Is thinc-apple-ops used when training ner component?

[kinghuang]

I'm training a ner component for the first time on a MacBook Pro with an Apple M1 Pro processor. I have the thinc-apple-ops module installed (via spacy[apple]). But, the training time of the ner component doesn't seem to differ with or without the module.

Should I expect a speedup when training the ner component with thinc-apple-ops installed? If so, is there something I need to configure to take advantage of it? The spaCy 3.2 blog post shows an 8✕ prediction speed improvement of the de_core_news_lg pipeline. Should I see something similar for training?

[adrianeboyd]

It should be similar because there's a large proportion of gemm in both: https://github.com/explosion/thinc-apple-ops#training

The only thing you might need to double-check is that you have a new enough version of thinc. thinc-apple-ops is used by default if available starting in v8.0.11.