Upgrade llama.cpp from b10423 to b10456

CLAUDE.md

@@ -6,7 +6,7 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
 
 Java bindings for [llama.cpp](https://github.com/ggerganov/llama.cpp) via JNI, providing a high-level API for LLM inference in Java. The Java layer communicates with a native C++ library through JNI.
 
-Current llama.cpp pinned version: **b10423**
+Current llama.cpp pinned version: **b10456**
 
 ## Upgrading CUDA Version
 
@@ -490,7 +490,7 @@ needs no extra step here, `build-webui` re-reads the tag and rebuilds the matchi
 ships no UI):
 ```bash
 # needs node/npm + network; embed.cpp is plain C++17 (no npm)
-git clone --depth 1 --branch b10423 https://github.com/ggml-org/llama.cpp /tmp/lc
+git clone --depth 1 --branch b10456 https://github.com/ggml-org/llama.cpp /tmp/lc
 ( cd /tmp/lc/tools/ui && npm ci && npm run build \
   && ( cd dist && find . -type f -not -path './_gzip/*' \
        | while read -r f; do mkdir -p "_gzip/$(dirname "$f")"; gzip -9 -c "$f" > "_gzip/$f"; done ) \
@@ -530,7 +530,7 @@ cache lives in **Depot Cache** over sccache's **WebDAV** backend:
 - `SCCACHE_WEBDAV_TOKEN: ${{ secrets.DEPOT_TOKEN }}` — a Depot **organization** token, stored
   as the repo secret **`DEPOT_TOKEN`**.
 
-Because `sccache` is **content-addressed** and llama.cpp is pinned (`GIT_TAG b10423`), the
+Because `sccache` is **content-addressed** and llama.cpp is pinned (`GIT_TAG b10456`), the
 ~280 upstream object files are byte-identical every run, so a warm cache recompiles only the
 *changed* files. Depot's cache is **shared across all branches** (unlike GitHub's
 per-branch `actions/cache`), so every branch builds incrementally; a `b<nnnn>` version bump
@@ -1357,7 +1357,7 @@ ctest --test-dir build --output-on-failure -R "ResultsToJson"
 
 #### Upstream source location (in CMake build tree)
 
-llama.cpp is fetched via CMake FetchContent, pinned to `GIT_TAG b10423`.
+llama.cpp is fetched via CMake FetchContent, pinned to `GIT_TAG b10456`.
 
 **GoogleTest** is a separate `BUILD_TESTING`-only FetchContent (`GIT_TAG v1.17.0`), used solely
 by the `jllama_test` C++ unit-test binary — not by the shipped library, and not coupled to the

README.md

@@ -11,7 +11,7 @@
 **Build:**  
 ![Java 8+](https://img.shields.io/badge/Java-8%2B-informational)  
 ![Platform](https://img.shields.io/badge/Platform-Linux%20%7C%20macOS%20%7C%20Windows%20%7C%20Android-lightgrey)  
-[![llama.cpp b10423](https://img.shields.io/badge/llama.cpp-%23b10423-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b10423)  
+[![llama.cpp b10456](https://img.shields.io/badge/llama.cpp-%23b10456-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b10456)  
 [![JPMS](https://img.shields.io/badge/JPMS-modular%20JAR-25A162)](https://openjdk.org/projects/jigsaw/)  
 ![JUnit](https://img.shields.io/badge/tested%20with-JUnit6-25A162)  
 [![JSpecify](https://img.shields.io/badge/JSpecify-1.0.0%20%40NullMarked-25A162)](https://jspecify.dev)  

docs/history/llama-cpp-breaking-changes.md

@@ -610,3 +610,12 @@ Used during `llama.cpp` version bumps: when upgrading, scan this file from the r
 | b10418–b10419 | `ggml/src/ggml-openvino/**` (**Qwen3.5 support, memory optimization, new ops, #26952**) | **No project-source change.** A single commit entirely inside the OpenVINO backend; no shared header touched, so only the `openvino-{linux,windows}-x86-64` classifier's compiled sources change. |
 | b10419–b10423 | `common/{common.h,common.cpp}` + `tools/completion/completion.cpp` (**shared CPU-parameter handling, #27026 — new `common_threadpools` RAII helper, `ggml_threadpool_params_from_cpu_params()` moved next to it**), `tools/ui/**` | **No project-source change (final step).** `ggml_threadpool_params_from_cpu_params()` moved within `common.h` but kept its signature; confirmed via grep that no project source mentions `threadpool` at all. |
