Upgrade llama.cpp from b10034 to b10056

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: **b10034**
+Current llama.cpp pinned version: **b10056**
 
 ## Upgrading CUDA Version
 
@@ -421,7 +421,7 @@ 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 b10034 https://github.com/ggml-org/llama.cpp /tmp/lc
+git clone --depth 1 --branch b10056 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 ) \
@@ -461,7 +461,7 @@ 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 b10034`), the
+Because `sccache` is **content-addressed** and llama.cpp is pinned (`GIT_TAG b10056`), 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
@@ -1254,7 +1254,7 @@ 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 b10034`.
+llama.cpp is fetched via CMake FetchContent, pinned to `GIT_TAG b10056`.
 
 **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

@@ -7,7 +7,7 @@
 **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 b10034](https://img.shields.io/badge/llama.cpp-%23b10034-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b10034)  
+[![llama.cpp b10056](https://img.shields.io/badge/llama.cpp-%23b10056-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b10056)  
 [![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

@@ -497,3 +497,5 @@ Used during `llama.cpp` version bumps: when upgrading, scan this file from the r
 | b9990–b10015 | upstream verification (sandbox) | **Six** patches re-verified against b10015 via a fresh `cmake -B build` (fail-loud `PATCH_COMMAND`): `0001`–`0003`, `0006`, `0008` apply **unchanged**; `0007` was **refreshed** for upstream #25655's CORS restructure (route bindings extracted to the shared helper, per-caller warnings kept in `llama_server()` — see the row above) and now applies clean. No OuteTTS generator anchor touched (`tools/tts/tts.cpp` unchanged — the generator extracted `tts.cpp @ b10015`). Full local verification: fresh configure (fail-loud patch apply + TTS extraction succeeded, `ggml commit 12127defd`) + full `cmake --build` (`jllama` + `jllama_test` link cleanly, confirming the refreshed `0007` `server.cpp` restructure compiles) + `ctest` **485/485 passing**; per-platform confirmation by the CI pipeline. |
 | b10015–b10034 | `ggml/src/{ggml-cpu/**,ggml-cuda/{concat.cu,ggml-cuda.cu,lightning-indexer.{cu,cuh},mmq.{cu,cuh},quantize.cu},ggml-metal/**,ggml-opencl/**,ggml-sycl/{fattn-onednn.{cpp,hpp},dequantize.hpp,element_wise.*,getrows.cpp,conv2d-dw.cpp,ggml-sycl.cpp,common.hpp,fattn.cpp},ggml-vulkan/**}` + `src/models/deepseek4.cpp` + `common/{arg.cpp,speculative.cpp}` + `tools/{server/server-tools.cpp,tokenize/tokenize.cpp}` + `tests/test-backend-ops.cpp` | **Additive/tuning-only, no public-API surface (46 files, ~100 KiB full / ~1.6k insertions excluding the auto-followed `tools/ui` WebUI, 19 commits).** All **eight** priority-8 headers (`common/{common.h,chat.h,speculative.h,arg.h,download.h}`, `tools/mtmd/mtmd.h`, `include/{llama.h,llama-cpp.h}`) are **byte-identical**. The only non-backend, non-tool changes are trivial: `common/arg.cpp` reflows the `--stdin` tokenize help string (one line, in `common_params_parser_init`, **outside** patch `0001`'s `common_params_parse` region); `common/speculative.cpp` nulls `batch.logits` on the draft-decode path (internal spec-decode fix, upstream-compiled TU); `tools/server/server-tools.cpp` tweaks the `read_file` tool description/line-prefix text (not a patch target); `tools/tokenize/tokenize.cpp` (a standalone tool, not compiled/shipped here) aligns with the `--stdin` change. `src/models/deepseek4.cpp` is an internal DeepSeek-V4 compute-graph refactor inside the upstream-compiled `llama` lib. The bulk (new CUDA lightning-indexer + concat/mmq/quantize tuning, Metal FA/ops additions, SYCL oneDNN flash-attention + dequant/getrows, Vulkan out_prod, OpenCL Adreno kernel gating) is entirely inside ggml backend TUs — this project's GPU classifiers are build-only, no source touched. No `tools/server/*` patch-target region or OuteTTS generator anchor (`tools/tts/tts.cpp` unchanged) touched; all **six** patches (`0001`–`0003`, `0006`–`0008`) apply unchanged. b10034 is the topmost release at bump time. |
 | b10015–b10034 | upstream verification (sandbox) | All **six** patches (`0001`–`0003`, `0006`–`0008`) re-verified against b10034: applied in filename order onto a clean b10034 checkout, all clean (`0001`'s `common/arg.cpp` region untouched by the `--stdin` help reflow; `0006`/`0007` anchor onto `0001`'s post-flip `tools/server/server.cpp`, which upstream left unchanged in this range). No OuteTTS generator anchor touched (`tools/tts/tts.cpp` unchanged — the generator extracts `tts.cpp @ b10034`). Per-platform build + `ctest` confirmation by the CI pipeline. |
+| b10034–b10056 | `ggml/{include/{ggml.h,ggml-cpu.h,ggml-rpc.h},src/{ggml-cpu/**,ggml-cuda/{dsv4-hc.{cu,cuh},ggml-cuda.cu,mm{f,id,q}.*,mmq-load-tiles.cuh,quantize.*,vecdotq.cuh,common.cuh},ggml-hexagon/**,ggml-opencl/**,ggml-vulkan/ggml-vulkan.cpp,ggml.c,ggml-backend-meta.cpp}}` + `src/{llama-context.cpp,llama-cparams.h,llama-graph.h,llama-model.cpp,llama-quant.cpp,models/{deepseek4.cpp,models.h}}` + `tools/server/{server-common.h,server-context.cpp,server-http.cpp}` + `tests/**` | **Additive/tuning-only, no public-API surface (89 files, ~397 KiB / ~4.8k insertions, 22 commits).** All **eight** priority-8 headers (`common/{common.h,chat.h,speculative.h,arg.h,download.h}`, `tools/mtmd/mtmd.h`, `include/{llama.h,llama-cpp.h}`) are **byte-identical**. `ggml/include/ggml.h` only **adds** three DeepSeek-V4 hyper-connection op enums (`GGML_OP_DSV4_HC_{COMB,PRE,POST}`) and `ggml-cpu.h` adds `ggml_cpu_has_sme2()` — additive, resolved inside upstream-compiled TUs. The three touched `tools/server/*` files are internal: `server-common.h` adds an inline `server_tokens::has_media()`; `server-context.cpp` re-gates SLOT_SAVE/RESTORE/ERASE on slot media content (`check_no_mtmd` → `check_slot_no_media`) — a different region than patches `0002`/`0003`/`0007`; `server-http.cpp` guards an empty CORS `Origin` header (a different region than `0007`'s route registration). The bulk is ggml backend kernel work (CUDA DSV4-HC + MoE scatter-quant + Q1_0 unpack, Arm SME2 + DSV4-HC CPU ops, a large Hexagon DSP rework, OpenCL Adreno kernels, Vulkan) inside upstream-compiled TUs — this project's GPU classifiers are build-only, no source touched. `src/{llama-*,models/deepseek4.cpp}` is internal DeepSeek-V4 compute-graph work inside the upstream-compiled `llama` lib. No `tools/tts/tts.cpp` change (OuteTTS generator anchors hold); all **six** patches (`0001`–`0003`, `0006`–`0008`) apply unchanged. b10056 is the topmost release at bump time. |
+| b10034–b10056 | upstream verification (sandbox) | Verified via a fresh `cmake -B build -DBUILD_TESTING=ON` (fail-loud `PATCH_COMMAND` + TTS extraction) on Linux x86_64: all **six** patches (`0001`–`0003`, `0006`–`0008`) apply clean against b10056 (ggml commit `b85833e93`) — `0001`'s `common/arg.{cpp,h}` region untouched upstream; `0002`/`0003`/`0007` anchor onto `tools/server/server-context.{cpp,h}` regions the slot-save/CORS edits do not overlap; `0006`/`0008` targets (`server.cpp`, `server-models.cpp`) unchanged. OuteTTS generator extracted `tts.cpp @ b10056` (all anchors held). Full native build (`cmake --build`, `libjllama.so` + `jllama_test` linked, incl. the changed server TUs) + C++ `ctest` **485/485 green**; per-platform model-backed confirmation by the CI pipeline. |

