Error using GPU with the accuracy-pipeline (AttributeError: module 'torch._C' has no attribute '_cuda_setDevice')

[mscherrmann]

How to reproduce the behaviour

I created a virtual environment and installed spacy with the following steps:

1. conda create --name spacy python=3.6
2. conda activate spacy
3. pip install -U pip setuptools wheel
4. pip install -U spacy[cuda102]
5. python -m spacy download en_core_web_trf

Then I tried to run the following code in the console within the created virtual environment:
import spacy
spacy.require_gpu()

I get the following error message:

Traceback (most recent call last):

File "", line 1, in
spacy.require_gpu()

File "C:\ProgramData\Anaconda3\envs\spacy\lib\site-packages\thinc\util.py", line 184, in require_gpu
set_active_gpu(gpu_id)

File "C:\ProgramData\Anaconda3\envs\spacy\lib\site-packages\thinc\util.py", line 151, in set_active_gpu
torch.cuda.set_device(gpu_id)

File "C:\ProgramData\Anaconda3\envs\spacy\lib\site-packages\torch\cuda_init_.py", line 263, in set_device
torch._C._cuda_setDevice(device)

AttributeError: module 'torch._C' has no attribute '_cuda_setDevice'

Note:

• The code runs fine when I execute it before running step 5. in the installation process
• Spacy works fine if I run it on CPU instead of GPU
• If I install the efficiency-pipeline instead, both the CPU and GPU versions work fine

I expect that there is some update of torch when installing the accuracy pipeline using "python -m spacy download en_core_web_trf". The console outputs among others:

Collecting torchcontrib<0.1.0,>=0.0.2
Using cached torchcontrib-0.0.2-py3-none-any.whl
Collecting ftfy<6.0.0,>=5.0.0
Using cached ftfy-5.9-py3-none-any.whl
Collecting spacy-alignments<1.0.0,>=0.7.2
Using cached spacy_alignments-0.7.2-cp36-cp36m-win_amd64.whl (183 kB)
Collecting transformers<4.3.0,>=3.1.0
Using cached transformers-4.2.2-py3-none-any.whl (1.8 MB)
Collecting torch>=1.5.0
Using cached torch-1.7.1-cp36-cp36m-win_amd64.whl (184.1 MB)

In the attachment you will find the installed packages of the virtual environment:
venv_packages.txt

What can I do to fix this error? Thank you a lot for your help in advance.

Your Environment

• Operating System: Windows 10
• Python Version Used: 3.6.12
• spaCy Version Used: 3.0.3
• Pipelines: en-core-web-trf 3.0.0
• Environment Information: Cuda 10.2 (output of nvcc -V), Anaconda 2020.07, Nvidia GeForce 2080Ti

[adrianeboyd]

Hi, it sounds like something when wrong with the torch installation, which is unfortunately common with a plain pip install torch. (torch is being installed as a dependency of en_core_web_trf.)

Uninstall all the torch-related packages (or just create a new conda env) and try the directions here for using the pytorch quickstart, using conda instead of pip to install torch after step 2 above: https://spacy.io/usage/embeddings-transformers#transformers-installation

[mscherrmann]

Hi, thank you very much for your quick response. You are absolutly right, now everything works very well.

For completeness: The only change to the installation process described above is that I added the command
conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch between steps 2. and 3.