 | b10360–b10423 | upstream verification (sandbox, final target) | All **6** patches re-verified against a clean b10423 checkout: sequential `git apply` (filename order, mirroring the `PATCH_COMMAND` applier) succeeded at **every** intermediate tag of the walk, and the fail-loud `PATCH_COMMAND` ran clean on a fresh `cmake -B build` at the target — no patch needed refreshing across the whole range. **Full local verification (mandatory for the final target):** fresh configure + full `cmake --build` (`libjllama.so` + `jllama_test` compile and link, `-O3`, no undefined references) + `ctest` **485/485 passing**. Walk of 10 steps (b10360→b10362→b10369→b10375→b10380→b10398→b10405→b10408→b10418→b10419→b10423) on branch `claude/update-b10423-ciih0h`. Four steps exceeded the 100 KiB chunking threshold unavoidably — upstream has **no** intermediate tags there (b10363–b10368, b10381–b10397, b10406–b10407 and b10420–b10422 do not exist), so the tag list itself, not the threshold, bounded the step size. |
+| b10423–b10430 | `tools/server/server-queue.{h,cpp}` + `tools/server/server-context.cpp` (**`/metrics` and `/slots` served during `llama_decode()`, #27041 — new `server_queue::yield_to_queue()` worker thread, and `on_new_task`'s callback type changed from `std::function<void(server_task &&)>` to `std::function<bool(server_task &&, bool)>`**), `src/llama.cpp` (**virtual iGPU devices, #26953**), `ggml/src/ggml-sycl/**`, `ggml/src/ggml-cpu/ggml-cpu.c` (**single thread on WASI, #25686**) | **No project-source change.** `on_new_task`'s signature change is the only breaking edit in a header the project compiles, and it is inert here: `server-context.cpp` (compiled in from upstream) is the sole registrant — confirmed via grep that no project source mentions `on_new_task`, `server_queue`, `queue_tasks` or `yield_to_queue`. First step of the b10423→b10456 walk. |
+| b10430–b10436 | `common/arg.cpp` (**additive `--reasoning-effort` / `LLAMA_ARG_REASONING_EFFORT`, plus a negative-value guard on `--spec-draft-n-max`**), `common/chat.cpp` + `common/jinja/{caps,runtime,string}.{h,cpp}` (**`reasoning_effort` passed to the chat template; quadratic-cost fix in `gather_string_parts`, #27034**), `tools/server/server-common.cpp` (**`oaicompat_chat_params_parse` now forwards a non-`"none"` `reasoning_effort` into `chat_template_kwargs` instead of ignoring it**), `tools/mtmd/{clip.cpp,clip-impl.h,mtmd-image.cpp,mtmd-helper-common.h}` (**various fixes, #27071 — `decode_embd_batch` index arithmetic widened to `size_t`**), `ggml/include/ggml.h` (**`ggml_ssm_scan()` gained a trailing `int64_t K` parameter, #26623**) | **No project-source change.** `ggml_ssm_scan` is a model-graph builder called only inside llama.cpp's own translation units (the project includes `ggml.h` for types, never this function); the `mtmd-helper-common.h` edits are `size_t` casts with no signature change; the `arg.cpp`/`chat.cpp`/`server-common.cpp` edits are additive or internal to upstream-compiled files. The `reasoning_effort` forwarding is behavioral, not API — worth noting because `test_utils.cpp` covers `oaicompat_chat_params_parse` (suite re-run green at the target). |
+| b10436–b10441 | `tools/mtmd/{clip.cpp,clip-impl.h,mtmd-image.{h,cpp},mtmd.cpp,models/granite4-vision.cpp,models/models.h}` (**Granite4-Vision image sequence assembly fix, #26653 — internal `mtmd-image.h`/`models.h` helper signatures changed**), `src/{llama-arch,llama-hparams,llama-model,llama-context,llama-model-loader}.*` + `src/models/minimax-01.cpp` (**new MiniMaxText01/MiniMaxM1 architectures, #27018**), `ggml/src/ggml-vulkan/ggml-vulkan.cpp` (**Intel Xe coopmat1 gating fix**), `conversion/`, `docs/`, `scripts/` | **No project-source change.** The mtmd edits are confined to the library's *internal* headers (`clip-impl.h`, `mtmd-image.h`, `models/models.h`) — the two public headers the project includes, `mtmd.h` and `mtmd-helper.h`, are byte-identical across the whole b10423→b10456 range. New model architectures are entirely inside upstream-compiled `src/` translation units. |