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        b10034
+	GIT_TAG        b10056
 	PATCH_COMMAND  ${CMAKE_COMMAND}
 		-DPATCH_DIR=${CMAKE_CURRENT_SOURCE_DIR}/patches
 		-DLLAMA_SRC=<SOURCE_DIR>
@@ -196,7 +196,7 @@ execute_process(
     COMMAND ${CMAKE_COMMAND}
         -DTTS_SRC=${llama.cpp_SOURCE_DIR}/tools/tts/tts.cpp
         -DOUT_CPP=${JLLAMA_TTS_GEN_CPP}
-        -DLLAMA_TAG=b10034
+        -DLLAMA_TAG=b10056
         -P ${CMAKE_CURRENT_SOURCE_DIR}/cmake/generate-tts-upstream.cmake
     RESULT_VARIABLE JLLAMA_TTS_GEN_RESULT
 )

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 "b10034"}) that mirrors the
+ * <p>{@link #LLAMA_CPP_VERSION} is a pure-Java string ({@code "b10056"}) 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 "b10034-505b1ed15"} — call
+ * plus the resolved upstream commit, e.g. {@code "b10056-b85833e93"} — 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 "b10034"}.
+     * The pinned llama.cpp release tag this library was built against, e.g. {@code "b10056"}.
      *
      * <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 = "b10034";
+    public static final String LLAMA_CPP_VERSION = "b10056";
 
     // Constants holder — not instantiable.
     private LlamaCppVersion() {}