From da7fdd54377e395b1deb17517a57f0b915524535 Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Thu, 8 Oct 2026 18:14:40 +0000
Subject: [PATCH 1/8] Docs: the CPU variants ship with 5.2.0, not 5.3.0

The reactor is at 5.2.0-SNAPSHOT (last release v5.1.0), so the first release
with the Linux CPU-variant builds is 5.2.0; CLAUDE.md and TODO.md named 5.3.0.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AytmJF9faEiQEVt6eetQS2
---
 CLAUDE.md | 2 +-
 TODO.md   | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/CLAUDE.md b/CLAUDE.md
index f8a424aed..0132e6f59 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -77,7 +77,7 @@ step, not by Maven.
 
 ## CPU variants (Linux x86-64 and aarch64): `JLLAMA_CPU_VARIANTS`
 
-The two Linux CPU jars, `cpu-linux-x86-64` and `cpu-linux-aarch64`, are built since 5.3.0 the way
+The two Linux CPU jars, `cpu-linux-x86-64` and `cpu-linux-aarch64`, are built since 5.2.0 the way
 upstream builds its own release binaries: `GGML_BACKEND_DL` + `GGML_CPU_ALL_VARIANTS`, one small
 `ggml-cpu-<level>` module per instruction-set level, of which `ggml_backend_load_best` loads the best
 at run time (it scores each module's own feature check against the running CPU). Before, each jar
diff --git a/TODO.md b/TODO.md
index 4436089ad..137444578 100644
--- a/TODO.md
+++ b/TODO.md
@@ -19,7 +19,7 @@ so everything below is genuinely still open.
 
 ### CPU variants (`JLLAMA_CPU_VARIANTS`) -- Windows, GPU modules, loader follow-ups
 
-Linux x86-64 and aarch64 ship the variant build since 5.3.0 (CLAUDE.md "CPU variants"). The rest,
+Linux x86-64 and aarch64 ship the variant build since 5.2.0 (CLAUDE.md "CPU variants"). The rest,
 with what a measurement on a Windows 11 machine (Ryzen 7 5800H, RTX 3070, JDK 21, upstream b11476
 binaries; 2026-10-08) established:
 

From c7b8cce6d35b4578481d4d7c1e309272dc553b9d Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Thu, 8 Oct 2026 19:09:44 +0000
Subject: [PATCH 2/8] Upgrade llama.cpp from b11476 to b11480

#29876 hands the chat parser the generated tokens next to the text:
common_chat_input replaces the std::string of common_chat_parse(), of
common_chat_parser_params::generation_prompt and of the server result
types' content. Its string constructor is explicit, so the C++ tests
wrap their literals; the wire format is unchanged.

#29887 adds a GPU cache for the MoE experts a --cpu-moe setup keeps in
host memory (--moe-cache-mib, EXPERIMENTAL). Exposed as
ModelParameters.setMoeCacheMib(int) with ModelOption.MOE_CACHE_MIB,
negative values rejected like upstream's parser does.

All ten patches apply unchanged (verified in order against the pristine
tag and by a fresh configure); see
docs/history/llama-cpp-breaking-changes.md for the range.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AytmJF9faEiQEVt6eetQS2
---
 CHANGELOG.md                                  |  3 +++
 CLAUDE.md                                     |  8 +++----
 README.md                                     |  2 +-
 docs/history/llama-cpp-breaking-changes.md    |  2 ++
 llama/CMakeLists.txt                          |  2 +-
 .../net/ladenthin/llama/args/ModelOption.java |  3 +++
 .../llama/parameters/ModelParameters.java     | 21 ++++++++++++++++++
 .../llama/value/LlamaCppVersion.java          |  8 +++----
 llama/src/test/cpp/test_server.cpp            | 22 +++++++++----------
 llama/src/test/cpp/test_utils.cpp             |  2 +-
 .../llama/parameters/ModelParametersTest.java | 21 ++++++++++++++++++
 11 files changed, 72 insertions(+), 22 deletions(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index 64c59759c..6cc3a6813 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -10,6 +10,9 @@ from version 5.0.0 onward. Pre-fork releases (`1.x`–`4.2.0`) were authored by
 ## [Unreleased]
 
 ### Added
+- **`ModelParameters.setMoeCacheMib(int)`** (`--moe-cache-mib`, llama.cpp b11480, EXPERIMENTAL upstream): a
+  GPU cache for the MoE expert weights a `setCpuMoeLayers` / `--cpu-moe` setup keeps in host memory,
+  split among several GPUs like the layers; `0` (the default) disables it.
 - **Kolibri-1 support** (Aleph Alpha, architecture `kolibri1`, 78B German/English reasoning MoE) ahead of upstream
   llama.cpp ([ggml-org/llama.cpp#29922](https://github.com/ggml-org/llama.cpp/issues/29922)), as the carried patch
   `0016-model-kolibri1.patch`. It combines the two community ports and, unlike either of them, loads the GGUFs of
diff --git a/CLAUDE.md b/CLAUDE.md
index 0132e6f59..98594ffd5 100644
--- a/CLAUDE.md
+++ b/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: **b11476**
+Current llama.cpp pinned version: **b11480**
 
 ## Natives jars: one directory per backend (`.github/natives.csv`)
 
@@ -787,7 +787,7 @@ 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 b11476 https://github.com/ggml-org/llama.cpp /tmp/lc
+git clone --depth 1 --branch b11480 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 \
@@ -827,7 +827,7 @@ 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 b11476`), the
+Because `sccache` is **content-addressed** and llama.cpp is pinned (`GIT_TAG b11480`), 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
@@ -1966,7 +1966,7 @@ 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 b11476`.
+llama.cpp is fetched via CMake FetchContent, pinned to `GIT_TAG b11480`.
 
