docker : fix CPU ARM build

[slaren]

This should fix the CPU docker images for ARM. The build will only be optimized for the machine running the build (GGML_NATIVE is enabled).

[ngxson]

Can be add back -DLLAMA_CURL=ON ?

[slaren]

Forgot about that, added.

[ngxson]

I triggered the CI for building docker: https://github.com/ggerganov/llama.cpp/actions/runs/12965742642

[slaren]

Do you know how the docker CI is able to build arm containers? When I try docker build --platform=linux/arm64 it fails on the first RUN with exec /bin/sh: exec format error.

Looks like QEMU needs to be installed first.

[ngxson]

Hmm ok the CI fails due to internal compiler error: Segmentation fault signal terminated program cc1

[ngxson]

I think it's missing the apt install qemu-user-static step. I'm not sure why but from my other project, I have it on top of docker/setup-qemu-action@v2. Testing it now..

[slaren]

It is an issue of the ARM build with gcc. The actual error is a bit earlier than that:
cc1: error: unknown value 'native+nodotprod+noi8mm+nosve' for '-mcpu'

I can fix it, but we need to choose an arm version to build for. The problem is that the only version that makes sense for a generic container is the most generic build without any features enabled, and performance is going to be very bad. So I am not sure that it is even worth to make ARM images, until support is added for building multiple versions of the CPU backend, as is done for x86. Until then, people should build their own images, optimized for their hardware.

[ngxson]

Hmm ok I don't have the hardware to test it either, so I think we can disable it now and see if the community can help.

If we ever bring this back in the future, probably target commonly known HW like raspberry pi or Ampere cpus.

[slaren]

I will change it so that there is an argument to select the ARM architecture to build for, set to armv8-a by default (seems to be the oldest supported in docker). Otherwise we would need to delete all current ARM images to stop people from trying to use a broken docker image.