feat!: upgrade llama.cpp from b10731 to b10797

[bernardladenthin]

Summary

Two commits, b10731 → b10792 → b10797. No project-source change was needed — and that was established, not assumed.

Headers on the priority review list. Every move in the whole range is additive or a widening const-qualification, so existing call sites still compile:

Header	Change
common/common.h	additive (common_params::preserve_reasoning_specified)
common/log.h	additive (common_log_get_verbosity)
ggml/include/ggml.h	additive (ggml_flash_attn_ext_set_n_kv_max)
tools/mtmd/mtmd.h	additive (mtmd_input_part, mtmd_tokenize_from_parts) + const-widening; bitmaps re-typed const T** → const T* const*
tools/mtmd/mtmd-helper.h	const-widening only, six functions
tools/mtmd/CMakeLists.txt	+models/deepseek4v.cpp

The project's only direct mtmd calls are mtmd_helper_bitmap_init_from_file + mtmd_helper_init_opt_default (tts_engine.cpp) and upstream's own tokenize_input_prompts (jllama.cpp); it never calls mtmd_tokenize itself, and that re-typed parameter would have been a legal qualification conversion anyway. mtmd_helper::gen_audio — the entire TTS surface — is untouched.

The server wire contract was checked mechanically, not by header diff. That is the failure class behind the getMetrics() payload break (b10408/b10519) and the repeat_last_n sentinel removal (b10273), where no signature moved and the build stayed green for hundreds of runs. server-schema.cpp's request-field set, its set_hard_limits bounds, and server-task.cpp's response keys in both emit forms are byte-identical across the range: zero added, zero removed, zero re-bounded.

Patches: all 8 still apply. For the first step this was not a free pass — 3 of their 42 target files are in b10731…b10792 (common/arg.cpp + mtmd-cli.cpp for 0001, server-context.cpp for 0002/0003/0010), so the applier had to run for real. 0010 and 0001 were additionally re-checked by hand against the rule in CLAUDE.md, because the fail-loud applier detects "does not apply" but never "upstream already fixed this, drop me": get_res_model_info() still emits an uncast enum for vocab_type, and common_params_parse_main still appears nowhere in arg.h while the #ifdef _WIN32 argv override is still in arg.cpp — upstream has not adopted #26416.

Upstream behaviour changes worth knowing (nothing broke)

1. preserve_reasoning now defaults to enabled (#28174). common_params_parse_ex sets it when the caller did not, where it previously followed the chat template's own default. On a template advertising supports_preserve_reasoning the full history now carries reasoning traces instead of only the last assistant message — better continuity, more prompt tokens. This reaches every entry point that parses argv: NativeServer in both modes and LlamaModel's own parameter parse. --no-reasoning-preserve restores the old behaviour.
2. data: URLs are now accepted for input_audio and input_video, not only images (#27735).
3. A big-endian correctness fix on a platform we ship in the default JAR (#28332): ggml_vec_dot_q5_1_q8_1's s390x path accumulated into an uninitialized v_acc, so Q5_1 inference on IBM Z could return garbage. Note what this says about our own coverage — build-linux-s390x runs the C++ suite under qemu-user, but that suite covers our endian-sensitive layer (tts_wav.hpp, the JSON/token transforms), not ggml's kernels. Exactly the class of bug that job cannot see.
4. A fix for Qwen3-TTS-0.6b (#28231), where an F16 ffn_down overflowed past the 65504 ceiling and turned the residual into NaN — on the TextToSpeech path this project ships.

One upstream build change was worth verifying rather than reading past: #28278 moves LLAMA_VERSION/LLAMA_COMMIT out of target_compile_definitions into a generated src/llama-version.h. That is the one thing in the range that could plausibly have reached us, since NativeLibraryLoadSmokeTest asserts the pin against the linked binary. It does not — jllama.cpp reads llama_build_info() from common/build-info.cpp — and the test passes at b10797.

