Script 'mail_helper' called by obssrc Hello community, here is the log from the commit of package llamacpp for openSUSE:Factory checked in at 2026-03-05 17:14:49 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ Comparing /work/SRC/openSUSE:Factory/llamacpp (Old) and /work/SRC/openSUSE:Factory/.llamacpp.new.561 (New) ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Package is "llamacpp" Thu Mar 5 17:14:49 2026 rev:26 rq:1336614 version:8189 Changes: -------- --- /work/SRC/openSUSE:Factory/llamacpp/llamacpp.changes 2026-01-29 17:48:48.315161962 +0100 +++ /work/SRC/openSUSE:Factory/.llamacpp.new.561/llamacpp.changes 2026-03-05 17:18:11.546212897 +0100 @@ -1,0 +2,35 @@ +Mon Mar 2 22:29:33 UTC 2026 - Robert Munteanu <[email protected]> + +- Update to version 8189: + * CUDA: CDNA3 MFMA support for FA MMA kernel, improved CUDA graph + capture, dequantization optimizations, and grid.y cap fix in + non-contiguous kernels. + * Vulkan: Intel/AMD tuning, FA scalar and coopmat1 refactors, + new ops (L2_NORM, GGML_OP_SET), overlap check before fusion, + and correctness fixes for rope, fp16 FA, and mul_mat_id. + * CPU and GGML: mxfp4 repack, SVE/RVV kernel additions, q5_K and + q6_K repack with dotprod and i8mm, tiled FA for prompt + processing, s390x optimizations, and extended bin bcast. + * Accelerator backends: new ops and optimizations across OpenCL, + Metal, WebGPU, SYCL, CANN, HIP/ROCm, Hexagon, ZenDNN, and a + new VirtGPU backend for Virglrenderer API remoting. + * Models and conversion: added Kimi-K2.5, Kimi Linear (MLA KV + cache), Jina Embeddings v5 Nano, Kanana-2, JAIS-2, PaddleOCR-VL, + GLM-OCR, GLM MoE DSA, Tiny Aya, full modern BERT, Step3.5-Flash, + Qwen3.5 series, and Devstral-2; graph deduplication and + conversion fixes across multiple architectures. + * Server: /v1/responses mirroring, multi-model alias support, + multi-modal prompt caching and context checkpoints, + max_completion_tokens property, and various stability fixes. + * KV cache, MTMD, and Jinja: M-RoPE shift fix, V-less cache + support, hybrid model size fix, multimodal tiling and padding + fixes, and multiple Jinja correctness and feature additions. + * WebUI: full-height code blocks, raw LLM output switcher, system + message injection, router mode fixes, and Svelte update. + * Misc: self-speculative decoding without a draft model, NetBSD + support, ggml bumped to 0.9.7, gguf-py to 0.18.0, updated + miniaudio, cpp-httplib, and BoringSSL. + * Full commit log: + https://github.com/ggml-org/llama.cpp/compare/b7789...b8189 + +------------------------------------------------------------------- Old: ---- llamacpp-7789.tar.gz New: ---- llamacpp-8189.tar.gz ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ Other differences: ------------------ ++++++ llamacpp.spec ++++++ --- /var/tmp/diff_new_pack.LEufmB/_old 2026-03-05 17:18:12.742262395 +0100 +++ /var/tmp/diff_new_pack.LEufmB/_new 2026-03-05 17:18:12.746262560 +0100 @@ -25,11 +25,11 @@ %global mtmd_sover 0.0.%{version} %global mtmd_sover_suffix 0 -%global ggml_sover 0.9.5 +%global ggml_sover 0.9.7 %global ggml_sover_suffix 0 Name: llamacpp -Version: 7789 +Version: 8189 Release: 0 Summary: Inference of Meta's LLaMA model (and others) in pure C/C++ License: MIT ++++++ llamacpp-7789.tar.gz -> llamacpp-8189.tar.gz ++++++ /work/SRC/openSUSE:Factory/llamacpp/llamacpp-7789.tar.gz /work/SRC/openSUSE:Factory/.llamacpp.new.561/llamacpp-8189.tar.gz differ: char 28, line 1
