feat!: upgrade llama.cpp from b10797 to b10817

[bernardladenthin]

Summary

No project-source change. The whole review surface across the paths this project links against is 7 files, +12 / −13; everything else in the 20-commit range is GPU-backend work (ggml-opencl +940 lines and a new 871-line Adreno SDPA kernel, Metal fa-vec tunings for M3/M3 Max, SYCL memtrace), a new model architecture (src/models/hy-v4.cpp), upstream CI and the WebUI.

Two entries needed checking rather than skimming:

• common/build-info.h — llama_print_build_info(const char *) became llama_print_build_info(const char *, FILE * = stderr) (#28322). A default argument, so every existing one-argument call still compiles; the project does not call it at all and reads llama_build_info(), which is untouched. Worth flagging anyway, because build-info is exactly what NativeLibraryLoadSmokeTest cross-checks against the pin — and this is the second consecutive range to move that file (b10792→b10797 carried #28278).
• ggml/include/ggml-backend.h REMOVES ggml_backend_op_alloc_size_may_expand (via the #28379 ggml sync). A deletion from a public header is the one shape that breaks a build with no diff hunk to notice, so the project source was grepped for it: zero references, in src/main/cpp and src/test/cpp alike.

The server contract was re-checked mechanically and is byte-identical (request fields, set_hard_limits bounds, response keys in both emit forms) — tools/server/ has no .cpp/.h change in the range at all. mtmd.cpp/clip.cpp change gemma4 vision handling and an image-token limit inside upstream translation units, with no API movement. The remaining three files are upstream's own version bump and a BoringSSL pin that only applies with LLAMA_BUILD_BORINGSSL, which this project leaves off.

Patches: zero intersection with the 42 files the eight patches touch, so no context can have moved. The applier was run for real anyway — fresh build dir, configure clean, stamp at head cd8cdf39, all eight hashes recorded.

Chunking: the full diff is 307 KB, over the runbook's 100 KiB threshold, but excluding ggml/src, tools/ui, docs, .github and conversion — every one outside the review surface the rule exists to bound — it is 68 KB, under it. Same call as the previous bump, and here the excluded-diff figure alone justifies it. Recorded in docs/history/ so the deviation stays auditable.

⚠️ This does NOT fix the macOS 15 regression

src/llama-model-loader.cpp is untouched in this range and hparams.n_expert_used_max() — the prime suspect from #28323, introduced at b10797 — is still called at both sites. Expect Java Tests macOS 15 arm64 (Metal) and (no Metal) to stay red. The bisect I proposed in #414 is still the open item; say the word and I'll push a probe branch pinned back to b10792.

Test plan

• Fresh rm -rf build && cmake -B build -DBUILD_TESTING=ON — the fail-loud PATCH_COMMAND ran for real; stamp cd8cdf397dc21d986e69069ce2180651d693fdff
• Full native build, zero warnings on the project TUs
• ctest 520/520
• mvn clean verify — 1486 Java tests, 0 failures
• .github/verify-bytecode-version.sh --max-major 52 llama/target clean
• CI is green on this branch — macOS 15 is expected to stay red, see above
• Docs updated — four pin files, a folded b10731 → b10817 CHANGELOG entry, two new rows in docs/history/llama-cpp-breaking-changes.md

A note on the verification, because it is the reason this is trustworthy rather than assumed: my first mvn clean verify on this branch failed, because a repo-wide git reset --hard I ran for an unrelated dependency audit had silently discarded the uncommitted pin edits while the native library was already built at b10817. NativeLibraryLoadSmokeTest.nativeBuildInfoMatchesPinnedVersionConstant caught exactly that drift and named it precisely (Linked build-info "b10817-cd8cdf397" must start with the pinned tag "b10797-"). The edits were restored and the full gate re-run green. The guard works.

Checklist

• I have read CONTRIBUTING.md and CODE_OF_CONDUCT.md
• My commits follow Conventional Commits
• No security-sensitive changes

🤖 Generated with Claude Code

https://claude.ai/code/session_01AnNYn8W1xuVxVJtyL34GyH

---

Generated by Claude Code

[claude]

Code Review Summary ✅

Version pins are consistent and correct:

• CMakeLists.txt: GIT_TAG b10817
• LlamaCppVersion.java: LLAMA_CPP_VERSION = "b10817"
• README.md: Badge updated
• CLAUDE.md: All 4 references updated
• CHANGELOG.md: Detailed documentation
• docs/history: Two comprehensive rows added

Verification Quality ⭐

• 520/520 C++ tests passing
• 1486 Java tests passing
• Patch applier verified on fresh build
• Bytecode gate clean
• Server contract byte-verified
• API removals checked (zero refs)

API Compatibility ✅

• common/build-info.h: llama_print_build_info() gains optional FILE * (backward compatible)
• ggml/include/ggml-backend.h: ggml_backend_op_alloc_size_may_expand removed (zero refs in project)

Known Limitations

• macOS-15 regression from b10797 correctly documented as unresolved

Recommendation: APPROVED — Thoroughly verified and well-documented. Ready for merge.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud