illegal hardware instruction/Apple M1: python -m spacy download

[superseoworld]

first of all: spacy is great software.
i have a problem installing a model. calling python -m spacy download throws the error illegal harware instruction. what can i do? thanks a lot regards thomas

[polm]

Hey, sorry you're having trouble with this. We've had one report of this problem, though the user didn't follow up so we haven't been able to confirm details.

Can you give us the output of python -m spacy info or does that give an error too?

What version of Python are you using, and how did you install spaCy?

[superseoworld]

Hi, thanks for your reply.

python -m spacy info produces the same error.

I have installed spacy as follows:

pip install -U spacy

Thanks. Regards Thomas

[polm]

Based on allenai/scispacy#372 this might be an issue with nmslib.

Can you import nmslib?

You can try this to fix it:

pip uninstall nmslib
pip install --no-binary :all: nmslib

And then try import nmslib and using spaCy again.

[superseoworld]

Thanks.

nmslib is installable and importable. but python -m spacy info returns the same result.

[polm]

I see, in that case I suggest you follow @adrianeboyd's suggestions.

[adrianeboyd]

I suspect you've installed spacy in emulation mode and you're trying to run it in native mode, or vice versa.

It's also possible that python and pip don't point to the same installation? What do you get for which python and which pip?

You can see if one of these works:

arch -x86_64 python -m spacy info
arch -arm64 python -m spacy info

It can be a bit difficult to keep python virtual environments straight given how seamlessly emulation runs in the background. I've found it easier to use a virtual environment with a python binary that isn't universal, so that you know that as long as you're in this venv, there's only one compatible platform.

I think the native miniforge OS X arm64 installer is a good choice if you're primarily using spacy (using binary conda-forge packages for spacy). Or if you're also using prodigy and need everything to run in x86_64 emulation mode, then miniconda x86_64 with python 3.8 would be a better option. It depends a bit on what you need to run, and these aren't the only options.