 **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
diff --git a/README.md b/README.md
index 9a4d3b262..b4c848dd1 100644
--- a/README.md
+++ b/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 b11476](https://img.shields.io/badge/llama.cpp-%23b11476-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b11476)  
+[![llama.cpp b11480](https://img.shields.io/badge/llama.cpp-%23b11480-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b11480)  
 [![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)  
diff --git a/docs/history/llama-cpp-breaking-changes.md b/docs/history/llama-cpp-breaking-changes.md
index abb5e2b05..f116a1d2f 100644
--- a/docs/history/llama-cpp-breaking-changes.md
+++ b/docs/history/llama-cpp-breaking-changes.md
@@ -836,3 +836,5 @@ Used during `llama.cpp` version bumps: when upgrading, scan this file from the r
 | b11462–b11474 | patches + upstream verification | **All ten patches apply unchanged**, verified by a fresh configure through the real `FetchContent` path. Two patch-target files are in the range -- `src/llama-vocab.cpp` and `src/models/models.h`, both touched by `0016` -- but the upstream hunks (the FIM token lists and the GLM5-Next structs) are away from `0016`'s, so only line offsets moved. The drop-checks hold: nothing under `tools/server/`, `include/`, `common/arg.cpp`, `common/log.{h,cpp}`, `src/llama-model.cpp` or `ggml/src/ggml-rpc/` changed, and `git grep -i kolibri` finds nothing under `src/` at b11474. No `.github/` change upstream: OpenVINO, CCCL (`v3.4.3`) and the ROCm wheels stay where b11457 left them. |
 | b11474–b11476 | Two commits, 4 files, ~60/5 lines (8 KiB, one step), Metal only. #30100 fixes the `MUL_MAT`+`ADD` fusion when the residual is itself a `MUL_MAT` output: `ggml_metal_op_mul_mat_mma` picked the residual as "the operand that is not a `MUL_MAT`", which chose the wrong one when both are; it now takes the other operand of the `ADD` by identity. #30065 extends the few-row MMA mat-mul (`mul_mv_mma`) to the remaining `src0` types (BF16, the K-quants Q2_K/Q3_K, MXFP4, the IQ and TQ types), with per-type minimum row counts. Plus `test-backend-ops` cases (not built here). Only the shipped `metal-macos-aarch64` build and the two test-only macOS Metal builds compile it. **No project source change.** This is the target of the bump. |
 | b11474–b11476 | patches + upstream verification | **All ten patches apply unchanged**, verified in order against pristine b11476 with `git apply`. Nothing outside `ggml/src/ggml-metal/` and `tests/` changed, so no patch-target file is in the range and every drop-check of the b11474 row holds unchanged (`0016` included: still no `kolibri` under `src/`). No `.github/` change upstream. |
+| b11476–b11480 | Four commits, 43 files, ~1068/199 lines (94 KiB, one step). **#29876 hands the chat parser text *and* tokens.** `common_chat_input` (`common/chat.h`: the text plus one token id per byte, `LLAMA_TOKEN_NULL` where a byte has none; `append`/`prepend`/`truncate`/`substr`, and `common_chat_input_tokenize()`, which re-tokenizes a text and keeps the ids only when the pieces reproduce it) replaces the `std::string` of `common_chat_parse()` / `common_chat_peg_parse()`, of `common_chat_parser_params::generation_prompt` (the schema tokenizes it against the vocab now), of the slot's `generated` and of `server_task_result_cmpl_{partial,final}::content`; the PEG parse context carries the tokens (`peg-parser.h`), so a parser can match token ids instead of text. Every `to_json()` emits `content.text` -- request fields, bounds and response keys are identical between the tags. **Project change:** the `std::string` constructor is `explicit`, so `test_server.cpp`'s ten `content = "..."` assignments, its one `generation_prompt = "..."` and `test_utils.cpp`'s `common_chat_parse(raw, ...)` wrap the string in `common_chat_input(...)`; `jllama.cpp` reads results through `to_json()` only. **#29887 adds a GPU cache for the MoE experts kept in host memory**: `llama_context_params.moe_cache_size` (after `type_v`, EXPERIMENTAL), `common_params.moe_cache_size`, the CLI option `--moe-cache-mib N` (env `LLAMA_ARG_MOE_CACHE_MIB`, negatives rejected; the draft context gets 0) and the new `src/llama-moe-cache.{cpp,h}`: experts a `--n-cpu-moe` / `--cpu-moe` setup leaves in RAM are cached on the device up to that budget. Exposed as **`ModelParameters.setMoeCacheMib(int)`** (`ModelOption.MOE_CACHE_MIB`). #30097 moves the trunk-only / MTP-only tensor flags the NextN models each copied into a `nextn_flags` helper of `llama_model_base` (six more architectures accept a trunk-only file); #29425 CUDA `pool2d`. Plus `test-chat.cpp`, `test-llama-archs.cpp` and `test-backend-ops.cpp`, not built here. |
+| b11476–b11480 | patches + upstream verification | **All ten patches apply unchanged**, verified in order against pristine b11480 with `git apply`, then by a fresh configure through the real `FetchContent` path. Patch-target files in the range: `common/arg.cpp` (`0001`, `0015`: the new `--moe-cache-mib` option is away from `common_params_parse` and `add_rpc_devices`), `tools/server/server-context.cpp` (`0002`, `0003`: the `generated` rename is away from `load_model` and the similarity getter/setter) and `src/llama-model.{cpp,h}` (`0012`, `0016`: the `nextn_flags` helper is away from `load_tensors`' split block and the Kolibri-1 registrations). Drop-checks against pristine b11480 unchanged: `0001` override present and `common_params_parse_main` absent, `0002` assignment unconditional, no `0003` getter, no `0006`-`0008` counterpart, no `0012` guard, no `0014` hook (`common/log.h` still names only `common_log_default_callback`), no `0015` `stop_server`, no `kolibri` under `src/`. No `.github/` change upstream anywhere in b11476..b11512: CCCL `v3.4.3`, OpenVINO 2026.4.1 and ROCm 10.0.0 stay. |
diff --git a/llama/CMakeLists.txt b/llama/CMakeLists.txt
index 9b992e80b..300a73b69 100644
--- a/llama/CMakeLists.txt
+++ b/llama/CMakeLists.txt
@@ -216,7 +216,7 @@ endif()
 FetchContent_Declare(
 	llama.cpp
 	GIT_REPOSITORY https://github.com/ggerganov/llama.cpp.git
-	GIT_TAG        b11476
+	GIT_TAG        b11480
 	PATCH_COMMAND  ${CMAKE_COMMAND}
 		-DPATCH_DIR=${CMAKE_CURRENT_SOURCE_DIR}/patches
 		-DLLAMA_SRC=<SOURCE_DIR>
diff --git a/llama/src/main/java/net/ladenthin/llama/args/ModelOption.java b/llama/src/main/java/net/ladenthin/llama/args/ModelOption.java
index 145b2b3d0..729cbc696 100644
--- a/llama/src/main/java/net/ladenthin/llama/args/ModelOption.java
+++ b/llama/src/main/java/net/ladenthin/llama/args/ModelOption.java
@@ -191,6 +191,9 @@ public enum ModelOption {
     /** CLI option {@code --model-url}. */
     MODEL_URL("--model-url"),
 
+    /** CLI option {@code --moe-cache-mib}. */
+    MOE_CACHE_MIB("--moe-cache-mib"),
+
     /** CLI option {@code --n-cpu-ffn}. */
     N_CPU_FFN("--n-cpu-ffn"),
 
diff --git a/llama/src/main/java/net/ladenthin/llama/parameters/ModelParameters.java b/llama/src/main/java/net/ladenthin/llama/parameters/ModelParameters.java
index dd8541cd2..8f46103b9 100644
--- a/llama/src/main/java/net/ladenthin/llama/parameters/ModelParameters.java
+++ b/llama/src/main/java/net/ladenthin/llama/parameters/ModelParameters.java
@@ -909,6 +909,27 @@ public ModelParameters setCpuMoeLayers(int layers) {
         return putScalar(ModelOption.N_CPU_MOE, layers);
     }
 
+    /**
+     * Give the MoE expert weights kept in host memory a cache on the GPU of {@code mib} MiB.
+     *
+     * <p>Experts a {@link #setCpuMoeLayers(int)} / {@code --cpu-moe} setup leaves in RAM are
+     * otherwise read from host memory on every use. With a cache, llama.cpp keeps the most recently
+     * used experts on the device up to this budget; with several GPUs the budget is split among
+     * them like the layers ({@link #setTensorSplit(String)}). Upstream {@code --moe-cache-mib}
+     * (llama.cpp b11480, marked EXPERIMENTAL there); {@code 0}, the default, disables the cache.
+     * Without a GPU backend the option has no effect.</p>
+     *
+     * @param mib the cache size in MiB, {@code 0} to disable
+     * @return this builder
+     * @throws IllegalArgumentException if {@code mib} is negative (upstream rejects it as well)
+     */
+    public ModelParameters setMoeCacheMib(int mib) {
+        if (mib < 0) {
+            throw new IllegalArgumentException("Invalid moe-cache-mib value: " + mib + " (must be >= 0; 0 = disabled)");
+        }
+        return putScalar(ModelOption.MOE_CACHE_MIB, mib);
+    }
+
     /**
      * Keep the dense FFN weights of the first {@code n} layers on the CPU.
      *
diff --git a/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java b/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java
index 353a820bc..cf60c06fb 100644
--- a/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java
+++ b/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java
@@ -9,13 +9,13 @@
  * 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 "b11476"}) that mirrors the
+ * <p>{@link #LLAMA_CPP_VERSION} is a pure-Java string ({@code "b11480"}) 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 "b11476-<commit>"} — call
+ * plus the resolved upstream commit, e.g. {@code "b11480-<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>
@@ -23,14 +23,14 @@
 public final class LlamaCppVersion {
 
     /**
-     * The pinned llama.cpp release tag this library was built against, e.g. {@code "b11476"}.
+     * The pinned llama.cpp release tag this library was built against, e.g. {@code "b11480"}.
      *
      * <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 = "b11476";
+    public static final String LLAMA_CPP_VERSION = "b11480";
 
     // Constants holder — not instantiable.
     private LlamaCppVersion() {}
diff --git a/llama/src/test/cpp/test_server.cpp b/llama/src/test/cpp/test_server.cpp
index da59edba2..5a194c325 100644
--- a/llama/src/test/cpp/test_server.cpp
+++ b/llama/src/test/cpp/test_server.cpp
@@ -250,7 +250,7 @@ TEST(SlotParamsToJson, GrammarValue_OutputFormatGrammarExtracted) {
 
 TEST(SlotParamsToJson, GenerationPrompt_ReflectsSyntaxField) {
     task_params p;
-    p.chat_parser_params.generation_prompt = "Think step by step:";
+    p.chat_parser_params.generation_prompt = common_chat_input("Think step by step:");
 
     const json j = p.to_json();
 
@@ -1090,7 +1090,7 @@ TEST(ServerTaskResultCmplPartial, NonOaicompat_CoreFields) {
     server_task_result_cmpl_partial p;
     p.is_updated = true;
     p.res_type = TASK_RESPONSE_TYPE_NONE;
-    p.content = "hello";
+    p.content = common_chat_input("hello");
     p.n_decoded = 3;
     p.n_prompt_tokens = 10;
 
