Script 'mail_helper' called by obssrc Hello community, here is the log from the commit of package llamacpp for openSUSE:Factory checked in at 2026-10-15 17:08:48 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ Comparing /work/SRC/openSUSE:Factory/llamacpp (Old) and /work/SRC/openSUSE:Factory/.llamacpp.new.3698736 (New) ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Package is "llamacpp" Thu Oct 15 17:08:48 2026 rev:37 rq:1384106 version:0.6.0 Changes: -------- --- /work/SRC/openSUSE:Factory/llamacpp/llamacpp.changes 2026-10-01 22:21:07.194631172 +0200 +++ /work/SRC/openSUSE:Factory/.llamacpp.new.3698736/llamacpp.changes 2026-10-15 17:08:52.039569765 +0200 @@ -1,0 +2,11 @@ +Tue Oct 06 00:00:00 UTC 2026 - Eyad Issa <[email protected]> + +- Update to version 0.6.0 + * Add the extended batch API, GLM-5.3-Flash and Clef models, and + Qwen4Exp MTP speculative decoding. + * Add the /v1/systemone API and a new Web UI model download + pipeline. + * Update ggml to 0.26.0 with new Metal and Vulkan flash attention + kernels. + +------------------------------------------------------------------- @@ -5,2 +16,2 @@ - * Add HRM-Text, MiMo-V2.6, and HunyuanOCR support; extend Nemotron - MTP and Nemotron-H handling. + * Add HRM-Text, MiMo-V2.6, and HunyuanOCR support; extend + Nemotron MTP and Nemotron-H handling. @@ -11,3 +22,4 @@ - and improve router lifecycle, token counting, and startup logging. - * Improve chat parsers, JSON schema handling, multimodal validation, - and model save/load support. + and improve router lifecycle, token counting, and startup + logging. + * Improve chat parsers, JSON schema handling, multimodal + validation, and model save/load support. @@ -15,3 +27,4 @@ - attention, IQ4_XS Vulkan kernels, and Metal f32 x bf16 matmul variants. - * Include the b11146 Web UI asset with WEBM support and responsive - layout, toast, and reasoning-menu fixes. + attention, IQ4_XS Vulkan kernels, and Metal f32 x bf16 matmul + variants. + * Include the b11146 Web UI asset with WEBM support and + responsive layout, toast, and reasoning-menu fixes. @@ -24,3 +37,5 @@ - * Improve JSON schema and chat parsing, structured JSONL logging, model - conversion, multimodal decoding, and server child-process management. - * Improve UI rendering and conversation export; include b10964 Web UI. + * Improve JSON schema and chat parsing, structured JSONL logging, + model conversion, multimodal decoding, and server child-process + management. + * Improve UI rendering and conversation export; include b10964 + Web UI. @@ -28,2 +43,2 @@ - * GGML gains precision-control APIs and major Vulkan, SYCL, Hexagon, - OpenCL, CUDA/HIP, Metal, CPU, and WebGPU improvements. + * GGML gains precision-control APIs and major Vulkan, SYCL, + Hexagon, OpenCL, CUDA/HIP, Metal, CPU, and WebGPU improvements. @@ -34 +49,2 @@ -- Own per-subpackage license directories so the SUSE filelist check passes. +- Own per-subpackage license directories so the SUSE filelist check + passes. @@ -39,2 +55,2 @@ -- Ship native and Web UI third-party license notices in every binary and - development subpackage. +- Ship native and Web UI third-party license notices in every + binary and development subpackage. @@ -47,7 +63,8 @@ - * Initial Qwen3.8-Flash-Next and NVIDIA Nemotron-3-Puzzle support. - * Add on-demand tensor reading, per-slot server context limits, and - video input options. - * GGML updated to 0.23.0 with sparse flash attention, asynchronous - backend APIs, and Apple RDMA transport. - * Improve CUDA, Vulkan, Metal, SYCL, OpenCL, and other accelerator - backends. + * Initial Qwen3.8-Flash-Next and NVIDIA Nemotron-3-Puzzle + support. + * Add on-demand tensor reading, per-slot server context limits, + and video input options. + * GGML updated to 0.23.0 with sparse flash attention, + asynchronous backend APIs, and Apple RDMA transport. + * Improve CUDA, Vulkan, Metal, SYCL, OpenCL, and other + accelerator backends. @@ -72,2 +89,3 @@ -- Add skip-sme-variants-when-unsupported.patch to skip SME CPU variants - when unsupported, allowing aarch64 builds on Leap 16.0 with GCC 13 +- Add skip-sme-variants-when-unsupported.patch to skip SME CPU + variants when unsupported, allowing aarch64 builds on Leap 16.0 + with GCC 13 @@ -92,5 +110,5 @@ - tuned flash-attn vec (M4, M4 Pro, M3 Max, M5, M5 Pro), - chunked SSD MMA for Mamba-2 prefill, dequantize q8_0 using - packed types, null-check buffer alloc fix, memory leak fix. - * CUDA: