I am working with tensorflowjs in node. when i run the code, "const tf = require('@tensorflow/tfjs-node');" it throw an error "zsh: illegal hardware instruction node app.js".
i am using macbook air M1 chip.
1 Answer
Official @tensorflow/tfjs-node does not support Apple's M1 CPU just yet
There are unofficial ways to make it work (by using M1 build of tensorflow.so and rebuilding NodeJS bindings), but that is anything but trivial
UPDATE: as requested, a bit more info
there is Apple's port of Tensorflow that works on M1, but Apple stopped developing it and last update is from ~2 years: https://github.com/apple/tensorflow_macos
and there are several 3rd party ports, for example https://towardsdatascience.com/installing-tensorflow-on-the-m1-mac-410bb36b776
also, suggest going through TF's GitHub for M1 issues: https://github.com/tensorflow/tensorflow/issues?q=M1
in either case, once you have tensorflow, you still need to manually rebuild @tensorflow/tfjs-node so it binds to it which is non-trivial.
4 Comments
Sarath Soman
Thanks for the response. Can you explain the one you mentioned, or can you share any links.
Yala Yala Herzlin
I find it funny that I get this error trying to run the node.js version of the examples of github.com/justadudewhohacks/face-api.js, but, the browser examples run smoothly and don't have any kind of errors... any hint of why this different behaviour? This makes me think its not a M1 related issue... or is it?
Vladimir Mandic
when you're running in browser, you're running JS compiled to pseudo-code interpreted and executed by browser (and in case of WebGL backend, GLSL code compiled to WebGL), never native instructions. so its normal that you see issue with
tfjs-node which tries to run native CPU instructions and not in browser.Yala Yala Herzlin
thanks for clarifying. i've hit a wall, do you recommend any of the ports above to be able to run it server side?