@@ -1192,7 +1192,7 @@ TEST(ServerTaskResultCmplFinal, IsStop_ReturnsTrue) {
 
 TEST(ServerTaskResultCmplFinal, NonOaicompat_StopAlwaysTrue) {
     server_task_result_cmpl_final f;
-    f.content = "done";
+    f.content = common_chat_input("done");
     f.n_decoded = 3;
     f.n_prompt_tokens = 7;
     const json j = f.to_json_non_oaicompat();
@@ -1307,7 +1307,7 @@ TEST(ServerTaskResultCmplFinal, UsageJsonOaicompat_TotalTokensIsSumOfBoth) {
 namespace {
 server_task_result_cmpl_final make_oai_final(const std::string &content = "hello") {
     server_task_result_cmpl_final f;
-    f.content = content;
+    f.content = common_chat_input(content);
     f.oaicompat_model = "test-model";
     f.oaicompat_cmpl_id = "cmpl-test";
     f.n_decoded = 3;
@@ -1575,7 +1575,7 @@ server_task_result_cmpl_partial make_partial(const std::string &content = "tok")
     server_task_result_cmpl_partial p;
     p.is_updated = true;
     p.res_type = TASK_RESPONSE_TYPE_OAI_CMPL;
-    p.content = content;
+    p.content = common_chat_input(content);
     p.oaicompat_model = "test-model";
     p.oaicompat_cmpl_id = "cmpl-part";
     return p;
@@ -1636,7 +1636,7 @@ TEST(CmplPartialToJsonDispatch, ResTypeNone_RoutesToNonOaicompat) {
     server_task_result_cmpl_partial p;
     p.is_updated = true;
     p.res_type = TASK_RESPONSE_TYPE_NONE;
-    p.content = "hello";
+    p.content = common_chat_input("hello");
     const json j = p.to_json(); // must not assert/abort
     // non-oaicompat shape has "content" directly
     EXPECT_EQ(j.at("content").get<std::string>(), "hello");
@@ -1646,7 +1646,7 @@ TEST(CmplPartialToJsonDispatch, ResTypeOaiCmpl_RoutesToOaicompat) {
     server_task_result_cmpl_partial p;
     p.is_updated = true;
     p.res_type = TASK_RESPONSE_TYPE_OAI_CMPL;
-    p.content = "hi";
+    p.content = common_chat_input("hi");
     p.oaicompat_model = "m";
     p.oaicompat_cmpl_id = "c";
     const json j = p.to_json();
@@ -1695,7 +1695,7 @@ server_task_result_cmpl_final make_dispatched_final(task_response_type rt, bool
     f.is_updated = true;
     f.res_type = rt;
     f.stream = stream;
-    f.content = "hi";
+    f.content = common_chat_input("hi");
     f.oaicompat_model = "m";
     f.oaicompat_cmpl_id = "id";
     return f;
@@ -2108,7 +2108,7 @@ TEST(ParamsFromJsonCmpl, PlainGrammar_NoGrammarType_SetsUserType) {
 
 TEST(CmplFinalResponseFields, EmptyList_AllFieldsPresent) {
     server_task_result_cmpl_final f;
-    f.content = "hi";
+    f.content = common_chat_input("hi");
     f.stop = STOP_TYPE_EOS;
     // response_fields is empty by default → full object returned
     const json j = f.to_json_non_oaicompat();
@@ -2119,7 +2119,7 @@ TEST(CmplFinalResponseFields, EmptyList_AllFieldsPresent) {
 
 TEST(CmplFinalResponseFields, NonEmptyList_OnlyRequestedFieldsPresent) {
     server_task_result_cmpl_final f;
-    f.content = "projected";
+    f.content = common_chat_input("projected");
     f.response_fields = {"content", "tokens_predicted"};
     const json j = f.to_json_non_oaicompat();
     EXPECT_TRUE(j.contains("content"));
@@ -2131,7 +2131,7 @@ TEST(CmplFinalResponseFields, NonEmptyList_OnlyRequestedFieldsPresent) {
 
 TEST(CmplFinalResponseFields, ContentValue_PreservedThroughProjection) {
     server_task_result_cmpl_final f;
-    f.content = "keep this";
+    f.content = common_chat_input("keep this");
     f.response_fields = {"content"};
     const json j = f.to_json_non_oaicompat();
     EXPECT_EQ(j.at("content").get<std::string>(), "keep this");
diff --git a/llama/src/test/cpp/test_utils.cpp b/llama/src/test/cpp/test_utils.cpp
index 135e25b7d..4cda471a7 100644
--- a/llama/src/test/cpp/test_utils.cpp
+++ b/llama/src/test/cpp/test_utils.cpp
@@ -1466,7 +1466,7 @@ namespace {
 // is_partial = false.
 std::string parse_content_only(const std::string &raw) {
     const common_chat_parser_params params;
-    return common_chat_parse(raw, /*is_partial=*/false, params).content;
+    return common_chat_parse(common_chat_input(raw), /*is_partial=*/false, params).content;
 }
 
 // U+FFFD REPLACEMENT CHARACTER, the byte sequence upstream substitutes for
diff --git a/llama/src/test/java/net/ladenthin/llama/parameters/ModelParametersTest.java b/llama/src/test/java/net/ladenthin/llama/parameters/ModelParametersTest.java
index 8500f4fcc..2b080a6f0 100644
--- a/llama/src/test/java/net/ladenthin/llama/parameters/ModelParametersTest.java
+++ b/llama/src/test/java/net/ladenthin/llama/parameters/ModelParametersTest.java
@@ -781,4 +781,25 @@ public void testSetDraftSamplingProbabilistic() {
         ModelParameters p = new ModelParameters().setDraftSampling(DraftSampling.PROBABILISTIC);
         assertThat(p.parameters.get("--spec-draft-sampling"), is("probabilistic"));
     }
+
+    // -------------------------------------------------------------------------
+    // setMoeCacheMib (llama.cpp b11480)
+    // -------------------------------------------------------------------------
+
+    @Test
+    public void testSetMoeCacheMib() {
+        ModelParameters p = new ModelParameters().setMoeCacheMib(2048);
+        assertThat(p.parameters.get("--moe-cache-mib"), is("2048"));
+    }
+
+    @Test
+    public void testSetMoeCacheMibZeroDisables() {
+        ModelParameters p = new ModelParameters().setMoeCacheMib(0);
+        assertThat(p.parameters.get("--moe-cache-mib"), is("0"));
+    }
+
+    @Test
+    public void testSetMoeCacheMibRejectsNegative() {
+        assertThrows(IllegalArgumentException.class, () -> new ModelParameters().setMoeCacheMib(-1));
+    }
 }

From 3398b7d23af4451686702ddea0e0adccb1e290fb Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Thu, 8 Oct 2026 19:10:17 +0000
Subject: [PATCH 3/8] Upgrade llama.cpp from b11480 to b11484

All ten patches apply unchanged (verified in order against the pristine
tag); see docs/history/llama-cpp-breaking-changes.md for the range.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AytmJF9faEiQEVt6eetQS2
---
 CLAUDE.md                                                 | 8 ++++----
 README.md                                                 | 2 +-
 docs/history/llama-cpp-breaking-changes.md                | 2 ++
 llama/CMakeLists.txt                                      | 2 +-
 .../java/net/ladenthin/llama/value/LlamaCppVersion.java   | 8 ++++----
 5 files changed, 12 insertions(+), 10 deletions(-)

diff --git a/CLAUDE.md b/CLAUDE.md
index 98594ffd5..dcc55e840 100644
--- a/CLAUDE.md
+++ b/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: **b11480**
+Current llama.cpp pinned version: **b11484**
 
 ## Natives jars: one directory per backend (`.github/natives.csv`)
 
@@ -787,7 +787,7 @@ 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 b11480 https://github.com/ggml-org/llama.cpp /tmp/lc
+git clone --depth 1 --branch b11484 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 \
@@ -827,7 +827,7 @@ 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 b11480`), the
+Because `sccache` is **content-addressed** and llama.cpp is pinned (`GIT_TAG b11484`), 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
@@ -1966,7 +1966,7 @@ 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 b11480`.
+llama.cpp is fetched via CMake FetchContent, pinned to `GIT_TAG b11484`.
 
