Install nightly release of pytorch to enable ML support for arm CPUs

[ollywelch]

Hi there, thought I'd create a quick PR to enable the ML image to be built on ARM CPUs. As per pytorch's website, their nightly builds have ARM64 support. Using this would be great for us Raspberry Pi users!

On a side note, love this project, and keen to contribute back. Let me know if there's anything else I need to do for this to be considered :)

[vercel[bot]]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

1 Ignored Deployment

Name	Status	Preview	Comments	Updated
immich	⬜️ Ignored (Inspect)			Feb 26, 2023 at 2:40PM (UTC)

[alextran1502]

Hello, what is the image size for this change? We use the cpu only image to keep the size small

[ollywelch]

Hello, what is the image size for this change? We use the cpu only image to keep the size small

1.6 GB when I built it on my rpi4 - quite beefy

[bo0tzz]

This is great, thank you! I was going to mention that the armv7 build fails, but I see you already got it :)

Alex: this is still cpu-only as far as I can tell. Olly: AFAIK future gpu support is as easy as swapping cpu for cu118, but just to be sure, do you know whether the Cuda builds also support arm64?

[ollywelch]

Not sure about the CUDA builds, would have to check

[bo0tzz]

It seems like the current machine-learning image is 749MB. @ollywelch is the 1.6GB you mention for an arm-only image, or a multiplatform build?

[ollywelch]

It seems like the current machine-learning image is 749MB. @ollywelch is the 1.6GB you mention for an arm-only image, or a multiplatform build?

@bo0tzz sorry think I quoted the wrong figure there - compressed size is 557MB (built on my rpi4), just pushed it to docker hub. Also realised I removed the --no-cache-dir from the first pip install, so I've added that back in now!

[ollywelch]

Make that 483MB with the latest change :)