Original ImportError was: ModuleNotFoundError("No module named 'amp_C')

[misslibra]

my env is CUDA 10.1, and pytorch 1.3,GPU v100,python 3.6,
and I get error:

Warning: multi_tensor_applier fused unscale kernel is unavailable, possibly because apex was installed without --cuda_ext --cpp_ext. Using Python fallback. Original ImportError was: ModuleNotFoundError("No module named 'amp_C'",)
Traceback (most recent call last):
File "train.py", line 104, in
lr_scheduler = config_lr_scheduler(args, optimizer)
File "/home/perception/Documents/apex-master/geely_yolo3D/utils/utils.py", line 553, in config_lr_scheduler
return torch.optim.lr_scheduler.StepLR(optimizer, args.lr_decay_epoch, args.gamma)
File "/home/perception/anaconda3/envs/new/lib/python3.6/site-packages/torch/optim/lr_scheduler.py", line 208, in init
super(StepLR, self).init(optimizer, last_epoch)
File "/home/perception/anaconda3/envs/new/lib/python3.6/site-packages/torch/optim/lr_scheduler.py", line 58, in init
self.optimizer.step = with_counter(self.optimizer.step)
File "/home/perception/anaconda3/envs/new/lib/python3.6/site-packages/torch/optim/lr_scheduler.py", line 40, in with_counter
instance_ref = weakref.ref(method.self)
AttributeError: 'function' object has no attribute 'self'

My code run well in env : CUDA 9,pytorch 1.2,GPU P4000,python 3.6

Please help me with this issue.

[mcarilli]

Should by fixed by #553. Try puling the latest master and reinstalling:

apex $ git pull
apex $ pip uninstall apex # repeat until you're sure it's gone
apex $ pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

[misslibra]

Thanks for your reply !
I download code zip yesterday ,which should be the latest version.
and I have fixed the problem by pip install using pip in conda--bin install of pip default .

[soodeh-nilforoushan]

Hi, I did not understand how you fix this error? I was stuck for two days for this error, if someone can help me, I would be happy!

[CRaimondo]

Also getting this error and not sure how you fixed @misslibra

[JoshLuxton]

Phew. this was hard to get working.

To the couple new commenters from the past week - this won't work in an environment with cuda-tookit-11+ you will need cuda-toolkit-10-2.

don't touch your cuda drivers just follow the instructions here: https://www.mathworks.com/matlabcentral/answers/687133-how-to-install-supported-cuda-toolkit-10-2-on-ubuntu-20-04-lts-linux

when it references: https://developer.nvidia.com/cuda-10.2-download-archive?target_os=Linux&target_arch=x86_64&target_distro=Ubuntu&target_version=1804&target_type=debnetwork

follow all steps from the archive EXCEPT when it says to install cuda (this will overwrite your drivers and cause incompatibility). Instead, install cuda-toolkit-10-2

this should update nvcc -V to reference 10.2, but if it doesnt check /usr/local/cuda:

ls -la /usr/local/cuda*

often multiple toolkits can live there.. and often you will see a symlink from /usrl/local/cuda to whichever version of cuda nvcc will reference for compile (you can replace the symlink to your desired version).

after that is setup and you follow the instructions for setting alternative gcc version you can proceed with the install listed in mcarills comment from Oct 31, 2019

[alexqdh]

After install apex, it still has some errors

1. ModuleNotFoundError: No module named 'packaging' Error occurs when build apex latest code. I checkout the previous one to fix it.

git clone https://github.com/NVIDIA/apex
cd apex
git checkout 22.04-dev
pip install -r requirements.txt
pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

2. ImportError: libc10.so. It should import torch before amp_C, use like this:

import torch
import amp_C

Install torch-1.12.1 gpu version:

pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 --extra-index-url https://download.pytorch.org/whl/cu113

[liaochuanlin]

ERROR: Could not build wheels for apex, which is required to install pyproject.toml-based projects

[ming-afk]

ERROR: Could not build wheels for apex, which is required to install pyproject.toml-based projects

maybe try this:
#1594 (comment)

[hayatkhan8660-maker]

my env is CUDA 10.1, and pytorch 1.3,GPU v100,python 3.6, and I get error:

Warning: multi_tensor_applier fused unscale kernel is unavailable, possibly because apex was installed without --cuda_ext --cpp_ext. Using Python fallback. Original ImportError was: ModuleNotFoundError("No module named 'amp_C'",) Traceback (most recent call last): File "train.py", line 104, in lr_scheduler = config_lr_scheduler(args, optimizer) File "/home/perception/Documents/apex-master/geely_yolo3D/utils/utils.py", line 553, in config_lr_scheduler return torch.optim.lr_scheduler.StepLR(optimizer, args.lr_decay_epoch, args.gamma) File "/home/perception/anaconda3/envs/new/lib/python3.6/site-packages/torch/optim/lr_scheduler.py", line 208, in init super(StepLR, self).init(optimizer, last_epoch) File "/home/perception/anaconda3/envs/new/lib/python3.6/site-packages/torch/optim/lr_scheduler.py", line 58, in init self.optimizer.step = with_counter(self.optimizer.step) File "/home/perception/anaconda3/envs/new/lib/python3.6/site-packages/torch/optim/lr_scheduler.py", line 40, in with_counter instance_ref = weakref.ref(method.self) AttributeError: 'function' object has no attribute 'self'

My code run well in env : CUDA 9,pytorch 1.2,GPU P4000,python 3.6

Please help me with this issue.

@misslibra
Have you solved this issue?

I am having the same issue..