 **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
diff --git a/README.md b/README.md
index b4c848dd1..43e6d3d7d 100644
--- a/README.md
+++ b/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 b11480](https://img.shields.io/badge/llama.cpp-%23b11480-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b11480)  
+[![llama.cpp b11484](https://img.shields.io/badge/llama.cpp-%23b11484-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b11484)  
 [![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)  
diff --git a/docs/history/llama-cpp-breaking-changes.md b/docs/history/llama-cpp-breaking-changes.md
index f116a1d2f..8dae6d5fc 100644
--- a/docs/history/llama-cpp-breaking-changes.md
+++ b/docs/history/llama-cpp-breaking-changes.md
@@ -838,3 +838,5 @@ Used during `llama.cpp` version bumps: when upgrading, scan this file from the r
 | b11474–b11476 | patches + upstream verification | **All ten patches apply unchanged**, verified in order against pristine b11476 with `git apply`. Nothing outside `ggml/src/ggml-metal/` and `tests/` changed, so no patch-target file is in the range and every drop-check of the b11474 row holds unchanged (`0016` included: still no `kolibri` under `src/`). No `.github/` change upstream. |
 | b11476–b11480 | Four commits, 43 files, ~1068/199 lines (94 KiB, one step). **#29876 hands the chat parser text *and* tokens.** `common_chat_input` (`common/chat.h`: the text plus one token id per byte, `LLAMA_TOKEN_NULL` where a byte has none; `append`/`prepend`/`truncate`/`substr`, and `common_chat_input_tokenize()`, which re-tokenizes a text and keeps the ids only when the pieces reproduce it) replaces the `std::string` of `common_chat_parse()` / `common_chat_peg_parse()`, of `common_chat_parser_params::generation_prompt` (the schema tokenizes it against the vocab now), of the slot's `generated` and of `server_task_result_cmpl_{partial,final}::content`; the PEG parse context carries the tokens (`peg-parser.h`), so a parser can match token ids instead of text. Every `to_json()` emits `content.text` -- request fields, bounds and response keys are identical between the tags. **Project change:** the `std::string` constructor is `explicit`, so `test_server.cpp`'s ten `content = "..."` assignments, its one `generation_prompt = "..."` and `test_utils.cpp`'s `common_chat_parse(raw, ...)` wrap the string in `common_chat_input(...)`; `jllama.cpp` reads results through `to_json()` only. **#29887 adds a GPU cache for the MoE experts kept in host memory**: `llama_context_params.moe_cache_size` (after `type_v`, EXPERIMENTAL), `common_params.moe_cache_size`, the CLI option `--moe-cache-mib N` (env `LLAMA_ARG_MOE_CACHE_MIB`, negatives rejected; the draft context gets 0) and the new `src/llama-moe-cache.{cpp,h}`: experts a `--n-cpu-moe` / `--cpu-moe` setup leaves in RAM are cached on the device up to that budget. Exposed as **`ModelParameters.setMoeCacheMib(int)`** (`ModelOption.MOE_CACHE_MIB`). #30097 moves the trunk-only / MTP-only tensor flags the NextN models each copied into a `nextn_flags` helper of `llama_model_base` (six more architectures accept a trunk-only file); #29425 CUDA `pool2d`. Plus `test-chat.cpp`, `test-llama-archs.cpp` and `test-backend-ops.cpp`, not built here. |
 | b11476–b11480 | patches + upstream verification | **All ten patches apply unchanged**, verified in order against pristine b11480 with `git apply`, then by a fresh configure through the real `FetchContent` path. Patch-target files in the range: `common/arg.cpp` (`0001`, `0015`: the new `--moe-cache-mib` option is away from `common_params_parse` and `add_rpc_devices`), `tools/server/server-context.cpp` (`0002`, `0003`: the `generated` rename is away from `load_model` and the similarity getter/setter) and `src/llama-model.{cpp,h}` (`0012`, `0016`: the `nextn_flags` helper is away from `load_tensors`' split block and the Kolibri-1 registrations). Drop-checks against pristine b11480 unchanged: `0001` override present and `common_params_parse_main` absent, `0002` assignment unconditional, no `0003` getter, no `0006`-`0008` counterpart, no `0012` guard, no `0014` hook (`common/log.h` still names only `common_log_default_callback`), no `0015` `stop_server`, no `kolibri` under `src/`. No `.github/` change upstream anywhere in b11476..b11512: CCCL `v3.4.3`, OpenVINO 2026.4.1 and ROCm 10.0.0 stay. |
+| b11480–b11484 | Four commits, 42 files, ~635/106 lines (95 KiB, one step). **#30088 names the tool and argument rules of the generated PEG parsers by index** instead of by the tool's and the arguments' names (`common/parsers`, 17 files, `chat-auto-parser-generator.cpp`, `chat-peg-parser.cpp`): a tool named `call` collided with a built-in rule and sent the parser into infinite recursion (#29967). The grammars the server emits are otherwise the same; `test_server.cpp`'s grammar assertions are literal user grammars and untouched, and tool calling through `ToolDefinition` / the OpenAI-compatible server stays covered by `ToolCallingIntegrationTest`. **#30110 adds the LiquidAI d1-3B decision model**: `COMMON_DECISION_TYPE_LFM2_D1` (`"lfm2-d1"`, openjev-style labels that depend on the question type), `server_task::decision::label_groups` (several label tokens per output, the output is their maximum -- reduced in `server-context.cpp` right after the logits are read), `server-decision.{h,cpp}`, and mtmd reads the LFM2 image-resize algorithm from the GGUF. `LlamaModel.handleSystemOne` and `ModelMeta.isDecisionModel()` (output modalities `["decisions"]`) serve it unchanged. #30062 adds the cohere2 vision projector to mtmd, #29100 CUDA FWHT kernels for block widths 1024-8192. **No project source change.** |
+| b11480–b11484 | patches + upstream verification | **All ten patches apply unchanged**, verified in order against pristine b11484 with `git apply`. Patch-target file in the range: `tools/server/server-context.cpp` (`0002`, `0003`: the `label_groups` reduction sits in the decision result path, away from `load_model` and the similarity getter/setter). Nothing under `common/arg.cpp`, `common/log.h`, `src/llama-model.cpp` or `ggml-rpc` changed, so every drop-check of the b11480 row holds unchanged. |
diff --git a/llama/CMakeLists.txt b/llama/CMakeLists.txt
index 300a73b69..ecfdead8a 100644
--- a/llama/CMakeLists.txt
+++ b/llama/CMakeLists.txt
@@ -216,7 +216,7 @@ endif()
 FetchContent_Declare(
 	llama.cpp
 	GIT_REPOSITORY https://github.com/ggerganov/llama.cpp.git
-	GIT_TAG        b11480
+	GIT_TAG        b11484
 	PATCH_COMMAND  ${CMAKE_COMMAND}
 		-DPATCH_DIR=${CMAKE_CURRENT_SOURCE_DIR}/patches
 		-DLLAMA_SRC=<SOURCE_DIR>
diff --git a/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java b/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java
index cf60c06fb..4a8d0ca6a 100644
--- a/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java
+++ b/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java
@@ -9,13 +9,13 @@
  * 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 "b11480"}) that mirrors the
+ * <p>{@link #LLAMA_CPP_VERSION} is a pure-Java string ({@code "b11484"}) 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 "b11480-<commit>"} — call
+ * plus the resolved upstream commit, e.g. {@code "b11484-<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>
@@ -23,14 +23,14 @@
 public final class LlamaCppVersion {
 
     /**
-     * The pinned llama.cpp release tag this library was built against, e.g. {@code "b11480"}.
+     * The pinned llama.cpp release tag this library was built against, e.g. {@code "b11484"}.
      *
      * <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 = "b11480";
+    public static final String LLAMA_CPP_VERSION = "b11484";
 
     // Constants holder — not instantiable.
     private LlamaCppVersion() {}

From e075e725b8226aa4e19aa6c102e4cedbfbbb70f7 Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Thu, 8 Oct 2026 19:10:17 +0000
Subject: [PATCH 4/8] Upgrade llama.cpp from b11484 to b11487

All ten patches apply unchanged (verified in order against the pristine
tag); see docs/history/llama-cpp-breaking-changes.md for the range.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AytmJF9faEiQEVt6eetQS2
---
 CLAUDE.md                                                 | 8 ++++----
 README.md                                                 | 2 +-
 docs/history/llama-cpp-breaking-changes.md                | 2 ++
 llama/CMakeLists.txt                                      | 2 +-
 .../java/net/ladenthin/llama/value/LlamaCppVersion.java   | 8 ++++----
 5 files changed, 12 insertions(+), 10 deletions(-)

diff --git a/CLAUDE.md b/CLAUDE.md
index dcc55e840..884be5775 100644
--- a/CLAUDE.md
+++ b/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: **b11484**
+Current llama.cpp pinned version: **b11487**
 
 ## Natives jars: one directory per backend (`.github/natives.csv`)
 
@@ -787,7 +787,7 @@ 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 b11484 https://github.com/ggml-org/llama.cpp /tmp/lc
+git clone --depth 1 --branch b11487 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 \
@@ -827,7 +827,7 @@ 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 b11484`), the
+Because `sccache` is **content-addressed** and llama.cpp is pinned (`GIT_TAG b11487`), 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
@@ -1966,7 +1966,7 @@ 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 b11484`.
+llama.cpp is fetched via CMake FetchContent, pinned to `GIT_TAG b11487`.
 
