spacy.require_gpu() is true but spacy cli uses CPU

[LarsNeR]

How to reproduce the behaviour

Use the Quickstart guide on https://spacy.io/usage/training with GPU

I run following code for testing GPU-availability:
import torch; torch.cuda.is_available() # --> True
import spacy; spacy.require_gpu() # --> True

But when I run python -m spacy train ./config.cfg ... like described in the Quickstart it says

"i Using CPU"

Tested that on my local windows machine and on AWS Sagemaker.

Your Environment

Windows machine:

• spaCy version: 3.0.3
• Platform: Windows-10-10.0.18362-SP0
• Python version: 3.8.6

AWS Sagemaker:

• spaCy version: 3.0.3
• Platform: Linux-4.14.214-118.339.amzn1.x86_64-x86_64-with-glibc2.9
• Python version: 3.6.10

The (important parts of the) config file look like that:

[system]
gpu_allocator="pytorch"

[nlp]
pipeline=["transformer", "ner"]

[components]

[components.transformer]
factory="transformer"

[components.transformer.model]
@architectures="spacy-transformers.TransformerModel.v1"
name="bert-base-german-cased"

I got the config from https://spacy.io/usage/training "Quickstart" selecting "ner" and "GPU" and I have not changed anything there. After that I did the init fill-config and then I tried the train resulting in the "i Using CPU"

I expected that when spacy.require_gpu() returns True the cli command would use the GPU.

[adrianeboyd]

This is a good type of question for our new discussions board, so I'll transfer it over there in just a minute. You'll get an automatic message saying this issue is locked, but the discussion can continue on the open thread there.

[adrianeboyd]

Whether the GPU is enabled during training is not stored in the config itself, since you might want to use the same config on machines with and without GPUs. There's a command-line flag for spacy train to select the GPU, e.g.:

python -m spacy train ./config.cfg --gpu-id 0

[LarsNeR]

Thanks that works, my mistake. But maybe that should be mentioned as a side-note at the Embeddings&Transformers Training or at the Quickstart?