MMVQ nwarps=8 for bs=1 dense models on DGX Spark, - unblock mmq for MoE on sm_60. + tuned flash-attn vec (M4, M4 Pro, M3 Max, M5, M5 Pro), chunked + SSD MMA for Mamba-2 prefill, dequantize q8_0 using packed + types, null-check buffer alloc fix, memory leak fix. + * CUDA: MMVQ nwarps=8 for bs=1 dense models on DGX Spark, unblock + mmq for MoE on sm_60. @@ -102,4 +120,4 @@ - * WebGPU: mulmat with overlapping src0/src1 (for MiniMax-01), - fix infinity handling in ARGSORT and TOP_K. - * RPC: implement event and async backend APIs, support Apple - RDMA as transport. + * WebGPU: mulmat with overlapping src0/src1 (for MiniMax-01), fix + infinity handling in ARGSORT and TOP_K. + * RPC: implement event and async backend APIs, support Apple RDMA + as transport. @@ -109,3 +127,3 @@ - ctx-per-slot (--kv-unified-per-slot), dedup-cache-models - preset option, dedup-cache-models, and reject prefilled - assistant messages with tool calls. + ctx-per-slot (--kv-unified-per-slot), dedup-cache-models preset + option, dedup-cache-models, and reject prefilled assistant + messages with tool calls. @@ -113,2 +131,2 @@ - support --models-dir loading MTP assistant models, share - thread pools when n_threads differ. + support --models-dir loading MTP assistant models, share thread + pools when n_threads differ. @@ -120 +138,2 @@ - Content from MCP tool result, alphabetical enum member ordering. + Content from MCP tool result, alphabetical enum member + ordering. @@ -122,3 +141,3 @@ - fixes (Granite4 Vision, LFM2 tiling threshold, video moov - atom, sha256 input hashing), grammar char class hyphen, - unicode collapsed symbol class. + fixes (Granite4 Vision, LFM2 tiling threshold, video moov atom, + sha256 input hashing), grammar char class hyphen, unicode + collapsed symbol class. @@ -128,2 +147,2 @@ - * OpenVINO: updated to 2026.3.1, whisper.cpp support, - Qwen3.5 on NPU, new ops. + * OpenVINO: updated to 2026.3.1, whisper.cpp support, Qwen3.5 on + NPU, new ops. @@ -137,5 +156,5 @@ - * New model support: Nanbeige 4.2, MiniMax-M3 (MSA sparse attention), - GLM-5.2-Vision, Laguna XS.2 and M.1. - * CUDA: GET_ROWS for quantized types, improved NVFP4 W4A4 activation - quantization, sqrt_softplus in top-k MoE for DeepSeek v4, and a - fix for external compilation of Q1_0 MMQ. + * New model support: Nanbeige 4.2, MiniMax-M3 (MSA sparse + attention), GLM-5.2-Vision, Laguna XS.2 and M.1. + * CUDA: GET_ROWS for quantized types, improved NVFP4 W4A4 + activation quantization, sqrt_softplus in top-k MoE for + DeepSeek v4, and a fix for external compilation of Q1_0 MMQ. @@ -145,5 +164,5 @@ - im2col support, activation ops update, and Windows crash fix when - op_poll is enabled. - * OpenCL: cache compiled cl_program binaries on disk, fix fused RMS - norm mul view offset, and do not treat NULL-mask flash attention - as causal. + im2col support, activation ops update, and Windows crash fix + when op_poll is enabled. + * OpenCL: cache compiled cl_program binaries on disk, fix fused + RMS norm mul view offset, and do not treat NULL-mask flash + attention as causal. @@ -153,7 +172,9 @@ - * CPU/ggml: enable BF16 tiled GEMM on PowerPC, fix offloading logic - for weight backend, update PowerPC backend variants on AIX. - * Server/API: support MCP stdio tool transport, add format argument - to datetime tool, support "reasoning_effort": "none" in OAI API, - improve error handling for null context and conversation IDs, fix - stream routes for model names containing a slash, and add missing - task parameters in generation_settings. + * CPU/ggml: enable BF16 tiled GEMM on PowerPC, fix offloading + logic for weight backend, update PowerPC backend variants on + AIX. + * Server/API: support MCP stdio tool transport, add format + argument to datetime tool, support "reasoning_effort": "none" + in OAI API, improve error handling for null context and + conversation IDs, fix stream routes for model names containing + a slash, and add missing task parameters in + generation_settings. @@ -162,8 +183,9 @@ - sampler, infer speculative type from draft repo sidecars, refactor - mlock/mmap/directio into load-mode, and add -lm mlock option. - * UI: symbolic math support via nerdamer in JS sandbox, fix symbolic - math tool prompt, fix system message edit box, detect conversation - import format from file contents, reduce per-token render cost, - remove render effects, fix MCP server display name conflicts, add - "Default" option for reasoning