 **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
diff --git a/README.md b/README.md
index 43e6d3d7d..01878bc90 100644
--- a/README.md
+++ b/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 b11484](https://img.shields.io/badge/llama.cpp-%23b11484-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b11484)  
+[![llama.cpp b11487](https://img.shields.io/badge/llama.cpp-%23b11487-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b11487)  
 [![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)  
diff --git a/docs/history/llama-cpp-breaking-changes.md b/docs/history/llama-cpp-breaking-changes.md
index 8dae6d5fc..5aced3253 100644
--- a/docs/history/llama-cpp-breaking-changes.md
+++ b/docs/history/llama-cpp-breaking-changes.md
@@ -840,3 +840,5 @@ Used during `llama.cpp` version bumps: when upgrading, scan this file from the r
 | b11476–b11480 | patches + upstream verification | **All ten patches apply unchanged**, verified in order against pristine b11480 with `git apply`, then by a fresh configure through the real `FetchContent` path. Patch-target files in the range: `common/arg.cpp` (`0001`, `0015`: the new `--moe-cache-mib` option is away from `common_params_parse` and `add_rpc_devices`), `tools/server/server-context.cpp` (`0002`, `0003`: the `generated` rename is away from `load_model` and the similarity getter/setter) and `src/llama-model.{cpp,h}` (`0012`, `0016`: the `nextn_flags` helper is away from `load_tensors`' split block and the Kolibri-1 registrations). Drop-checks against pristine b11480 unchanged: `0001` override present and `common_params_parse_main` absent, `0002` assignment unconditional, no `0003` getter, no `0006`-`0008` counterpart, no `0012` guard, no `0014` hook (`common/log.h` still names only `common_log_default_callback`), no `0015` `stop_server`, no `kolibri` under `src/`. No `.github/` change upstream anywhere in b11476..b11512: CCCL `v3.4.3`, OpenVINO 2026.4.1 and ROCm 10.0.0 stay. |
 | b11480–b11484 | Four commits, 42 files, ~635/106 lines (95 KiB, one step). **#30088 names the tool and argument rules of the generated PEG parsers by index** instead of by the tool's and the arguments' names (`common/parsers`, 17 files, `chat-auto-parser-generator.cpp`, `chat-peg-parser.cpp`): a tool named `call` collided with a built-in rule and sent the parser into infinite recursion (#29967). The grammars the server emits are otherwise the same; `test_server.cpp`'s grammar assertions are literal user grammars and untouched, and tool calling through `ToolDefinition` / the OpenAI-compatible server stays covered by `ToolCallingIntegrationTest`. **#30110 adds the LiquidAI d1-3B decision model**: `COMMON_DECISION_TYPE_LFM2_D1` (`"lfm2-d1"`, openjev-style labels that depend on the question type), `server_task::decision::label_groups` (several label tokens per output, the output is their maximum -- reduced in `server-context.cpp` right after the logits are read), `server-decision.{h,cpp}`, and mtmd reads the LFM2 image-resize algorithm from the GGUF. `LlamaModel.handleSystemOne` and `ModelMeta.isDecisionModel()` (output modalities `["decisions"]`) serve it unchanged. #30062 adds the cohere2 vision projector to mtmd, #29100 CUDA FWHT kernels for block widths 1024-8192. **No project source change.** |
 | b11480–b11484 | patches + upstream verification | **All ten patches apply unchanged**, verified in order against pristine b11484 with `git apply`. Patch-target file in the range: `tools/server/server-context.cpp` (`0002`, `0003`: the `label_groups` reduction sits in the decision result path, away from `load_model` and the similarity getter/setter). Nothing under `common/arg.cpp`, `common/log.h`, `src/llama-model.cpp` or `ggml-rpc` changed, so every drop-check of the b11480 row holds unchanged. |
+| b11484–b11487 | Three commits, 14 files, ~919/27 lines (46 KiB, one step). **#30096 fixes the Jinja parser for TranslateGemma** (`common/chat.cpp`, three parser files, the template under `models/templates/`, `test-chat.cpp`) and logs a missing `source_lang_code` / `target_lang_code` instead of failing silently -- a chat-template fix that reaches the embedded server through `--jinja` without any project change. The rest is the Hexagon backend (tiled Q4_K/Q6_K `GET_ROWS` #30115, GELU accuracy #30104), which no natives jar builds. **No project source change.** |
+| b11484–b11487 | patches + upstream verification | **All ten patches apply unchanged**, verified in order against pristine b11487 with `git apply`. No patch-target file in the range; nothing under `src/`, `tools/server/`, `common/arg.cpp`, `common/log.h` or `include/` changed, so every drop-check of the b11480 row holds unchanged. |
diff --git a/llama/CMakeLists.txt b/llama/CMakeLists.txt
index ecfdead8a..4d3f9e795 100644
--- a/llama/CMakeLists.txt
+++ b/llama/CMakeLists.txt
@@ -216,7 +216,7 @@ endif()
 FetchContent_Declare(
 	llama.cpp
 	GIT_REPOSITORY https://github.com/ggerganov/llama.cpp.git
-	GIT_TAG        b11484
+	GIT_TAG        b11487
 	PATCH_COMMAND  ${CMAKE_COMMAND}
 		-DPATCH_DIR=${CMAKE_CURRENT_SOURCE_DIR}/patches
 		-DLLAMA_SRC=<SOURCE_DIR>
diff --git a/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java b/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java
index 4a8d0ca6a..6c135f537 100644
--- a/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java
+++ b/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java
@@ -9,13 +9,13 @@
  * 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 "b11484"}) that mirrors the
+ * <p>{@link #LLAMA_CPP_VERSION} is a pure-Java string ({@code "b11487"}) 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 "b11484-<commit>"} — call
+ * plus the resolved upstream commit, e.g. {@code "b11487-<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>
@@ -23,14 +23,14 @@
 public final class LlamaCppVersion {
 
     /**
-     * The pinned llama.cpp release tag this library was built against, e.g. {@code "b11484"}.
+     * The pinned llama.cpp release tag this library was built against, e.g. {@code "b11487"}.
      *
      * <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 = "b11484";
+    public static final String LLAMA_CPP_VERSION = "b11487";
 
     // Constants holder — not instantiable.
     private LlamaCppVersion() {}

From 2e387365fea3d41c691ea2f7d6ab3c8be06e70f7 Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Thu, 8 Oct 2026 19:10:17 +0000
Subject: [PATCH 5/8] Upgrade llama.cpp from b11487 to b11491

All ten patches apply unchanged (verified in order against the pristine
tag); see docs/history/llama-cpp-breaking-changes.md for the range.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AytmJF9faEiQEVt6eetQS2
---
 CLAUDE.md                                                 | 8 ++++----
 README.md                                                 | 2 +-
 docs/history/llama-cpp-breaking-changes.md                | 2 ++
 llama/CMakeLists.txt                                      | 2 +-
 .../java/net/ladenthin/llama/value/LlamaCppVersion.java   | 8 ++++----
 5 files changed, 12 insertions(+), 10 deletions(-)

diff --git a/CLAUDE.md b/CLAUDE.md
index 884be5775..97ff69d69 100644
--- a/CLAUDE.md
+++ b/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: **b11487**
+Current llama.cpp pinned version: **b11491**
 
 ## Natives jars: one directory per backend (`.github/natives.csv`)
 
@@ -787,7 +787,7 @@ 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 b11487 https://github.com/ggml-org/llama.cpp /tmp/lc
+git clone --depth 1 --branch b11491 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 \
@@ -827,7 +827,7 @@ 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 b11487`), the
+Because `sccache` is **content-addressed** and llama.cpp is pinned (`GIT_TAG b11491`), 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
@@ -1966,7 +1966,7 @@ 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 b11487`.
+llama.cpp is fetched via CMake FetchContent, pinned to `GIT_TAG b11491`.
 
