From 78e0e0207c758d2a4bb06ba36a6305200e8f125d Mon Sep 17 00:00:00 2001 From: Niko Maroulis Date: Mon, 6 Jul 2026 13:14:59 -0400 Subject: [PATCH] Bump llama.cpp to cb295bf (b9888) for v0.8.33 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Fold the latest upstream binding into the unreleased v0.8.33. Forward bump from 2d973636e (b9870, v0.8.32) — 18 commits. All headers the binding compiles against are unchanged in this range; the diff only touches llama core src and the statically-linked ggml backends (CUDA/HIP/Vulkan/CPU), so no NIF changes were required. CHANGELOG updated only; @version stays 0.8.33 and checksum.exs is left for CI to regenerate on the release tag. --- CHANGELOG.md | 8 ++++++++ vendor/llama.cpp | 2 +- 2 files changed, 9 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 5c04580..d00c031 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -57,6 +57,14 @@ is 1.6x faster than the stateless path. Full suite: 252 tests, 0 failures. ### Changed +- **llama.cpp submodule** — Updated from 2d973636e to cb295bf59 (18 commits, tag b9888). No NIF changes were required: every header the binding compiles against — `include/llama.h`, `ggml/include/ggml.h`, `ggml/include/ggml-backend.h`, `common/common.h`, `common/chat.h`, `common/json-schema-to-grammar.h`, `common/sampling.h`, and `common/speculative.h` — has zero diff in this range (nothing under `include/` or `common/` changed at all). The range only touches `llama` core `src/` and the statically-linked `ggml` backend implementations (CUDA, HIP, Vulkan, CPU) plus webui/scripts/server-test code the NIF does not link against, so the bump is picked up purely by resyncing the static libraries on rebuild. + - **llama core**: guard the K/V rotation input when the buffer is unallocated (#25215); fix stale tensor-split params for draft models (#24814). + - **ggml**: fix a tensor-parallel + `-ncmoe` crash on MoE models (#25028); abort on a multi-buffer in `ggml-backend-meta` (#25276); fix the broken CPU concat implementation for quantized types (#25247). + - **CUDA**: extend K-type validation to V-types for flash attention (#24403); add a concat implementation for quantized types (#25303); optimize `conv_transpose_1d` indexing (#25310); VMM-pool allocation Turing P2P access fix (#24491). + - **ggml-cpu**: use a UE4M3 LUT in the ARM NVFP4 dot product (#25331); enable tiled matmul on AIX (#25199). + - **ggml-hip**: enable `-ffast-math` for HIP builds (#23862). + - **vulkan**: fix a 32-bit integer overflow in `CEIL_DIV` (#25245). + - **ui/scripts/tests** (not linked into the binding): restore the Ctrl+B sidebar toggle (#25307), fake a 200 for proxy `DELETE` requests (#25298), and add sync blocks so display/behavior settings honor `--ui-config-file` (#25132); use `HF_TOKEN` when downloading UI assets (#25280); temporarily skip the model-downloading API test (#25355). - **Server hot loop** — one fused dirty-CPU NIF per tick (`batch_eval_sample`: decode + per-slot sampling + detokenization + EOG check) replaces the previous 1 dirty + (2 normal-scheduler NIFs + send) × generating-slots pattern; streamed diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 2d97363..cb295bf 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 2d973636e292ee6f75fadcf08d29cb33511f509f +Subproject commit cb295bf59663cd3577389315636772f4060bd1f5