Publish builds for aarch64 Linux

[Delamcode]

Could be get a Linux Arm64 binary? Thank you!

[krixoko]

would also appreciate

[Night3890]

I would also like this so I can run GPT4All in termux

[knigth95]

+1，插眼

[Euro-pol]

+1

[mael-bomane]

+1 🥰

[beeemT]

+1

[Orgoth]

+1

[ManjTi]

+1

[canyonrobins]

Is there any workaround for building and installing on Linux arm64?

[Xatinni]

+1, jetson nano cuba

[netandreus]

+1, waiting for this

[slavekrouta]

+1

[devloic]

+1

[baselakasha]

Is there any workaround for building and installing on Linux arm64?

I was able to build the backend and the python package by following this documentation: https://github.com/nomic-ai/gpt4all/blob/main/gpt4all-bindings/python/README.md

[benhaotang]

Hi, I build this on the Jetson dev kit with Ubuntu 20.04 aarch64 with following requirements

• qt 6.7.0 (using aqt installer to install linux_arm64 version of qt 6.7.0 with all the dependencies, e.g. aqt install-qt linux_arm64 desktop 6.7.0 -m all) now only 6.7.0 seems to have everything needed for compiling this.
• g++ and GCC 10.5.0 (you can either build it or find it in some test repo)
• cmake >= 3.22.0 build from source

when you compile the GPT4All, you need to provide the path to cudatoolkit as the newly compiled cmake 3.22 don't able to find it by itself for some reason ~at least for me~ (e.g. cmake . -DCUDAToolkit_ROOT=/usr/local/cuda-11.4)

go to the build folder for "gpt4all-chat", note that the qt creator come with the 20.04 system don't work properly(~might be my problem~), so I use cli instead,

qmake -project # make sure it's the 6.7.0 version instead of the system one!!!
qmake
cmake . -DCUDAToolkit_ROOT=/usr/local/cuda-11.4 # might be redundant for you
make -j4

and everything should compile and run normally, the cuda acceleration works very well!

Note: for 8G RAM devkit, if you want to use 7b models, please build and use python backend only, using QT chat UI only support 4Gram models without crashing due to low memory(even headless startup).

Note2: it can also built without cuda, so I think this should also work for raspberry pi.

Note3: please compile from the main branch or at least v2.8.0 and above, as tested v2.7.5 and below will not compile with this method.