Chunking: one deviation, stated openly

The second step (b10792 → b10797) is a straight bump by the runbook's rule: 46 KB across 5 commits, inside the 100 KiB threshold, and its review surface is literally empty — zero files under common/, include/, tools/server/, tools/mtmd/, ggml/include/ or the top-level CMakeLists.txt.

The first step (b10731 → b10792) deliberately did not follow the rule. Its full diff is 4.19 MB, but docs/ops/Hexagon.csv alone is 19 792 generated lines and the GPU backends account for most of the rest — both outside the review surface the rule exists to bound (ggml/src/** is on the "safe to skip" list). Across the paths that are actually reviewed the range is 35 files, +758 / −175, reviewable in one pass; chunking would have produced ~44 commits without adding a line of review. The measured figures are recorded in docs/history/llama-cpp-breaking-changes.md so the deviation is auditable rather than silent.

Test plan

• Fresh rm -rf build && cmake -B build -DBUILD_TESTING=ON at both tags — the fail-loud PATCH_COMMAND ran for real; stamps at head c5a5535e (b10792) and head 832fd6f1 (b10797)
• Full native build, zero warnings on the project TUs
• ctest 520/520 at both tags
• mvn clean verify — 1474 Java tests, 0 failures, incl. NativeLibraryLoadSmokeTest, which cross-checks the new LLAMA_CPP_VERSION against the actually-linked getLlamaCppBuildInfo() (clean is required here; the constant is inlined into the test class — see the runbook gotcha)
• .github/verify-bytecode-version.sh --max-major 52 llama/target clean
• CI is green on this branch
• Docs updated — the four pin files, a folded b10731 → b10797 CHANGELOG entry, and four new rows in docs/history/llama-cpp-breaking-changes.md

Model-backed Java tests were not run locally (HuggingFace is blocked in this sandbox); they self-skip here and are covered by CI.

Related issues / PRs

Follows #412. A stale remote branch claude/llama-cpp-b10792 is left over from the first step — the delete push kept failing at the proxy. Nothing tracks it; safe to delete from the GitHub UI.

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]

Review Summary

This is a well-executed llama.cpp version bump from b10731 to b10797 (two steps). No issues found.

✅ Verification Checklist

Version pin consistency — all four required locations updated correctly:

• ✅ llama/CMakeLists.txt: GIT_TAG b10731 → b10797
• ✅ CLAUDE.md: documentation updated (4 references)
• ✅ README.md: badge updated to b10797
• ✅ LlamaCppVersion.java: constant set to "b10797"

Documentation quality:

• ✅ CHANGELOG entry is comprehensive and explains: no project-source changes, header changes are additive/const-widening only, server contract verified mechanically, all 8 patches still apply
• ✅ Breaking-changes history records both range details and chunking rationale
• ✅ All upstream behavior changes documented (preserve_reasoning default, data: URLs, Qwen3-TTS fix, s390x correctness fix)

Verification evidence:

• ✅ Fresh cmake builds at both tags with patch applier verification
• ✅ 520/520 C++ tests passing
• ✅ 1474/1474 Java tests passing
• ✅ Bytecode version verification passing
• ✅ Breaking-changes history includes audit trail of what was mechanically verified

Follows bump procedure per CLAUDE.md:

• ✅ All four pin files updated in lockstep
• ✅ README version snippets updated
• ✅ CHANGELOG and history documentation completed
• ✅ Commits use Conventional Commits format

Analysis

The PR demonstrates careful upgrade methodology:

• No silent assumptions: documentation explicitly states what was checked mechanically (server contract fields, response keys) vs. what a header diff can miss
• Chunking decision well-reasoned: b10731→b10792 deliberately deviated from the 100 KiB rule because GPU backends dominate the diff, not the review surface. Recorded as auditable deviation rather than silence.
• Patch verification: not just "hoped they apply" — ran the fail-loud applier for real and recorded stamps; additionally hand-verified two patches (0010, 0001) against the "upstream already fixed this?" rule

Ready to 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