[bernardladenthin]

Summary

• Bug-hunt and pre-release fixes (every finding has its id, reproduction and red/green test in BUGHUNT-FOUND.md, the user-visible ones are listed in CHANGELOG.md): values that were ignored or misread are refused when set (TrainingParameters bounds, setPredict below -1, adapter paths with a colon, empty JSON fragments, NaN biases); the repeatable adapter options add up instead of replacing; one request is one prompt and one completion on every native path (LlamaModel, the C API, llama-ffm, every OpenAiCompatServer route); slot files are names inside --slot-save-path as llama.cpp's slot route resolves them; LlamaLoader.initialize() is idempotent; a malformed stream envelope is an error, not the end of the stream (the translation lives in StreamEnvelope so its test runs without the natives, F263); infill fields and messages on the raw completion paths take effect (template stops, reasoning budget, preserved tokens, upstream's multimodal prompt object) or are refused; both servers default to llama.cpp's port 9931; chatBatch / completeBatch leave the common ForkJoinPool; ServerMetrics.getCumulativeUsage() counts cached prompt tokens; a LoRA change no longer answers from the old slot cache (new patches/0021); a vocab-only model refuses server operations instead of crashing; the classes jar carries Main-Class: ServerLauncher again (lost with the fat-jar assembly, so every documented jbang --deps <natives jar> net.ladenthin:llama:<v> line had no entry point; smoke-natives-jars.sh reads it back from the built jar, F269); jllama_rpc_server_stop() is for the serve calls in progress when it is made (every serve call has an id, the stop waits for the serves up to its target and issues ggml's stop under the registration mutex, so a serve started while the stop still waits is not ended too; C API and llama-ffm only, the Java RpcServer was not affected; red/green over a new stop seam, F271); eleven tests compared two greedy generations across a prompt-cache hit (the second request re-evaluates one token over the cached prefix, upstream [TAG_PROMPT_LOGITS], whose logits are not bit-identical to the batched prompt evaluation; measured on the random test model over 4000 prompts: the first token's logprob differs in 3999 first-vs-second pairs and the text in 4, while the second and the third cached request agree in all 4000; the three that the MSVC job's failure named and eight more found by a sweep of every test tree now compare the second with the third cached request and assert the reuse through the timings, or run both requests without the cache and prove that through the timings; CLAUDE.md records the rule and names the helpers, F272).
• CI: the Windows sign job runs on every run instead of being skipped (ported as a fresh change from the unmerged branch claude/jolly-brahmagupta-0ydzby, d10f211). A skipped job skips every job behind it in the dependency chain past any if: in between: since the job existed (2026-10-15) every run without the sign_windows switch ended "success" with package green and the smoke sets, the report and both publish jobs skipped (run 38870889665 attempt 2 on main, and every run since) -- no smoke set and no publish job has run since. Now the job always runs, its steps are gated on the switch, the environment is maven-central only when signing and the unprotected unsigned otherwise, package is back on the default condition, and check-natives.py refuses a job-level if: on the sign job (unit test, proven red with the old condition put back).
• OpenAiCompatServer parity with upstream and the protocols: the token-count routes (/v1/messages/count_tokens, /v1/responses/input_tokens, /v1/chat/completions/input_tokens), logit_bias / n_probs / response_fields forwarded, Ollama /api/generate with Ollama's template semantics, format and images, Anthropic's named tool_choice forced and id_slot / extra_params passed through, TEI-style rerank requests, a failure after a stream started reported as the route's error event, wrong-typed number fields and unknown input items answered 400 instead of silently replaced. New builder methods: withLogitBiasJson, withResponseFields, the reasoning-budget tags, enableContextShift, RouterModel.getAliases() / Status.INCOMPLETE, OpenAiServerConfig.totalSlots.
• llama.cpp b11639 → b11815 in eight reviewable steps (b11760, b11761, b11799, b11800, b11803, b11809, b11814, b11815; docs/history/llama-cpp-breaking-changes.md has a row per step): patch 0016 (Kolibri-1) dropped for upstream's own implementation (#30673, community GGUFs need reconversion, documented in the README), patch 0019 dropped (this project's ggml-cpu : select the NEON fp32->fp16 conversion ggml-org/llama.cpp#30505 merged at b11761), patch 0007 refreshed for POST /tts and the /v1/realtime WebSocket, server-connect.cpp compiled in, tts_engine.cpp moved to the gen_audio init/valid API, test_kolibri1.cpp rewritten against upstream's converter format; the b11803 step brings Flint, SenseNova U1 (with its vision projector) and a jinja namespace fix; the last three steps (2026-10-22) bring LoRA on prism.hadamard folded weights, an OpenCL ARGMAX kernel the three OpenCL module jars embed, the tokenizer.ggml.clean_spaces GGUF key with changed detokenizer defaults for ten pre-tokenizers, the Pixtral pre-tokenizer moved to Tekken, KleidiAI v1.32.0 (not built here) and a SYCL CONV_2D layout fix -- no source change here in any of them, and all 12 patches apply as they are (no patch-target file changed in b11803..b11815, the drop-checks re-read at b11815).
• README audited against the code (one reader per section, every finding adversarially verified, 63 confirmed and applied; the CHANGELOG "Changed" entry lists them): no 32-bit Windows natives, the snapshot channel is a manual Publish run, OpenVINO opt-in wherever the additive rule is described, the static library jars, the release assets, the overridden build options, GPU backends as module builds (-DLLAMA_CURL is gone), the loader's search order and the four Windows extras, handleTokenize / handleDetokenize arguments, streaming /v1/completions, OpenAiServerConfig.builder() and x-api-key, the router statuses, RPC in the library jars, the LangChain4j adapter's mapped features, the agent's prompt and status line, chatWithTools' round cap, the infill prompt position, video input shapes, the reactive cancellation semantics, the Kotlin wrappers, the Akka snippet, the logging verbosity split, the Android ABIs, emulator coverage and from-source recipe, the Windows import table, the PIT scope, and three test properties added to the reference table.
• CI: both SYCL natives jobs install upstream's toolchain (Intel Deep Learning Essentials 2026.1.4 + oneDNN 2026.0.2 offline installers + Level Zero, pinned by URL as llama.cpp's ubuntu-24-sycl / windows-sycl jobs do) instead of the unpinned apt repository and the 2025.3 bundle; the shipped SYCL modules now import ze_loader and the Windows one needs the oneAPI 2026 runtime, both named in the README's classifier table. The OpenSSL half of upstream's move does not apply: the SYCL jars are module-only builds with no cpp-httplib.

