CMake: declare BUILD_TESTING before any subproject; llama.cpp b11457 → b11462

CHANGELOG.md

@@ -45,11 +45,11 @@ from version 5.0.0 onward. Pre-fork releases (`1.x`–`4.2.0`) were authored by
   core's; `check-natives.py` fails when they differ.
 
 ### Changed
-- **Upgraded the pinned llama.cpp from b11320 to b11457**, in 21 reviewed steps, each ending at a tag.
+- **Upgraded the pinned llama.cpp from b11320 to b11462**, in 23 reviewed steps, each ending at a tag.
   Every carried patch that broke was traced to the one upstream commit that broke it, and the step
   containing that commit ends at the first tag after it: `0007` at #29818 (b11361) and #29895 (b11401), `0014` at #29895, `0008`
   at #29987 (the commit just before b11429), `0015` at #26610 (b11450). Each refresh moved context only,
-  and all nine patches are still needed. The new
+  and every patch is still needed. The new
   upstream features this binding now exposes are listed under *Added*; the build follows upstream's
   CUDA CCCL pin (v3.4.3) and OpenVINO 2026.4.1. Per-step record:
   `docs/history/llama-cpp-breaking-changes.md`.

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: **b11457**
+Current llama.cpp pinned version: **b11462**
 
 ## Natives jars: one directory per backend (`.github/natives.csv`)
 
@@ -214,6 +214,12 @@ To change the CUDA version, update the following places:
    `DeviceTopK` path needs CCCL >= 3.4.3 and falls back to a sort below it, and CUDA 13.4 bundles an
    older 3.4. **Drop both flags once the toolkit is 13.5 or newer** (it bundles CCCL 3.5); follow
    upstream's `release.yml` matrix comment, which says the same.
+   **The pin makes ggml fetch CCCL, and CCCL calls `include(CTest)` unconditionally**, which creates
+   `BUILD_TESTING` as a cache variable defaulting to ON. That is why `llama/CMakeLists.txt` declares
+   `option(BUILD_TESTING ... OFF)` before its first `FetchContent_MakeAvailable()`: declared after,
+   the option was a no-op, both CUDA jobs built `jllama_test`, and its gtest discovery failed on the
+   GPU-less runners (no `libcuda.so.1` / `nvcuda.dll`) -- the first Publish run after the bump
+   (37550451678). Keep the option there.
 5. **`CLAUDE.md`** — the "Current CUDA version" line above.
 
