Upgrade llama.cpp from b9990 to b10015

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: **b9990**
+Current llama.cpp pinned version: **b10015**
 
 ## Upgrading CUDA Version
 
@@ -421,7 +421,7 @@ needs no extra step here, `build-webui` re-reads the tag and rebuilds the matchi
 ships no UI):
 ```bash
 # needs node/npm + network; embed.cpp is plain C++17 (no npm)
-git clone --depth 1 --branch b9990 https://github.com/ggml-org/llama.cpp /tmp/lc
+git clone --depth 1 --branch b10015 https://github.com/ggml-org/llama.cpp /tmp/lc
 ( cd /tmp/lc/tools/ui && npm ci && npm run build \
   && ( cd dist && find . -type f -not -path './_gzip/*' \
        | while read -r f; do mkdir -p "_gzip/$(dirname "$f")"; gzip -9 -c "$f" > "_gzip/$f"; done ) \
@@ -461,7 +461,7 @@ cache lives in **Depot Cache** over sccache's **WebDAV** backend:
 - `SCCACHE_WEBDAV_TOKEN: ${{ secrets.DEPOT_TOKEN }}` — a Depot **organization** token, stored
   as the repo secret **`DEPOT_TOKEN`**.
 
-Because `sccache` is **content-addressed** and llama.cpp is pinned (`GIT_TAG b9990`), the
+Because `sccache` is **content-addressed** and llama.cpp is pinned (`GIT_TAG b10015`), 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
@@ -572,7 +572,7 @@ Current patches:
 | `0001-win32-arg-parse-embed-guard.patch` | Windows JNI regression from llama.cpp **#24779** (introduced b9739): on Windows `common_params_parse` re-derived argv from the **process** command line (`GetCommandLineW`) and adopted it, so an embedded/JNI caller (`java.exe`) lost its `--model …` args → "Failed to parse model parameters". b9789 narrowed the unconditional override to a **count-guard** (`if (static_cast<int>(utf8.buf.size()) == argc) { argv = utf8.ptrs.data(); }`), but that is exactly the variant the project already found breaks its Windows server-integration tests (when the embedded argv length coincides with `java.exe`'s). The patch carries the **complete upstream change** (so it can be submitted to llama.cpp verbatim and then dropped here): **(1)** `common_params_parse` parses **exactly the argv it is given** (no `GetCommandLineW` magic) and a new `common_params_parse_main()` wrapper holds the UTF-8 recovery for the standalone tools' `main()` (`common/arg.{cpp,h}`); **(2)** the **~34 standalone `main()` call sites** (every `common_params_parse(argc, argv, …)` across `tools/*`, `examples/*` and the `tests/*` programs) flip to `common_params_parse_main()`; **(3)** a `tests/test-arg-parser.cpp` regression case pins that `common_params_parse` honors a caller-supplied argv. The embedded caller (`jllama.cpp`) keeps calling `common_params_parse` and is never overridden. **Our subproject build compiles only the `arg.{cpp,h}` core** — `LLAMA_BUILD_TOOLS`/`LLAMA_BUILD_TESTS` are OFF for a FetchContent subproject — so the flips + test are applied-but-not-compiled here; they were validated via a one-off `-DLLAMA_BUILD_TOOLS=ON -DLLAMA_BUILD_TESTS=ON` build (the new test compiles and its asserts pass; `test-arg-parser`'s only red there is the live `ggml.ai` download check, which is sandbox-network, not the patch). Because it spans **37 files** it must be refreshed on every llama.cpp bump (the applier fails loud). |
 | `0002-server-preserve-caller-load-progress-callback.patch` | Load-progress-callback regression introduced in llama.cpp **b9789**: `server_context::load_model` (`tools/server/server-context.cpp`) now **unconditionally** installs the server's own load-progress reporter on `params_base.load_progress_callback` immediately before `common_init_from_params`, clobbering any callback the embedding caller already set. libjllama's `LoadProgressCallback` feature wires `common_params.load_progress_callback` to a JNI trampoline *before* calling `load_model`, so the bump silently killed it — `LoadProgressCallbackTest` saw zero progress updates and the abort-on-`false` path never threw. The patch guards the assignment with `if (params_base.load_progress_callback == nullptr)`, so the server installs its own reporter **only when the caller hasn't** — a caller-supplied callback survives and fires during load. Standalone `llama-server` (no caller callback, so the field is null) is unaffected. Same JNI-vs-standalone divergence class as `0001`. |