 **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
diff --git a/README.md b/README.md
index 01878bc90..208ed55d2 100644
--- a/README.md
+++ b/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 b11487](https://img.shields.io/badge/llama.cpp-%23b11487-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b11487)  
+[![llama.cpp b11491](https://img.shields.io/badge/llama.cpp-%23b11491-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b11491)  
 [![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)  
diff --git a/docs/history/llama-cpp-breaking-changes.md b/docs/history/llama-cpp-breaking-changes.md
index 5aced3253..c45ca241f 100644
--- a/docs/history/llama-cpp-breaking-changes.md
+++ b/docs/history/llama-cpp-breaking-changes.md
@@ -842,3 +842,5 @@ Used during `llama.cpp` version bumps: when upgrading, scan this file from the r
 | b11480–b11484 | patches + upstream verification | **All ten patches apply unchanged**, verified in order against pristine b11484 with `git apply`. Patch-target file in the range: `tools/server/server-context.cpp` (`0002`, `0003`: the `label_groups` reduction sits in the decision result path, away from `load_model` and the similarity getter/setter). Nothing under `common/arg.cpp`, `common/log.h`, `src/llama-model.cpp` or `ggml-rpc` changed, so every drop-check of the b11480 row holds unchanged. |
 | b11484–b11487 | Three commits, 14 files, ~919/27 lines (46 KiB, one step). **#30096 fixes the Jinja parser for TranslateGemma** (`common/chat.cpp`, three parser files, the template under `models/templates/`, `test-chat.cpp`) and logs a missing `source_lang_code` / `target_lang_code` instead of failing silently -- a chat-template fix that reaches the embedded server through `--jinja` without any project change. The rest is the Hexagon backend (tiled Q4_K/Q6_K `GET_ROWS` #30115, GELU accuracy #30104), which no natives jar builds. **No project source change.** |
 | b11484–b11487 | patches + upstream verification | **All ten patches apply unchanged**, verified in order against pristine b11487 with `git apply`. No patch-target file in the range; nothing under `src/`, `tools/server/`, `common/arg.cpp`, `common/log.h` or `include/` changed, so every drop-check of the b11480 row holds unchanged. |
+| b11487–b11491 | Four commits, 28 files, ~1104/70 lines (92 KiB, one step). **#30114 adds the LiquidAI d1-omni-600M decision model** (audio and image input): `COMMON_DECISION_TYPE_LFM2_D1_OMNI` (`"lfm2-d1-omni"`, laya-style with another prompt layout), a memory-less LFM2 graph for the non-causal case (`src/models/lfm2.cpp`, +295; `models.h`, `llama-model.cpp`), mtmd's conformer GLU sigmoid kept on CUDA and new `mtmd-audio` helpers, and in the server audio arrives through `images` / `input_audio`, with `images` now an alias of `files` on `/v1/systemone` (`server-decision.{h,cpp}`, +233; README). `LlamaModel.handleSystemOne` hands the request body through, so the alias and the audio input are available to Java callers as they are. #29507 removes duplicate SYCL block-size defines; #30126/#30121 Hexagon (`alloc_buffer_n`, Q6_K dequant). **No project source change.** |
+| b11487–b11491 | patches + upstream verification | **All ten patches apply unchanged**, verified in order against pristine b11491 with `git apply`. Patch-target files in the range: `src/llama-model.cpp` and `src/models/models.h` (`0012`, `0016`: the LFM2 additions are away from `load_tensors`' split block and from the Kolibri-1 registration and struct) and `tools/server/server-context.cpp` (`0002`, `0003`; a two-line change in the decision path). Drop-checks unchanged: still no `split_sum` guard, still no `kolibri` under `src/`. |
diff --git a/llama/CMakeLists.txt b/llama/CMakeLists.txt
index 4d3f9e795..22383897b 100644
--- a/llama/CMakeLists.txt
+++ b/llama/CMakeLists.txt
@@ -216,7 +216,7 @@ endif()
 FetchContent_Declare(
 	llama.cpp
 	GIT_REPOSITORY https://github.com/ggerganov/llama.cpp.git
-	GIT_TAG        b11487
+	GIT_TAG        b11491
 	PATCH_COMMAND  ${CMAKE_COMMAND}
 		-DPATCH_DIR=${CMAKE_CURRENT_SOURCE_DIR}/patches
 		-DLLAMA_SRC=<SOURCE_DIR>
diff --git a/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java b/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java
index 6c135f537..b0047b1ea 100644
--- a/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java
+++ b/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java
@@ -9,13 +9,13 @@
  * 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 "b11487"}) that mirrors the
+ * <p>{@link #LLAMA_CPP_VERSION} is a pure-Java string ({@code "b11491"}) 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 "b11487-<commit>"} — call
+ * plus the resolved upstream commit, e.g. {@code "b11491-<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>
@@ -23,14 +23,14 @@
 public final class LlamaCppVersion {
 
     /**
-     * The pinned llama.cpp release tag this library was built against, e.g. {@code "b11487"}.
+     * The pinned llama.cpp release tag this library was built against, e.g. {@code "b11491"}.
      *
      * <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 = "b11487";
+    public static final String LLAMA_CPP_VERSION = "b11491";
 
     // Constants holder — not instantiable.
     private LlamaCppVersion() {}

From a6559ec01cf4740e587d57d4b669429a40aa34b3 Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Thu, 8 Oct 2026 19:10:18 +0000
Subject: [PATCH 6/8] Upgrade llama.cpp from b11491 to b11495

All ten patches apply unchanged (verified in order against the pristine
tag); see docs/history/llama-cpp-breaking-changes.md for the range.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AytmJF9faEiQEVt6eetQS2
---
 CLAUDE.md                                                 | 8 ++++----
 README.md                                                 | 2 +-
 docs/history/llama-cpp-breaking-changes.md                | 2 ++
 llama/CMakeLists.txt                                      | 2 +-
 .../java/net/ladenthin/llama/value/LlamaCppVersion.java   | 8 ++++----
 5 files changed, 12 insertions(+), 10 deletions(-)

diff --git a/CLAUDE.md b/CLAUDE.md
index 97ff69d69..5a566cb11 100644
--- a/CLAUDE.md
+++ b/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: **b11491**
+Current llama.cpp pinned version: **b11495**
 
 ## Natives jars: one directory per backend (`.github/natives.csv`)
 
@@ -787,7 +787,7 @@ 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 b11491 https://github.com/ggml-org/llama.cpp /tmp/lc
+git clone --depth 1 --branch b11495 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 \
@@ -827,7 +827,7 @@ 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 b11491`), the
+Because `sccache` is **content-addressed** and llama.cpp is pinned (`GIT_TAG b11495`), 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
@@ -1966,7 +1966,7 @@ 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 b11491`.
+llama.cpp is fetched via CMake FetchContent, pinned to `GIT_TAG b11495`.
 
 **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
diff --git a/README.md b/README.md
index 208ed55d2..8e1482971 100644
--- a/README.md
+++ b/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 b11491](https://img.shields.io/badge/llama.cpp-%23b11491-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b11491)  
+[![llama.cpp b11495](https://img.shields.io/badge/llama.cpp-%23b11495-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b11495)  
 [![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)  
diff --git a/docs/history/llama-cpp-breaking-changes.md b/docs/history/llama-cpp-breaking-changes.md
index c45ca241f..428d167ab 100644
--- a/docs/history/llama-cpp-breaking-changes.md
+++ b/docs/history/llama-cpp-breaking-changes.md
@@ -844,3 +844,5 @@ Used during `llama.cpp` version bumps: when upgrading, scan this file from the r
 | b11484–b11487 | patches + upstream verification | **All ten patches apply unchanged**, verified in order against pristine b11487 with `git apply`. No patch-target file in the range; nothing under `src/`, `tools/server/`, `common/arg.cpp`, `common/log.h` or `include/` changed, so every drop-check of the b11480 row holds unchanged. |
 | b11487–b11491 | Four commits, 28 files, ~1104/70 lines (92 KiB, one step). **#30114 adds the LiquidAI d1-omni-600M decision model** (audio and image input): `COMMON_DECISION_TYPE_LFM2_D1_OMNI` (`"lfm2-d1-omni"`, laya-style with another prompt layout), a memory-less LFM2 graph for the non-causal case (`src/models/lfm2.cpp`, +295; `models.h`, `llama-model.cpp`), mtmd's conformer GLU sigmoid kept on CUDA and new `mtmd-audio` helpers, and in the server audio arrives through `images` / `input_audio`, with `images` now an alias of `files` on `/v1/systemone` (`server-decision.{h,cpp}`, +233; README). `LlamaModel.handleSystemOne` hands the request body through, so the alias and the audio input are available to Java callers as they are. #29507 removes duplicate SYCL block-size defines; #30126/#30121 Hexagon (`alloc_buffer_n`, Q6_K dequant). **No project source change.** |
 | b11487–b11491 | patches + upstream verification | **All ten patches apply unchanged**, verified in order against pristine b11491 with `git apply`. Patch-target files in the range: `src/llama-model.cpp` and `src/models/models.h` (`0012`, `0016`: the LFM2 additions are away from `load_tensors`' split block and from the Kolibri-1 registration and struct) and `tools/server/server-context.cpp` (`0002`, `0003`; a two-line change in the decision path). Drop-checks unchanged: still no `split_sum` guard, still no `kolibri` under `src/`. |
+| b11491–b11495 | Four commits, 19 files, ~1461/43 lines (97 KiB, one step). **#29692 makes the reranker classifier activation a model property**: the new GGUF key `<arch>.classifier.activation` (`LLM_KV_CLASSIFIER_ACTIVATION`, `src/llama-arch.{h,cpp}`) sets `hparams.act_cls` (`"gelu"` → `GELU_ERF`, `"silu"`, `"tanh"`; anything else asserts at load), and `build_pooling` applies it to the RANK head instead of the hard-coded ModernBERT-gelu / otherwise-tanh choice; the default stays `tanh`, ModernBERT without the key keeps `gelu_erf`, so a GGUF converted before this (the CI jina reranker included) scores exactly as before, and `handleRerank` is unchanged. #27920 teaches the converter Qwen3.5 embedding models (Python plus a pooling tweak in `src/models`); #30087 CUDA GDN state columns per warp; #29245 SYCL grouped MoE XMX GEMM. **No project source change.** |
+| b11491–b11495 | patches + upstream verification | **All ten patches apply unchanged**, verified in order against pristine b11495 with `git apply`. Patch-target files in the range: `src/llama-arch.{cpp,h}` and `src/llama-model.cpp` (`0016` adds the Kolibri-1 arch entries next to the new classifier key; `0012`'s split block is untouched). Drop-checks unchanged. |
diff --git a/llama/CMakeLists.txt b/llama/CMakeLists.txt
index 22383897b..e0f1f6333 100644
--- a/llama/CMakeLists.txt
+++ b/llama/CMakeLists.txt
@@ -216,7 +216,7 @@ endif()
 FetchContent_Declare(
 	llama.cpp
 	GIT_REPOSITORY https://github.com/ggerganov/llama.cpp.git
-	GIT_TAG        b11491
+	GIT_TAG        b11495
 	PATCH_COMMAND  ${CMAKE_COMMAND}
 		-DPATCH_DIR=${CMAKE_CURRENT_SOURCE_DIR}/patches
 		-DLLAMA_SRC=<SOURCE_DIR>
diff --git a/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java b/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java
index b0047b1ea..8ef3a5055 100644
--- a/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java
+++ b/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java
@@ -9,13 +9,13 @@
  * 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 "b11491"}) that mirrors the
+ * <p>{@link #LLAMA_CPP_VERSION} is a pure-Java string ({@code "b11495"}) 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 "b11491-<commit>"} — call
+ * plus the resolved upstream commit, e.g. {@code "b11495-<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>
@@ -23,14 +23,14 @@
 public final class LlamaCppVersion {
 
