ci : split build.yml + server.yml

[ggerganov]

cont #20540

• Split build.yml + server.yml into parts and move some of the workflows in the new parts
• Continue to run build.yml + server.yml on all PRs and master branch
• The new build parts run only on master branch
• The new build parts can be manually triggered on PRs
• Fork and use setup-xcode action
• Rename workflow: drop -cmake, improve consistency
• Rename ggml-ci-arm64-graviton4-kleidiai -> ggml-ci-arm64-cpu-kleidiai-graviton4
• Remove ubuntu-llguidance-riscv64-native
• Remove ubuntu-cmake-rpc-riscv64-native

Note: not sure how the GA cache works, but I suspect that when we run this new CI it will trash the caches because of the new workflow names

Next PRs

• Move ggml-ci-*-cpu-* workflows to the self-hosted runners to free some GH cache
• Drop spacemit GH cache: ci : split build.yml + server.yml #20546 (comment)

[CISC]

Note: not sure how the GA cache works, but I suspect that when we run this new CI it will trash the caches because of the new workflow names

Since you also renamed the keys, yes, but since we just about can't fit a whole run in the 10GB we have available we are partially thrashing the cache already. This will just leave older PRs in the dark until they catch up, not a major issue I think.

[ggerganov]

I moved the sanitizer workflows in separate *-sanitize.yml files to be executed manually on PRs.

In the next PR, I think we can move some of the ggml-ci CPU workflows to self-hosted runners since we have both x86 and arm64 machines enrolled now. This will probably free up some GH cache.

[ggerganov]

I think we should drop the spacemit cache:

The build-riscv.yml should run only on changes to ggml/src/ggml-cpu/spacemit/* and should not use the GH cache.

[CISC]

I think we should drop the spacemit cache:

The build-riscv.yml should run only on changes to ggml/src/ggml-cpu/spacemit/* and should not use the GH cache.

It's not running on riscv, it's a crossbuild, IIRC it's a special arch.

[ggerganov]

The #20585 and #20588 should free quite some GH cache too.

[ggerganov]

• Extracted the android builds to run only on master
• For the split workflows, they now also run in the PRs if files are touched directly. For example, modifying ggml-metal sources will trigger the build-apple.yml workflow even in the PRs

[ggerganov]

I think I'll yolo this into master and see what happens.

[CISC]

I think I'll yolo this into master and see what happens.

https://github.com/ggml-org/llama.cpp/actions/runs/23111057141/job/67128454775

[ggerganov]

Should be fixed now - I incorrectly forked only the master branch earlier.