mcarilli]

python setup.py install --cuda_ext --cpp_ext is not the recommended install command anymore. Try

$ pip uninstall apex
$ pip uninstall apex # repeat until you're sure it's gone, different install methods may create redundant installs
$ cd apex_repo_dir
$ rm -rf build # if "build" exists"
$ rm -rf apex.egg-info # if "apex.egg-info" exists
$  pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" .

I updated https://github.com/NVIDIA/apex/issues/156#issuecomment-465301976 to show the new command as well.

[lemonhu]

Thank you for your reply.
I don't know why it still has the above problems in the environment of gcc 4.8.5. However, when I upgrade the gcc environment to 4.9.2, re-clone the repository and follow your commands, it works for me.

[mcarilli]

I'm also not sure why gcc 4.8.5 fails, but thank you for the information, I'll remember this if future users have the same issue.