 | `0003-pr22393-server-add-slot-prompt-similarity-getter-setter.patch` | **Upstream-PR carry** of [ggml-org/llama.cpp#22393](https://github.com/ggml-org/llama.cpp/pull/22393) ("server : add slot_prompt_similarity getter/setter") while it is still open upstream. Purely additive: adds `server_context::get_slot_prompt_similarity()` / `set_slot_prompt_similarity(float)` (`tools/server/server-context.{cpp,h}`) so an embedding/JNI caller can query and tune the slot-selection threshold at runtime without reloading the model. Verbatim copy of the PR — drop it once a pinned `b<nnnn>` includes the change. |
-| `0007-server-attach-http-frontend.patch` | **Adds `llama_server_attach(argc, argv, server_context&)`** so the `NativeServer` *attach mode* can serve an **already-loaded `LlamaModel`** over the full upstream HTTP frontend — no second model load, no `start_loop()`; the LlamaModel's worker keeps driving the shared `server_context` and the HTTP routes post tasks to its queue (the queue is the synchronization point). Mechanically: (1) extracts the common route table + CORS-proxy/tools blocks out of `llama_server()` into `llama_server_register_common_routes(...)` (shared verbatim, so the entry points cannot drift; returns `false` on tools-setup failure); (2) adds `llama_server_attach`, which parses only the HTTP-side argv via `common_params_parse`, starts `g_stream_sessions` GC + `server_http_context`, registers the common routes plus the non-router resumable-streaming handlers, marks ready immediately (model already loaded), and blocks on the HTTP thread until `llama_server_request_shutdown()` — never calling `common_init()`, backend init, `ctx_server.terminate()` or `llama_backend_free()` (the embedding caller owns those). Applies after `0001`+`0006` (same file); closes the "NativeServer — reuse an already-loaded LlamaModel" TODO. Upstream-submittable ("server: let embedding callers attach the HTTP frontend to an existing server_context"). |
+| `0007-server-attach-http-frontend.patch` | **Adds `llama_server_attach(argc, argv, server_context&)`** so the `NativeServer` *attach mode* can serve an **already-loaded `LlamaModel`** over the full upstream HTTP frontend — no second model load, no `start_loop()`; the LlamaModel's worker keeps driving the shared `server_context` and the HTTP routes post tasks to its queue (the queue is the synchronization point). Mechanically: (1) extracts the common route table + the CORS-proxy/tools **route bindings** out of `llama_server()` into `llama_server_register_common_routes(...)` (shared, so the entry points cannot drift; returns `false` on tools-setup failure) — the per-caller **experimental-feature warnings** (upstream's b10015/#25655 `cors_origins` security warning + the router/MCP/tools `warn_names` block) stay in `llama_server` since they depend on router state the shared helper does not carry; (2) adds `llama_server_attach`, which parses only the HTTP-side argv via `common_params_parse`, starts `g_stream_sessions` GC + `server_http_context`, registers the common routes plus the non-router resumable-streaming handlers, marks ready immediately (model already loaded), and blocks on the HTTP thread until `llama_server_request_shutdown()` — never calling `common_init()`, backend init, `ctx_server.terminate()` or `llama_backend_free()` (the embedding caller owns those). Applies after `0001`+`0006` (same file); closes the "NativeServer — reuse an already-loaded LlamaModel" TODO. Upstream-submittable ("server: let embedding callers attach the HTTP frontend to an existing server_context"). |
