docker : add GGML_CPU_ARM_ARCH arg to select ARM architecture

[slaren]

cont #11403

[ngxson]

I'm testing it on my cloned repo

[ngxson]

https://github.com/ngxson/llama.cpp/actions/runs/12966267528

[ngxson]

It worked, so I'm merging it now. Thanks.

[ngxson]

Not entirely sure why, but CI to build docker still fails with c++: internal compiler error: Segmentation fault signal terminated program collect2 while it works perfectly on my clone repo (which is auto-sync with ggerganov/llama.cpp master branch every 5 minutes)

At this point, the only explanation I can think of is that something happening at runner's hardware level

[slaren]

That's weird, I also tested it locally and it completes successfully on my machine with QEMU.

[ngxson]

Hmm sorry I overlooked a bit, the scheduled build failed on my clone repo too: https://github.com/ngxson/llama.cpp/actions/workflows/docker.yml

I'm re-running another build from xsn/test_docker_arm branch which is known to work yesterday. If it doesn't work, there is definitely something to do with runner's hardware: https://github.com/ngxson/llama.cpp/actions/runs/12976884398

[slaren]

If that doesn't work, may be worth to try disabling the cache, or building each platform separately in sequence.

[ngxson]

It fails, I'm running another one with AMD/ARM as separated jobs: https://github.com/ngxson/llama.cpp/actions/runs/12977065608