[willsheldon]

Moves the llama-cpp-sys-2/llama.cpp submodule from v0.6.0 (d812350) to v0.7.0 (620d222).

The range includes ggml-org/llama.cpp#30054 (4fbc76d), which adds the gemma-embedding2 architecture. Without it, loading embeddinggemma-2-Q8_0.gguf fails with unknown model architecture: 'gemma-embedding2'. 4fbc76d is an ancestor of v0.7.0.

API change ported

llama_perf_context_data gains n_reallocs_same and n_reallocs_diff. LlamaTimings::new builds the struct literal, so it now initialises both fields to 0. Its public signature does not change. The C++ wrappers compile unchanged.

Behaviour change to note

ggml-org/llama.cpp#30445 and #30255 stop computing logits for pooled embedding contexts (embeddings enabled and a pooling type other than None, set explicitly or taken from the GGUF metadata). On such a context, after decode, LlamaContext::get_logits and get_logits_ith now panic, and LlamaSampler::sample aborts on GGML_ASSERT(logits != nullptr). With v0.6.0 these calls returned logits. The in-repo embeddings and reranker examples read only embeddings and are not affected. This PR does not change the wrappers for this; it may be worth a note in the with_embeddings docs.

Testing

macOS arm64, Rust 1.98.0, with the commands from llama-cpp-rs-check.yml:

• cargo fmt --check: pass
• cargo clippy: pass (warnings only, none from this change)
• cargo build --features sampler,mtmd,llguidance: pass
• cargo test --features sampler,mtmd,llguidance: pass (23 unit, 4 integration, 94 doc tests; 3 doc tests ignored)
• cargo publish --workspace --dry-run: pass

Functional check with ggml-org/embeddinggemma-2-GGUF embeddinggemma-2-Q8_0.gguf (sha256 2188ac1deca4b77dffefd603c2776a9d76d9d74ec01841392982ebb840b09135), using cargo run --release -p embeddings -- local <model>:

• At v0.7.0 the model loads (arch = gemma-embedding2) and the example returns a 768-dim embedding (n_embd_out = 768, n_embd = 512).
• At v0.6.0 the same command fails with unknown model architecture: 'gemma-embedding2'.

Linux, Windows, CUDA and Emscripten builds were not run locally; CI covers them. The dynamic-link feature was not built locally and is not in CI. ggml-org/llama.cpp#30014 hides non-GGML_API symbols in shared ggml builds; every ggml function the crate calls is GGML_API, so no break is expected.

🤖 Generated with Claude Code

[willsheldon]

Branch updated with current main (cc 1.8.0). This bumps llama.cpp to v0.7.0, which adds EmbeddingGemma 2 support. All CI targets passed on the previous head. Could a maintainer review when convenient? We are pinned to this rev downstream until a release includes it.

[madsmtm]

Please expose the new fields in LlamaTimings::new, we don't worry too much about breaking API changes (yet).

[willsheldon]

Done in 424d1ff: LlamaTimings::new now takes n_reallocs_same and n_reallocs_diff. Thanks for the review.