 | `0008-server-models-worker-cmd-override.patch` | **Makes router mode usable in-JVM.** The router (`server-models.cpp`) spawns each model worker by re-executing its own binary (`get_server_exec_path()` = `/proc/self/exe` & friends) — inside a JVM that binary is `java`, not a llama-server, so embedded router workers could never start. The patch adds env `LLAMA_SERVER_WORKER_CMD` (whitespace-split; read in `server_model_meta::update_args`) which replaces only the leading binary-path token of the rendered worker args, letting an embedding host relaunch workers through its own bootstrap — e.g. `java -cp app.jar net.ladenthin.llama.server.NativeServer` (each worker is then a fresh JVM running the classic single-model `NativeServer`). Exposed in Java as `NativeServer.setWorkerCommand(String...)` (JNI `setenv`); exercised by `RouterModeIntegrationTest` (Linux CI). Upstream-submittable (also useful for containerized/wrapped deployments). |
 | `0006-server-embed-native-server-jni.patch` | **Makes `server.cpp`'s `llama_server` embeddable in the JVM** so the `NativeServer` JNI bridge can run the full upstream HTTP server (WebUI included) inside `libjllama` — see "Two server modes" below. b9870 already exposes `int llama_server(int, char**)` (non-static; no `main` in the file), so the patch only adds embedded-mode support: (1) a `g_llama_server_embedded` flag + `llama_server_set_embedded()` / `llama_server_request_shutdown()` (declared in the committed `src/main/cpp/native_server_bridge.h`); (2) skips installing the process-wide SIGINT/SIGTERM handlers when embedded (they would hijack the JVM's); (3) in embedded mode parses the **forwarded** argv via `common_params_parse` instead of `common_params_parse_main` (whose `GetCommandLineW` recovery would pick up `java.exe`'s command line — the same Windows class of bug `0001` fixes). `llama_server_request_shutdown()` mirrors the SIGTERM path (invokes the installed `shutdown_handler` → `ctx_server.terminate()` unblocks `start_loop()`), giving JNI an out-of-band stop since `ctx_server` is loop-local. Applies **after `0001`** (which flips this call site to `common_params_parse_main`), so its context is the post-`0001` tree; regenerate against `0001`+source on a bump. Only touches `tools/server/server.cpp`. |
 
@@ -1254,7 +1254,7 @@ ctest --test-dir build --output-on-failure -R "ResultsToJson"
 
 #### Upstream source location (in CMake build tree)
 
-llama.cpp is fetched via CMake FetchContent, pinned to `GIT_TAG b9990`.
+llama.cpp is fetched via CMake FetchContent, pinned to `GIT_TAG b10015`.
 
 **GoogleTest** is a separate `BUILD_TESTING`-only FetchContent (`GIT_TAG v1.17.0`), used solely
 by the `jllama_test` C++ unit-test binary — not by the shipped library, and not coupled to the

README.md

@@ -7,7 +7,7 @@
 **Build:**  
 ![Java 8+](https://img.shields.io/badge/Java-8%2B-informational)  
 ![Platform](https://img.shields.io/badge/Platform-Linux%20%7C%20macOS%20%7C%20Windows%20%7C%20Android-lightgrey)  
-[![llama.cpp b9990](https://img.shields.io/badge/llama.cpp-%23b9990-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b9990)  
+[![llama.cpp b10015](https://img.shields.io/badge/llama.cpp-%23b10015-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b10015)  
 [![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

@@ -491,3 +491,7 @@ Used during `llama.cpp` version bumps: when upgrading, scan this file from the r