     /**
-     * The pinned llama.cpp release tag this library was built against, e.g. {@code "b11491"}.
+     * The pinned llama.cpp release tag this library was built against, e.g. {@code "b11495"}.
      *
      * <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 = "b11491";
+    public static final String LLAMA_CPP_VERSION = "b11495";
 
     // Constants holder — not instantiable.
     private LlamaCppVersion() {}

From 80517a1cbfa9886de87289606fed46d3a005c32a Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Thu, 8 Oct 2026 19:10:18 +0000
Subject: [PATCH 7/8] Upgrade llama.cpp from b11495 to b11505

All ten patches apply unchanged (verified in order against the pristine
tag); see docs/history/llama-cpp-breaking-changes.md for the range.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AytmJF9faEiQEVt6eetQS2
---
 CLAUDE.md                                                 | 8 ++++----
 README.md                                                 | 2 +-
 docs/history/llama-cpp-breaking-changes.md                | 2 ++
 llama/CMakeLists.txt                                      | 2 +-
 .../java/net/ladenthin/llama/value/LlamaCppVersion.java   | 8 ++++----
 5 files changed, 12 insertions(+), 10 deletions(-)

diff --git a/CLAUDE.md b/CLAUDE.md
index 5a566cb11..6a43f6051 100644
--- a/CLAUDE.md
+++ b/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: **b11495**
+Current llama.cpp pinned version: **b11505**
 
 ## Natives jars: one directory per backend (`.github/natives.csv`)
 
@@ -787,7 +787,7 @@ 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 b11495 https://github.com/ggml-org/llama.cpp /tmp/lc
+git clone --depth 1 --branch b11505 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 \
@@ -827,7 +827,7 @@ 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 b11495`), the
+Because `sccache` is **content-addressed** and llama.cpp is pinned (`GIT_TAG b11505`), 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
@@ -1966,7 +1966,7 @@ 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 b11495`.
+llama.cpp is fetched via CMake FetchContent, pinned to `GIT_TAG b11505`.
 
 **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
diff --git a/README.md b/README.md
index 8e1482971..8b6c3e090 100644
--- a/README.md
+++ b/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 b11495](https://img.shields.io/badge/llama.cpp-%23b11495-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b11495)  
+[![llama.cpp b11505](https://img.shields.io/badge/llama.cpp-%23b11505-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b11505)  
 [![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)  
diff --git a/docs/history/llama-cpp-breaking-changes.md b/docs/history/llama-cpp-breaking-changes.md
index 428d167ab..e70d83138 100644
--- a/docs/history/llama-cpp-breaking-changes.md
+++ b/docs/history/llama-cpp-breaking-changes.md
@@ -846,3 +846,5 @@ Used during `llama.cpp` version bumps: when upgrading, scan this file from the r
 | b11487–b11491 | patches + upstream verification | **All ten patches apply unchanged**, verified in order against pristine b11491 with `git apply`. Patch-target files in the range: `src/llama-model.cpp` and `src/models/models.h` (`0012`, `0016`: the LFM2 additions are away from `load_tensors`' split block and from the Kolibri-1 registration and struct) and `tools/server/server-context.cpp` (`0002`, `0003`; a two-line change in the decision path). Drop-checks unchanged: still no `split_sum` guard, still no `kolibri` under `src/`. |
 | b11491–b11495 | Four commits, 19 files, ~1461/43 lines (97 KiB, one step). **#29692 makes the reranker classifier activation a model property**: the new GGUF key `<arch>.classifier.activation` (`LLM_KV_CLASSIFIER_ACTIVATION`, `src/llama-arch.{h,cpp}`) sets `hparams.act_cls` (`"gelu"` → `GELU_ERF`, `"silu"`, `"tanh"`; anything else asserts at load), and `build_pooling` applies it to the RANK head instead of the hard-coded ModernBERT-gelu / otherwise-tanh choice; the default stays `tanh`, ModernBERT without the key keeps `gelu_erf`, so a GGUF converted before this (the CI jina reranker included) scores exactly as before, and `handleRerank` is unchanged. #27920 teaches the converter Qwen3.5 embedding models (Python plus a pooling tweak in `src/models`); #30087 CUDA GDN state columns per warp; #29245 SYCL grouped MoE XMX GEMM. **No project source change.** |
 | b11491–b11495 | patches + upstream verification | **All ten patches apply unchanged**, verified in order against pristine b11495 with `git apply`. Patch-target files in the range: `src/llama-arch.{cpp,h}` and `src/llama-model.cpp` (`0016` adds the Kolibri-1 arch entries next to the new classifier key; `0012`'s split block is untouched). Drop-checks unchanged. |
+| b11495–b11505 | Ten commits, 19 files, ~821/286 lines (95 KiB, one step), backends and the vendored HTTP library. **cpp-httplib 0.60.1** (#30134): a `Range` suffix longer than the representation selects the whole representation (RFC 9110), the Windows certificate-chain check no longer rejects on `dwErrorStatus`, a `WebSocketClient` friend declaration -- compiled into `jllama` through `httplib.cpp`, none of it on the embedded server's path (no range requests, no TLS). **Vulkan:** `TOP_K` no longer hangs or returns wrong indices for +inf/NaN inputs and orders negative values correctly for k = 1 (#30107; a correctness fix for the Vulkan jars), sparse flash attention extended to coopmat2 (#30003). **CUDA:** norm-family kernels beyond the grid-dimension limits (#28175), arbitrary strides for f16/f32/bf16 unary ops (#29781) and `ROLL` (#29547). **SYCL:** FWHT (#29605), the fused delta-net alpha gate (#29687), model loading staged through a pinned ring buffer (#29608). Hexagon mmadd alignment (#30133). **No project source change.** |
+| b11495–b11505 | patches + upstream verification | **All ten patches apply unchanged**, verified in order against pristine b11505 with `git apply`. No patch-target file in the range; nothing under `src/`, `common/`, `tools/server/` or `include/` changed, so every drop-check of the b11480 row holds unchanged. |
diff --git a/llama/CMakeLists.txt b/llama/CMakeLists.txt
index e0f1f6333..9b2e06016 100644
--- a/llama/CMakeLists.txt
+++ b/llama/CMakeLists.txt
@@ -216,7 +216,7 @@ endif()
 FetchContent_Declare(
 	llama.cpp
 	GIT_REPOSITORY https://github.com/ggerganov/llama.cpp.git
-	GIT_TAG        b11495
+	GIT_TAG        b11505
 	PATCH_COMMAND  ${CMAKE_COMMAND}
 		-DPATCH_DIR=${CMAKE_CURRENT_SOURCE_DIR}/patches
 		-DLLAMA_SRC=<SOURCE_DIR>
diff --git a/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java b/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java
index 8ef3a5055..ba01e2b35 100644
--- a/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java
+++ b/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java
@@ -9,13 +9,13 @@
  * 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 "b11495"}) that mirrors the
+ * <p>{@link #LLAMA_CPP_VERSION} is a pure-Java string ({@code "b11505"}) 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 "b11495-<commit>"} — call
+ * plus the resolved upstream commit, e.g. {@code "b11505-<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>
@@ -23,14 +23,14 @@
 public final class LlamaCppVersion {
 
     /**
-     * The pinned llama.cpp release tag this library was built against, e.g. {@code "b11495"}.
+     * The pinned llama.cpp release tag this library was built against, e.g. {@code "b11505"}.
      *
      * <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 = "b11495";
+    public static final String LLAMA_CPP_VERSION = "b11505";
 
     // Constants holder — not instantiable.
     private LlamaCppVersion() {}

From 51be68665056d8ce4c192742b177966f83c8efa3 Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Thu, 8 Oct 2026 19:10:18 +0000
Subject: [PATCH 8/8] Upgrade llama.cpp from b11505 to b11512