Test plan

• Affected unit / integration tests pass locally: the C++ suite built from a fresh tree and run at every bump step (b11799, b11800, b11803, b11809, b11814, b11815: 686/686 at the last three, the rewritten cache-path tests included), the Java pin smoke (NativeLibraryLoadSmokeTest, 7/7 against the b11815 library), the Java classes touched by each change (AnthropicApiSupportTest, RouterModelTest, OpenAiCompatServerHttpTest, StreamEnvelopeTest, the mapper and translator tests, the loader tests, the ten rewritten cache-path Java tests against the random tiny model, ...), the buildcheck unit tests (145), spotbugs:check, spotless:check, check-patches.py, check-natives.py (red with the sign job's old if: put back, green without), check-shared-files.py, check-release-gate.py, check-run-scripts.py and the PyYAML parse of the workflows; the classes jar's manifest read back after mvn package; the full model-backed suites and the native matrix are left to CI on purpose.
• CI is green on this branch -- the run on d10f211 is the first since 2026-10-16 in which the smoke sets, the report and the publish-job gates can run at all (the skipped sign job had silently skipped everything behind package), so it is also the first exercise of the Main-Class smoke (F269), the natives-jar smokes and the RPC smoke on the published jars at b11815. Run 39702748614 (head 272c8fc) was green in 69 of 72 jobs: the two verify-signing-key jobs are red by design on a branch, and the MSVC Windows Java job failed in ChatScenarioTest#testChatCompleteTextMatchesChatCompleteContent (F272, fixed in 85ebbfd and f64a42e). The run on the b11815 head is also the proof for the natives builds at b11815, for the SYCL installer URLs (unreachable from the sandbox, mirrored verbatim from upstream's release.yml), for the new manifest check in the package job and for the model-backed forms of the rewritten tests; the two grammar tests of LlamaModelTest (F270) were green in 39702748614.
• Docs / CHANGELOG updated: CHANGELOG.md, README.md, CLAUDE.md, TODO.md, docs/patches.md (rendered), docs/history/llama-cpp-breaking-changes.md, docs/history/dropped-llama-patches.md, docs/upgrade/llama-cpp-version-bump.md, BUGHUNT-FOUND.md / BUGHUNT-CHECKED.md.

Related issues / PRs

• Refs ggml-cpu : select the NEON fp32->fp16 conversion ggml-org/llama.cpp#30505 (merged at b11761, patch 0019 dropped), model : add Kolibri-1 (Aleph-Alpha) support ggml-org/llama.cpp#30673 (Kolibri-1 upstream, patch 0016 dropped), Misc. bug: Windows: common_params_parse discards the caller's argv ggml-org/llama.cpp#26416 (patch 0001 stays downstream until answered).
• Supersedes the unmerged branch claude/jolly-brahmagupta-0ydzby (its two commits are ported as d10f211; the branch can be deleted).
• Sibling PRs of the same branch: BitcoinAddressFinder, streambuffer (merged), srcmorph and workspace (the bug-hunt loop and its learnings).

Checklist

• I have read CONTRIBUTING.md and CODE_OF_CONDUCT.md
• My commits follow Conventional Commits -- they are imperative one-line subjects in the repository's existing style (Refuse ... (F266), Upgrade llama.cpp from b11814 to b11815), not type(scope): prefixed.
• No security-sensitive changes beyond the hardening described above (slot names with a directory part or .. are refused, F251), which needs no private disclosure.

---

Generated by Claude Code

[sonarqubecloud]

Quality Gate failed

Failed conditions
E Reliability Rating on New Code (required ≥ A)

See analysis details on SonarQube Cloud

Catch issues before they fail your Quality Gate with our IDE extension SonarQube for IDE