 | b9982–b9984 | upstream verification (sandbox) | All **six** patches (`0001`–`0003`, `0006`–`0008`) re-verified against b9984: applied in filename order onto a clean b9984 checkout via a fresh `cmake -B build` (fail-loud `PATCH_COMMAND`), all clean. No OuteTTS generator anchor touched (`tools/tts/tts.cpp` unchanged — the generator extracted `tts.cpp @ b9984`). Full local verification: fresh configure (fail-loud patch apply + TTS extraction succeeded, `ggml commit e474bba7a`) + full `cmake --build` (`jllama` + `jllama_test` link cleanly) + `ctest` **485/485 passing**; per-platform confirmation by the CI pipeline. |
 | b9984–b9986 | `common/{chat-auto-parser-generator.cpp,chat-diff-analyzer.cpp}` + `ggml/src/ggml-sycl/{backend.hpp,common.hpp,ggml-sycl.cpp,topk-moe.{cpp,hpp}}` + `tests/test-chat.cpp` | **Additive/tuning-only, no public-API surface (9 files, ~32 KiB excluding the auto-followed `tools/ui` WebUI rework, 2 commits).** `chat-auto-parser-generator.cpp`/`chat-diff-analyzer.cpp` fix a whitespace-trimming and preserved-tokens bug in the NVIDIA-Nemotron-Nano-v2 auto-parser profile (`reasoning.start` trimmed before use; `tools_array_wrapped` flag added; a stale `<SPECIAL_12>` preserved-token entry dropped) — neither file is a patch target or in the priority-8 header list, and this project links `chat.cpp`/`chat.h` (unchanged) rather than these newer split-out analyzer files directly, but they compile into the same upstream-compiled `llama-common` static lib either way. `tests/test-chat.cpp` gains matching regression cases for the fixed template output (not a patch target). `ggml/src/ggml-sycl/*` gains new top-k MoE (mixture-of-experts) kernels — entirely inside the SYCL backend's upstream-compiled TUs (this project's SYCL classifiers are build-only, no source touched). No `common/arg.*`, `tools/server/*`, or OuteTTS generator anchor (`tools/tts/tts.cpp` unchanged) touched; all eight priority-8 headers byte-identical. All **six** remaining patches (`0001`–`0003`, `0006`–`0008`) apply unchanged — confirmed via the upstream `b9984...b9986` compare diff. This chunk reaches the latest release at the time of the bump (b9986). |
 | b9984–b9986 | upstream verification (sandbox) | All **six** patches (`0001`–`0003`, `0006`–`0008`) re-verified against b9986: applied in filename order onto a clean b9986 checkout via a fresh `cmake -B build` (fail-loud `PATCH_COMMAND`), all clean. No OuteTTS generator anchor touched (`tools/tts/tts.cpp` unchanged — the generator extracted `tts.cpp @ b9986`). Full local verification: fresh configure (fail-loud patch apply + TTS extraction succeeded, `ggml commit 91c631b21`) + full `cmake --build` (`jllama` + `jllama_test` link cleanly) + `ctest` **485/485 passing**; per-platform confirmation by the CI pipeline. |
+| b9986–b9990 | `ggml/{include/gguf.h,src/gguf.cpp}` + `src/models/minimax-m2.cpp` + `tests/**` | **Additive-only, no public-API surface (8 files, ~2 KiB excluding the auto-followed `tools/ui` WebUI + upstream `README.md`, 4 commits).** `ggml/{include/gguf.h,src/gguf.cpp}` add a GGUF **tensor-shape accessor** (#24405) — purely additive API on the GGUF reader; `ggml/include/gguf.h` is not in the priority-8 review list and no existing signature changed. `src/models/minimax-m2.cpp` adds Minimax2 eagle3 speculative-decoding support (a model TU inside the upstream-compiled `llama` lib). `tests/**` (`test-alloc`/`test-backend-ops`/`test-gguf` header harmonization + a shape-accessor case) are upstream tools, not compiled or shipped here. No `common/arg.*`, `tools/server/*`, or OuteTTS generator anchor (`tools/tts/tts.cpp` unchanged) touched; all eight priority-8 headers byte-identical. All **six** patches (`0001`–`0003`, `0006`–`0008`) apply unchanged. |
+| b9986–b9990 | upstream verification (compare diff) | All **six** patches (`0001`–`0003`, `0006`–`0008`) apply unchanged against b9990: the `b9986...b9990` compare diff touches no patch-target file (`common/arg.*`, `tools/server/*`) and no OuteTTS generator anchor (`tools/tts/tts.cpp` unchanged), and all eight priority-8 headers are byte-identical. b9990 was this session's starting pin (landed via PR #341) and configures/builds cleanly as the base from which the b9990→b10015 bump proceeded; the full local `cmake --build` + `ctest` **485/485** was run at b10015 (the rows below), not separately at b9990. Per-platform confirmation by the CI pipeline. (Backfilled — the b9986→b9990 bump omitted its history rows when it landed.) |
