Bump llama.cpp to b6002

[fellhorn]

Bump llama.cpp to b6002.

[fellhorn]

A prerequisite for multimodal support - a couple of bug fixes have been added in the meantime.

[MarcusDunn]

Thanks for the PR, https://github.com/utilityai/llama-cpp-rs/actions/runs/16556434445/job/46873866455?pr=786 should be passing before merge.

[MarcusDunn]

merging main may fix.

[fellhorn]

There is a chicken-egg issue in CI: cargo publish --dry-run --package llama-cpp-2 requires that the new version of llama-cpp-sys-2 has already been published. The step fails at it uses the currently published version instead of the workspace version.

How was that solved before?

I also noticed that llama-cpp-sys-2 is not pinned as a dependency when publishing the crate. Maybe cargo-release could simplify that process.

[fellhorn]

The other failure seems to be a temporary (?) issue when pulling nvcr.io/nvidia/cuda:12.3.1-devel-ubuntu22.04

[MarcusDunn]

There is a chicken-egg issue in CI: cargo publish --dry-run --package llama-cpp-2 requires that the new version of llama-cpp-sys-2 has already been published. The step fails at it uses the currently published version instead of the workspace version.

How was that solved before?

I also noticed that llama-cpp-sys-2 is not pinned as a dependency when publishing the crate. Maybe cargo-release could simplify that process.

Yeah, I don't think we actually need to test publishing llama-cpp-2 (llama-cpp-sys-2 is always the problem one). This has been yet to be solved. I'm currently ignoring it until someone is more bothered than I am.

It seems the CUDA images are now behind some auth, which is unfortunate. I'll likely remove those checks - don't worry about those.

However, the tests need to pass: https://github.com/utilityai/llama-cpp-rs/actions/runs/16556434445/job/46873866455?pr=786

[fellhorn]

The multi-package publishing function for cargo would be needed here. Unfortunately it is still only available in nightly, see rust-lang/cargo#15636

cargo publish --workspace --dry-run runs through on my branch.

But the cargo publish --package llama-cpp-2 --verbose --dry-run, used in CI, does not as it uses llama-cpp-sys2 v0.1.112 from crates.io (logs) instead of the version in the workspace.

cargo test works in CI and locally.

[fellhorn]

There are other release tools out there to work around the multi-package publishing issue - or we just wait for cargo publish --workspace to reach the stable channel.