 Available CUDA versions for RHEL8/Manylinux_2_28 can be browsed at:
@@ -704,7 +710,7 @@ needs no extra step here, `build-webui` re-reads the tag and rebuilds the matchi
 ships no UI):
 ```bash
 # needs node/npm + network for the asset build; the embed step is plain cmake -P
-git clone --depth 1 --branch b11457 https://github.com/ggml-org/llama.cpp /tmp/lc
+git clone --depth 1 --branch b11462 https://github.com/ggml-org/llama.cpp /tmp/lc
 ( cd /tmp/lc/tools/ui && npm ci && npm run build )
 mkdir -p webui-generated /tmp/ui-gen
 cmake -DUI_SOURCE_DIR=/tmp/lc/tools/ui -DUI_BINARY_DIR=/tmp/ui-gen \
@@ -744,7 +750,7 @@ 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 b11457`), the
+Because `sccache` is **content-addressed** and llama.cpp is pinned (`GIT_TAG b11462`), 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
@@ -1867,7 +1873,7 @@ 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 b11457`.
+llama.cpp is fetched via CMake FetchContent, pinned to `GIT_TAG b11462`.
 
 **GoogleTest** is a separate `BUILD_TESTING`-only FetchContent (`GIT_TAG v1.18.0`), 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 b11457](https://img.shields.io/badge/llama.cpp-%23b11457-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b11457)  
+[![llama.cpp b11462](https://img.shields.io/badge/llama.cpp-%23b11462-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b11462)  
 [![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

@@ -828,3 +828,7 @@ Used during `llama.cpp` version bumps: when upgrading, scan this file from the r
 | b11450–b11454 | patches + upstream verification | **All nine patches apply unchanged**, verified in order against pristine b11454 with `git apply`. Patch-target files in the range: `src/llama-model.{cpp,h}` (`0012`; the two new architectures are registered away from `load_tensors`' split block and the helper declarations). Drop-checks against pristine b11454 unchanged: `0001` override present and `common_params_parse_main` absent, `0002` assignment unconditional, no `0003` getter, no `0006`-`0008` counterpart, no `0012` guard (`split_sum` still unguarded), no `0014` hook, no `0015` stop function. `tools/server/` unchanged in the range. |
 | b11454–b11457 | Three commits, 9 files, ~122/27 lines. **#30045 adds the PLaMo-3 tokenizer pre-segmentation**, a new `LLAMA_VOCAB_TYPE_PLAMO3 = 8` (`include/llama.h`, appended to the enum, so no existing value moves). `ModelMeta.getVocabType()` returns the number as is (`8` for such a model); the project has no Java enum of vocab types to extend. #28782 uses a per-thread CUDA stream for the buffer-init padding memset; #29955 adds BF16 to CUDA's `XIELU`. **No project source change.** This is the target of the bump: b11457 was the newest llama.cpp release when it was done (b11458 had no release assets). |
 | b11454–b11457 | patches + upstream verification | **All nine patches apply unchanged**, verified in order against pristine b11457 with `git apply`. No patch-target file in the range; nothing under `common/` or `tools/server/` changed. Drop-checks against pristine b11457 -- the final state of the bump: `common_params_parse` still carries the `#ifdef _WIN32` `argv = utf8.ptrs.data()` override and `common_params_parse_main` is absent (`0001` needed); `load_progress_callback` still assigned unconditionally (`0002`); no slot-similarity getter (`0003`); no `llama_server_attach`, `LLAMA_SERVER_WORKER_CMD` or embedded-shutdown counterpart (`0006`-`0008`); no `split_sum == 0` guard (`0012`); no callback hook in `common/log.h` (`0014`); no `stop_server` and a `GGML_ABORT("Failed to connect` still present (`0015`). **Every carried patch is still needed.** Across the whole bump (b11320 → b11457) four patches broke, each at exactly one upstream commit: `0007` at #29818 (b11361) and #29895 (b11401), `0014` at #29895 (b11401), `0008` at #29987 (`4d60b4d08`, before b11429), `0015` at #26610 (b11450); every refresh moved context only. |
+| b11457–b11458 | One commit, 9 files, ~978/733 lines, **#30067 only**: the Hexagon backend's CPY/CONCAT/CONT/DUP overhaul (DMA/HVX for every path, `htp/hex-cpy-dma.h` renamed to `dma-copy.h`) plus its Snapdragon developer guide. No natives jar builds Hexagon, so **nothing in it reaches a shipped library**. At 112 KiB it is over the 100 KiB step threshold, but a single commit cannot be split further. **No project source change.** |
+| b11457–b11458 | patches + upstream verification | **All ten patches apply unchanged** (`0001`-`0016`, now including the Kolibri-1 carry `0016`), verified in order against pristine b11458 with `git apply`. No patch-target file and nothing under `src/`, `common/`, `tools/server/` or `include/` in the range. |
+| b11458–b11462 | Four commits, 11 files, ~119/192 lines, backends only. **SYCL:** #27689 removes the separate flash-attention KV buffers (`fattn-buffers.{cpp,hpp}` deleted; `ggml-sycl/CMakeLists.txt` globs its sources, so nothing to wire) and #29500 adds an IQ3_S multi-column MMVQ kernel. **Vulkan:** #30049 skips the direct host read of uncached host-visible memory on AMD UMA devices (iGPUs), which is write-combined and slow to read, and takes the device-to-host copy path instead. **WebGPU:** #27069 (not built here). **No project source change.** This is the target of the bump: b11462 is the newest llama.cpp release. |
+| b11458–b11462 | patches + upstream verification | **All ten patches apply unchanged**, verified in order against pristine b11462 with `git apply`. No patch-target file and nothing under `src/`, `common/`, `tools/server/` or `include/` in the range, so every drop-check of the b11457 row still holds unchanged; `git grep -i kolibri` finds nothing under `src/` at b11462, so `0016` is still needed too. No `.github/` change upstream either: OpenVINO, CCCL (`v3.4.3`) and the ROCm wheels stay where b11457 left them. |

llama/CMakeLists.txt

@@ -36,6 +36,14 @@ set(LLAMA_CURL OFF)
 
 option(LLAMA_VERBOSE	"llama: verbose output"		OFF)
 
+# Declared here, before any FetchContent_MakeAvailable(), and not next to the tests below: a
+# subproject that calls include(CTest) creates BUILD_TESTING as a cache variable defaulting to ON,
+# after which this option() would be a no-op. CCCL does exactly that, unconditionally, and ggml
+# fetches it whenever GGML_CUDA_CCCL_VERSION is set (our CUDA builds pin v3.4.3) -- the CUDA job
+# then built jllama_test, whose gtest discovery needs libcuda.so.1 and failed the build on the
+# GPU-less runner. -DBUILD_TESTING=ON still enables the tests.
+option(BUILD_TESTING "Build C++ unit tests for jni_helpers / json_helpers / utils" OFF)
+
 #################### json ####################
 
 FetchContent_Declare(
@@ -182,7 +190,7 @@ set(LLAMA_BUILD_APP OFF CACHE BOOL "" FORCE)
 FetchContent_Declare(
 	llama.cpp
 	GIT_REPOSITORY https://github.com/ggerganov/llama.cpp.git
-	GIT_TAG        b11457
+	GIT_TAG        b11462
 	PATCH_COMMAND  ${CMAKE_COMMAND}
 		-DPATCH_DIR=${CMAKE_CURRENT_SOURCE_DIR}/patches
 		-DLLAMA_SRC=<SOURCE_DIR>
@@ -537,8 +545,7 @@ endif()
 
 #################### C++ unit tests ####################
 
-option(BUILD_TESTING "Build C++ unit tests for jni_helpers / json_helpers / utils" OFF)
-
+# BUILD_TESTING is declared at the top of this file, before any subproject (see there).
 if(BUILD_TESTING)
     FetchContent_Declare(
         googletest

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

@@ -9,28 +9,28 @@
  * 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 "b11457"}) that mirrors the
+ * <p>{@link #LLAMA_CPP_VERSION} is a pure-Java string ({@code "b11462"}) 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 "b11457-<commit>"} — call
+ * plus the resolved upstream commit, e.g. {@code "b11462-<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 "b11457"}.
+     * The pinned llama.cpp release tag this library was built against, e.g. {@code "b11462"}.
      *
      * <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 = "b11457";
+    public static final String LLAMA_CPP_VERSION = "b11462";
 
     // Constants holder — not instantiable.
     private LlamaCppVersion() {}