+| b9990–b10015 | `common/{arg.cpp,common.h}` + `tools/server/{server.cpp,server-context.cpp,server-http.cpp,server-task.{cpp,h}}` + `ggml/src/{ggml-cpu/kleidiai/**,ggml-cuda/mmq*,ggml-metal/**,ggml-sycl/**,ggml-vulkan/**,ggml-opencl/**,ggml-hexagon/**}` + `src/{llama-arch.{cpp,h},llama-model.cpp,models/hy-v3.cpp}` + `tests/**` | **Additive + one patch-target restructure (85 files, ~7.7k insertions excluding the auto-followed `tools/ui` WebUI, 25 commits — the first 5-digit `b1nnnn` build number).** Priority-8 headers: `common/common.h` is **additive-only** (new `LLAMA_EXAMPLE_TOKENIZE` enumerator + `cors_*` server params for #25655 + `tokenize_*` params for #25516 — no field removed/renamed); `common/arg.cpp` gains the `--cors-*` options (#25655), the tokenize-CLI common-args alignment (#25516) and a one-line `common_log_flush(common_log_main())` before `exit(0)` (#25504) — all **outside** patch `0001`'s `common_params_parse`/`common_params_parse_main` region, so `0001` applies unchanged. **One patch-target restructure — `0007` REFRESHED:** upstream #25655 ("server: add --cors-* options") rewrote the CORS-proxy/tools warning region of `tools/server/server.cpp` that `0007`'s deletion hunk relocated (introducing a `warn_names` vector, an `is_router_server` branch and a new `cors_origins=='*'` security warning), so `0007` no longer applied. It was refreshed so `llama_server_register_common_routes()` now extracts only the **route bindings** (health…slots + GCP + `res_403` + CORS-proxy + tools setup), while the per-caller **experimental-feature warnings** (incl. upstream's new `cors_origins` warning) stay in `llama_server()` — they depend on router state the shared helper does not carry, and keeping them there preserves #25655 with no upstream-warning regression. **#25649 ("server: refactor prompt cache state ownership")** rewrote `server-context.cpp`'s `server_slot`/private-method regions — a different region than patch `0002`'s `load_model` load-progress-callback guard and `0003`'s `get/set_slot_prompt_similarity` getters (both untouched), so `0002`/`0003` apply unchanged. `server-http.cpp` (CORS-header plumbing) and `server-task.{cpp,h}` (prompt-cache-state fields) are not patch targets. The remaining ggml backend churn (kleidiai SME2 f32, CUDA MMQ config refactor, Metal Q2_0, SYCL Battlemage, Vulkan native mxfp4/nvfp4, OpenCL OOB/dp4a fixes, Hexagon) + new Hunyuan Hy3 (`hy_v3`) model support are inside upstream-compiled TUs. b10015 is the topmost release at bump time; the `b[0-9]+` bump tooling and the manual's `b<nnnn>` placeholder are digit-agnostic and handled the 4→5-digit crossover unchanged. |
+| b9990–b10015 | upstream verification (sandbox) | **Six** patches re-verified against b10015 via a fresh `cmake -B build` (fail-loud `PATCH_COMMAND`): `0001`–`0003`, `0006`, `0008` apply **unchanged**; `0007` was **refreshed** for upstream #25655's CORS restructure (route bindings extracted to the shared helper, per-caller warnings kept in `llama_server()` — see the row above) and now applies clean. No OuteTTS generator anchor touched (`tools/tts/tts.cpp` unchanged — the generator extracted `tts.cpp @ b10015`). Full local verification: fresh configure (fail-loud patch apply + TTS extraction succeeded, `ggml commit 12127defd`) + full `cmake --build` (`jllama` + `jllama_test` link cleanly, confirming the refreshed `0007` `server.cpp` restructure compiles) + `ctest` **485/485 passing**; per-platform confirmation by the CI pipeline. |