+| b10441–b10447 | `tools/server/server-queue.{h,cpp}` + `tools/server/server-context.cpp` (**`yield_to_queue` thread model re-designed, #27133 — follow-up to #27041 from the b10423→b10430 step**), `vendor/cpp-httplib/{httplib.h,httplib.cpp,CMakeLists.txt}` (**cpp-httplib 0.53.0 → 0.53.1, #27103: `CPPHTTPLIB_REGEX_ROUTE_PATH_MAX_LENGTH` stack-exhaustion guard + IPv6-literal URL parsing fix; BoringSSL pin 0.20260803.0 → 0.20260813.0, #27099**), `common/preset.cpp` (**`--models-dir` loads MTP assistant models, #24431**), `common/imatrix-loader.cpp` (**GGUF array-type check, #27075**), `tools/mtmd/clip.cpp`, `ggml/src/ggml-vulkan/**` | **No project-source change.** Same `server_queue` reasoning as the b10423→b10430 row (no project source touches that API). The vendored `httplib.cpp` **is** compiled into `jllama`, but the point release is source-compatible — the two behavioral fixes are internal to `parse_url`/regex-route dispatch, and the new macro has a default. The BoringSSL pin is inert here: `LLAMA_BUILD_BORINGSSL` is off and `CPPHTTPLIB_OPENSSL_SUPPORT` is never defined (plain-HTTP by design). |
+| b10447–b10448 | `common/chat.cpp` (**new `common_chat_params_init_kimi_k3()` + a dispatch arm in `common_chat_try_specialized_template()`**), `src/{llama-arch,llama-graph,llama-hparams,llama-model,llama-model-saver,llama-context}.*` + `src/models/kimi-k3.cpp` + `models/templates/Kimi-K3.jinja`, `tests/test-chat.cpp` (**Kimi-K3 text model, #26185**) | **No project-source change.** A single-commit model addition: purely additive inside upstream-compiled translation units, and `common/chat.h` is unchanged (no new enum value or signature the project would have to switch on). Step forced above the 100 KiB threshold (114 KiB) — upstream has no tag between b10447 and b10448, so the tag list, not the threshold, bounded it. |
+| b10448–b10454 | `common/chat.cpp` + `common/jinja/caps.cpp` (**`supports_string_content` / `supports_typed_content` handling refactored, #27130**), `src/llama-adapter.cpp` (**LoRA tensor data bounds-checked against the file, #27056**), `src/models/**` (**some `ggml_concat` removed, #27176**), `tools/ui/**` (**API-key masking, MCP `structuredContent`**) | **No project-source change.** The chat refactor stays inside `chat.cpp`/`caps.cpp` — `common/chat.h` and `common/jinja/caps.h` are unchanged, so nothing the project compiles against moved. The LoRA hardening is a defensive check inside upstream's adapter loader; `tools/ui` is auto-followed (CI rebuilds the Svelte WebUI from the pinned `GIT_TAG`). |
+| b10454–b10455 | `ggml/src/ggml-sycl/{ggml-sycl.cpp,opt-step.cpp,opt-step.hpp}` (**SYCL `OPT_STEP_ADAMW` / `OPT_STEP_SGD`, #25268**), `docs/ops/SYCL.csv` + `docs/ops.md` + `examples/sycl/update-ops-doc.sh` (**regenerated op-support table**) | **No project-source change.** A single commit inside the SYCL backend, so only the `sycl-{fp16,fp32}-linux-x86-64` / `sycl-windows-x86-64` classifiers' compiled sources change. Step forced above the 100 KiB threshold (3.2 MiB) — but ~3.2 MiB of that is the regenerated `docs/ops/SYCL.csv` alone (20 646 lines rewritten); the reviewable code is ~150 lines, and upstream has no tag between b10454 and b10455 anyway. |
+| b10455–b10456 | `ggml/src/ggml-sycl/cpy.cpp` (**thread/block count fixed in the quantized cpy kernel launches, #27160**) | **No project-source change (final step).** A single one-commit fix inside the SYCL backend; no shared header touched, so only the `sycl-{fp16,fp32}-linux-x86-64` / `sycl-windows-x86-64` classifiers' compiled sources change. |
