Unsupported system error when installing on M1 / Apple Silicon

[nickplee]

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): macOS 11.0.1 (20B29)
• TensorFlow.js installed from (npm or script link): npm
• TensorFlow.js version: 2.8.3

Describe the problem

Installation fails due to a missing precompiled libtensorflow for darwin / arm64.

Apple has their own fork of TensorFlow supporting their chips-- is there a way to tell the installation script to use their libtensorflow.dylib?

Provide the exact sequence of commands / steps that you executed before running into the problem

npm install @tensorflow/tfjs-node --save

Any other info / logs

npm ERR! code 1
npm ERR! path /Users/nicklee/Documents/Development/<redacted>/src/<redacted>/node_modules/@tensorflow/tfjs-node
npm ERR! command failed
npm ERR! command sh -c node scripts/install.js
npm ERR! CPU-darwin-2.8.3.tar.gz
npm ERR! * Downloading libtensorflow
npm ERR! /Users/nicklee/Documents/Development/<redacted>/src/<redacted>/node_modules/@tensorflow/tfjs-node/scripts/install.js:100
npm ERR!     throw new Error(`Unsupported system: ${libType}-${platform}-${os.arch()}`);
npm ERR!           ^
npm ERR! 
npm ERR! Error: Unsupported system: cpu-darwin-arm64
npm ERR!     at getPlatformLibtensorflowUri (/Users/nicklee/Documents/Development/<redacted>/src/<redacted>/node_modules/@tensorflow/tfjs-node/scripts/install.js:100:11)
npm ERR!     at downloadLibtensorflow (/Users/nicklee/Documents/Development/<redacted>/src/<redacted>/node_modules/@tensorflow/tfjs-node/scripts/install.js:134:7)
npm ERR!     at async run (/Users/nicklee/Documents/Development/<redacted>/src/<redacted>/node_modules/@tensorflow/tfjs-node/scripts/install.js:197:5)

npm ERR! A complete log of this run can be found in:
npm ERR!     /Users/nicklee/.npm/_logs/2021-01-09T02_53_09_673Z-debug.log

[pyu10055]

@nickplee thank you for reporting this, can you help to test it out following the instruction of custom TF build
https://github.com/tensorflow/tfjs/tree/master/tfjs-node#optional-build-optimal-tensorflow-from-source

@yhwang any suggestions on this? thanks.

[yhwang]

since the tfjs-node@2.8.3 is using tensorflow shared libs v1.15.0, I think it's better to build the dependent shared libs by using the link @pyu10055 posted above. However, the tricky part would be tweaking the tensorflow to build on darwin arm64. It's okay by using ubuntu with arm64 to build tensorflow shared libs. But I am not sure about darwin with arm64.

For the tensorflow apple provides, you can verify whether you can find all files in this tarball from the binaries that apple provides. If yes, you can just put them under node_modules/@tensorflow/tfjs-node/lib/deps where you install the tfjs-node npm package. Then you can run npm run build-addon-from-source under node_modules/@tensorflow/tfjs-node. It will build the node binding for you. If everything goes well, you can try to use tfjs-node. Again, seems apple provides newer version then 1.15.0, I don't know if you can run tfjs-node without any issue even you can build the node binding successfully.

[yhwang]

@pyu10055 a side question: any idea of supporting more platforms and architectures officially?

[nickplee]

@yhwang @pyu10055

I actually already tried subbing in the Apple library and recompiling the binding. Once you install it, it includes the dylib as well as the headers.

I was hopeful, but ran in to two issues:

• TF_StringDecode, TF_StringEncode, and TF_StringEncodedSize are used by the binding and were deprecated/removed from the TensorFlow C API in 2.4.0
• After removing the usages of those functions (and replacing them with what I think is the right alternative implementation) the compiled binding crashed in weird ways. I don't know enough about TensorFlow to be able to explain in more detail unfortunately.

[pyu10055]

@yhwang It would definitely be useful to official support other platforms officially, but no plan as this moment, might be a good project for the SIG, any suggestions are welcome.

@nickplee Do you mind contriubting your fix attempt as an PR?

[google-ml-butler]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed in 7 days if no further activity occurs. Thank you.

[patrickhulce]

Perhaps with tensorflow/tensorflow#45404 this is more feasible? Has anyone been able to get tfjs-node to work on an M1 device?

We're extremely interested in this as well. If anyone has been able to get it working, we would very, very much appreciate some pointers :)

[google-ml-butler]

Closing as stale. Please @mention us if this needs more attention.

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No

[patrickhulce]

Can this be reopened or someone point us in the direction of where we should go to help? Support for the new M1 macs is essential for any applications running edge TF models.