All ten patches apply unchanged (verified in order against the pristine
tag); see docs/history/llama-cpp-breaking-changes.md for the range.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AytmJF9faEiQEVt6eetQS2
---
 CHANGELOG.md                                              | 8 ++++++++
 CLAUDE.md                                                 | 8 ++++----
 README.md                                                 | 2 +-
 docs/history/llama-cpp-breaking-changes.md                | 2 ++
 llama/CMakeLists.txt                                      | 2 +-
 .../java/net/ladenthin/llama/value/LlamaCppVersion.java   | 8 ++++----
 6 files changed, 20 insertions(+), 10 deletions(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index 6cc3a6813..990055e6b 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -79,6 +79,14 @@ from version 5.0.0 onward. Pre-fork releases (`1.x`–`4.2.0`) were authored by
   GitHub retires the `macos-14` image by 2026-11-02 and fails every `macos-14` job during its October
   brownouts. The label is pinned rather than `macos-latest`, and it is not `macos-15`, which the
   shipped Metal build and the no-Metal build already run on.
+- **Upgraded the pinned llama.cpp from b11476 to b11512**, in 7 reviewed steps, each ending at a tag.
+  No carried patch needed a refresh, and every drop-check still finds its defect. The chat parser now
+  receives the generated tokens next to the text (#29876, `common_chat_input`) -- a change in the C++
+  tests here, not on the wire: request fields, their bounds and the response keys are identical across
+  the range. Two things a user can notice: the embedded server speaks cpp-httplib 0.60.1 (b11505), and
+  a slot state file written from b11512 on carries the context checkpoints (#26004) in an appendix an
+  older file simply lacks, so files from earlier releases still restore. The new `--moe-cache-mib` is
+  under *Added*. Per-step record: `docs/history/llama-cpp-breaking-changes.md`.
 - **Upgraded the pinned llama.cpp from b11320 to b11476**, in 25 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`
diff --git a/CLAUDE.md b/CLAUDE.md
index 6a43f6051..98436f021 100644
--- a/CLAUDE.md
+++ b/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: **b11505**
+Current llama.cpp pinned version: **b11512**
 
 ## Natives jars: one directory per backend (`.github/natives.csv`)
 
@@ -787,7 +787,7 @@ 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 b11505 https://github.com/ggml-org/llama.cpp /tmp/lc
+git clone --depth 1 --branch b11512 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 \
@@ -827,7 +827,7 @@ 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 b11505`), the
+Because `sccache` is **content-addressed** and llama.cpp is pinned (`GIT_TAG b11512`), 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
@@ -1966,7 +1966,7 @@ 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 b11505`.
+llama.cpp is fetched via CMake FetchContent, pinned to `GIT_TAG b11512`.
 
 **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
diff --git a/README.md b/README.md
index 8b6c3e090..1fe4e1cd9 100644
--- a/README.md
+++ b/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 b11505](https://img.shields.io/badge/llama.cpp-%23b11505-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b11505)  
+[![llama.cpp b11512](https://img.shields.io/badge/llama.cpp-%23b11512-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b11512)  
 [![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)  
diff --git a/docs/history/llama-cpp-breaking-changes.md b/docs/history/llama-cpp-breaking-changes.md
index e70d83138..8f9646489 100644
--- a/docs/history/llama-cpp-breaking-changes.md
+++ b/docs/history/llama-cpp-breaking-changes.md
@@ -848,3 +848,5 @@ Used during `llama.cpp` version bumps: when upgrading, scan this file from the r
 | b11491–b11495 | patches + upstream verification | **All ten patches apply unchanged**, verified in order against pristine b11495 with `git apply`. Patch-target files in the range: `src/llama-arch.{cpp,h}` and `src/llama-model.cpp` (`0016` adds the Kolibri-1 arch entries next to the new classifier key; `0012`'s split block is untouched). Drop-checks unchanged. |
 | b11495–b11505 | Ten commits, 19 files, ~821/286 lines (95 KiB, one step), backends and the vendored HTTP library. **cpp-httplib 0.60.1** (#30134): a `Range` suffix longer than the representation selects the whole representation (RFC 9110), the Windows certificate-chain check no longer rejects on `dwErrorStatus`, a `WebSocketClient` friend declaration -- compiled into `jllama` through `httplib.cpp`, none of it on the embedded server's path (no range requests, no TLS). **Vulkan:** `TOP_K` no longer hangs or returns wrong indices for +inf/NaN inputs and orders negative values correctly for k = 1 (#30107; a correctness fix for the Vulkan jars), sparse flash attention extended to coopmat2 (#30003). **CUDA:** norm-family kernels beyond the grid-dimension limits (#28175), arbitrary strides for f16/f32/bf16 unary ops (#29781) and `ROLL` (#29547). **SYCL:** FWHT (#29605), the fused delta-net alpha gate (#29687), model loading staged through a pinned ring buffer (#29608). Hexagon mmadd alignment (#30133). **No project source change.** |
 | b11495–b11505 | patches + upstream verification | **All ten patches apply unchanged**, verified in order against pristine b11505 with `git apply`. No patch-target file in the range; nothing under `src/`, `common/`, `tools/server/` or `include/` changed, so every drop-check of the b11480 row holds unchanged. |
+| b11505–b11512 | Seven commits, 25 files, ~716/202 lines (79 KiB, the final step). **#26004 keeps the context checkpoints across a slot save and restore**: the slot file gains an appendix after the llama state payload (magic `SCKP`, version 1, a count, then per checkpoint `n_tokens`/`pos_min`/`pos_max` and the target, draft and speculative blobs), counted in the `n_written` / `n_read` of the save and restore results; on restore it is read bounds-checked -- a missing, truncated or foreign appendix is ignored with a warning, never fatal, so a slot file written before this restores as before -- and one draft checkpoint is test-loaded, the draft data dropped when it does not fit the current draft context, where `common_prompt_checkpoint::load_dft()` used to `GGML_ABORT` on a size mismatch (`load_tgt`/`load_dft` return `bool` now, `common/common.h`). `LlamaModel.handleSlotAction(SAVE/RESTORE)` carries the same payload; the response keys are unchanged. **#30112 splits the MoE expert cache across several GPUs** like the layers (`--tensor-split`; the `--moe-cache-mib` help text, `common_params`, `llama-moe-cache.cpp`) -- `setMoeCacheMib`'s Javadoc already says so. **CUDA:** MMQ out-of-bounds reads (#29953), a looped PAD kernel beyond 65535 rows or slices (#30147), and the CCCL version guard in `argsort.cu` compares the packed `CCCL_VERSION` (#29453: the per-component compare turned `STRIDED_ITERATOR_AVAILABLE` off on a major rollover -- with the pinned CCCL v3.4.3 it is true before and after, so no change for the CUDA jars until CCCL 4). #30111 reads DFlash's tied output weights from the GGUF metadata; #29668 WebUI settings on the first visit in router mode (auto-followed). **No project source change.** |
+| b11505–b11512 | patches + upstream verification | **All ten patches apply unchanged**, verified in order against pristine b11512 with `git apply` and by a fresh configure through the real `FetchContent` path. Patch-target files in the range: `common/arg.cpp` (`0001`, `0015`: a help-text change only) and `tools/server/server-context.cpp` (`0002`, `0003`: the checkpoint appendix helpers are private members added away from `load_model` and the similarity getter/setter). Drop-checks against pristine b11512 -- 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`); the load-progress assignment is unconditional (`0002`); no `slot_prompt_similarity` getter (`0003`); no embedded-mode, attach or worker-command counterpart (`0006`-`0008`); no `split_sum == 0` guard (`0012`); `common/log.h` names only `common_log_default_callback`, no sink (`0014`); no `stop_server` and the connect failure still aborts in `ggml-rpc.cpp` (`0015`); no `kolibri` under `src/` (`0016`). No `.github/` change upstream in the whole range. |
diff --git a/llama/CMakeLists.txt b/llama/CMakeLists.txt
index 9b2e06016..7371cc830 100644
--- a/llama/CMakeLists.txt
+++ b/llama/CMakeLists.txt
@@ -216,7 +216,7 @@ endif()
 FetchContent_Declare(
 	llama.cpp
 	GIT_REPOSITORY https://github.com/ggerganov/llama.cpp.git
-	GIT_TAG        b11505
+	GIT_TAG        b11512
 	PATCH_COMMAND  ${CMAKE_COMMAND}
 		-DPATCH_DIR=${CMAKE_CURRENT_SOURCE_DIR}/patches
 		-DLLAMA_SRC=<SOURCE_DIR>
diff --git a/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java b/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java
index ba01e2b35..54f6aea8e 100644
--- a/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java
+++ b/llama/src/main/java/net/ladenthin/llama/value/LlamaCppVersion.java
@@ -9,13 +9,13 @@
  * 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 "b11505"}) that mirrors the
+ * <p>{@link #LLAMA_CPP_VERSION} is a pure-Java string ({@code "b11512"}) 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 "b11505-<commit>"} — call
+ * plus the resolved upstream commit, e.g. {@code "b11512-<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>
@@ -23,14 +23,14 @@
 public final class LlamaCppVersion {
 
     /**
-     * The pinned llama.cpp release tag this library was built against, e.g. {@code "b11505"}.
+     * The pinned llama.cpp release tag this library was built against, e.g. {@code "b11512"}.
      *
      * <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 = "b11505";
+    public static final String LLAMA_CPP_VERSION = "b11512";
 
     // Constants holder — not instantiable.
     private LlamaCppVersion() {}