+| b10423–b10456 | upstream verification (sandbox, final target) | All **6** patches re-verified against a clean b10456 checkout: sequential `git apply` (filename order, mirroring the `PATCH_COMMAND` applier) succeeded at **every** intermediate tag of the walk, and the fail-loud `PATCH_COMMAND` ran clean on a fresh `cmake -B build` at the target — no patch needed refreshing across the whole range. **Full local verification (mandatory for the final target):** fresh configure (resolved `ggml commit f275595dd` = b10456 HEAD; all six patch markers confirmed present in the fetched tree) + full `cmake --build` (`libjllama.so` + `jllama_test` compile and link, `-O3`, no undefined references) + `ctest` **485/485 passing**. **API surface:** of the 22 upstream headers project source `#include`s directly, only three changed across the whole range — `tools/server/server-queue.h` (breaking: `on_new_task`'s callback type became `std::function<bool(server_task &&, bool)>`, plus the new `yield_to_queue()`; inert here, upstream's own `server-context.cpp` is the sole registrant), `ggml/include/ggml.h` (`ggml_ssm_scan()` gained a trailing `int64_t K`; called only inside llama.cpp's own TUs) and `tools/mtmd/mtmd-helper-common.h` (`size_t` widening, no signature change). `mtmd.h`, `mtmd-helper.h`, `common.h`, `chat.h`, `llama.h`, `arg.h` and every `tools/server/*.h` the project includes are byte-identical b10423→b10456; no files were added or removed under `tools/server`/`tools/mtmd`/`vendor` and no upstream `CMakeLists.txt` changed, so the b10154 `server-mcp.cpp` missing-`target_sources` failure class does not recur. Walk of 8 steps (b10423→b10430→b10436→b10441→b10447→b10448→b10454→b10455→b10456) on branch `claude/update-b10456-8a7j58`. Two steps exceeded the 100 KiB chunking threshold unavoidably — upstream has **no** intermediate tags there (b10449–b10453 and b10456 aside, nothing sits between b10447/b10448 or b10454/b10455) — and the 3.2 MiB b10454→b10455 figure is almost entirely the regenerated `docs/ops/SYCL.csv` (20 646 lines rewritten), leaving ~150 lines of reviewable code. |

llama/CMakeLists.txt

@@ -173,7 +173,7 @@ set(LLAMA_BUILD_APP OFF CACHE BOOL "" FORCE)
 FetchContent_Declare(
 	llama.cpp
 	GIT_REPOSITORY https://github.com/ggerganov/llama.cpp.git
-	GIT_TAG        b10423
+	GIT_TAG        b10456
 	PATCH_COMMAND  ${CMAKE_COMMAND}
 		-DPATCH_DIR=${CMAKE_CURRENT_SOURCE_DIR}/patches
 		-DLLAMA_SRC=<SOURCE_DIR>

llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java

@@ -10,28 +10,28 @@
  * library was compiled against, exposed as a compile-time constant so callers can render a badge or
  * emit a startup log line without loading the native library.
  *
- * <p>{@link #LLAMA_CPP_VERSION} is a pure-Java string ({@code "b10423"}) that mirrors the
+ * <p>{@link #LLAMA_CPP_VERSION} is a pure-Java string ({@code "b10456"}) that mirrors the
  * {@code GIT_TAG} in {@code llama/CMakeLists.txt}. It is available even when {@code libjllama} is
  * absent (pure-Java checkout, before {@code System.load}), which is what makes it suitable for a
  * lightweight version badge in Android or other UIs.</p>
  *
  * <p>For the <em>authoritative</em> value that is baked into the native binary — the build number
- * plus the resolved upstream commit, e.g. {@code "b10423-<commit>"} — call
+ * plus the resolved upstream commit, e.g. {@code "b10456-<commit>"} — call
  * {@link net.ladenthin.llama.LlamaModel#getLlamaCppBuildInfo()} instead; that reads llama.cpp's own
  * {@code build-info} through JNI and therefore cannot drift from the compiled library (but requires
  * the native library to be loaded).</p>
  */
 public final class LlamaCppVersion {
 
     /**
-     * The pinned llama.cpp release tag this library was built against, e.g. {@code "b10423"}.
+     * The pinned llama.cpp release tag this library was built against, e.g. {@code "b10456"}.
      *
      * <p>Kept in lockstep with {@code GIT_TAG} in {@code llama/CMakeLists.txt} — see the
      * "Upgrading/Downgrading llama.cpp Version" checklist in {@code CLAUDE.md}. This is the
      * compile-time pin; use {@link net.ladenthin.llama.LlamaModel#getLlamaCppBuildInfo()} for the
      * value actually linked into the native binary.</p>
      */
-    public static final String LLAMA_CPP_VERSION = "b10423";
+    public static final String LLAMA_CPP_VERSION = "b10456";
 
     // Constants holder — not instantiable.
     private LlamaCppVersion() {}