selector, and fix context gauge - regressions. + sampler, infer speculative type from draft repo sidecars, + refactor mlock/mmap/directio into load-mode, and add -lm mlock + option. + * UI: symbolic math support via nerdamer in JS sandbox, fix + symbolic math tool prompt, fix system message edit box, detect + conversation import format from file contents, reduce per-token + render cost, remove render effects, fix MCP server display name + conflicts, add "Default" option for reasoning selector, and fix + context gauge regressions. @@ -171,2 +193,2 @@ - GLM-5.2-Vision, add GLM 5.2 Indexer support, fix Android build, use - align_corners for Qwen3-VL interpolation, and use RAII for + GLM-5.2-Vision, add GLM 5.2 Indexer support, fix Android build, + use align_corners for Qwen3-VL interpolation, and use RAII for @@ -175,2 +197,3 @@ - NomicBert GGUF conversion, HunyuanVL XD-RoPE config, use-after-free - when loading LoRA adapter fails, and various llama bugs. + NomicBert GGUF conversion, HunyuanVL XD-RoPE config, + use-after-free when loading LoRA adapter fails, and various + llama bugs. @@ -183,2 +206,2 @@ -- llama.cpp now enables SME, found in ARMv9.2. Only gcc>=16 knows about SME, - so require it on arm64. +- llama.cpp now enables SME, found in ARMv9.2. Only gcc>=16 knows + about SME, so require it on arm64. @@ -198,4 +221,8 @@ - * OpenCL: optimize Q4_K/Q5_K flat matrix-vector kernels by loading quants as uint - * OpenCL: improve Q4_K performance with transposed no-shuffle scales for coalesced reads - * OpenCL: optimize MoE DP4A activation tile access using 128-bit local memory transfers - * OpenCL: add broadcast support for Adreno MUL_MAT and honor view_offs + * OpenCL: optimize Q4_K/Q5_K flat matrix-vector kernels by + loading quants as uint + * OpenCL: improve Q4_K performance with transposed no-shuffle + scales for coalesced reads + * OpenCL: optimize MoE DP4A activation tile access using 128-bit + local memory transfers + * OpenCL: add broadcast support for Adreno MUL_MAT and honor + view_offs @@ -204 +231,2 @@ - * DFlash/Eagle3: automatically download Hugging Face sidecar files + * DFlash/Eagle3: automatically download Hugging Face sidecar + files @@ -206,2 +234,4 @@ - * Quantization: exclude DeepSeek V4 FFN routing table (ffn_gate_tid2eid) from quantization - * Vulkan: refactor queue handling to use per-instance mutexes and unique handles + * Quantization: exclude DeepSeek V4 FFN routing table + (ffn_gate_tid2eid) from quantization + * Vulkan: refactor queue handling to use per-instance mutexes and + unique handles @@ -211 +241,2 @@ - * UI: enable agentic flow when only the JavaScript sandbox is active + * UI: enable agentic flow when only the JavaScript sandbox is + active @@ -288,7 +319,9 @@ - * Server and API work added model-load progress reporting, improved - batch construction, and fixed several long-context and KV-cache - correctness issues. - * Vulkan gained several matmul, Flash Attention, and buffer-transfer - optimizations plus better memory-property handling, which should - mainly help prompt processing for quantized models. - * Additional multimodal, router, and WebUI updates landed upstream. + * Server and API work added model-load progress reporting, + improved batch construction, and fixed several long-context and + KV-cache correctness issues. + * Vulkan gained several matmul, Flash Attention, and + buffer-transfer optimizations plus better memory-property + handling, which should mainly help prompt processing for + quantized models. + * Additional multimodal, router, and WebUI updates landed + upstream. @@ -305,3 +338,3 @@ - * Improved accelerator backends with faster CUDA MTP and MoE paths, - SYCL multi-column MMVQ, Vulkan Intel FWHT support with safer - feature detection, OpenCL kernel tuning, and WASM SIMD128 + * Improved accelerator backends with faster CUDA MTP and MoE + paths, SYCL multi-column MMVQ, Vulkan Intel FWHT support with + safer feature detection, OpenCL kernel tuning, and WASM SIMD128 @@ -309,7 +342,8 @@ - * Extended server and API behavior with multimodal token counting, - a new `*/input_tokens` endpoint, reduced unnecessary checkpoint - restores, and disabled on-device speculative checkpoints. - * Updated build and packaging bits: CMake now skips CPU-only tools - when the CPU backend is disabled, XCFramework header and - module-map generation was fixed, Docker CUDA 13 moved to 13.3.0, - and MSYS CI dependencies were slimmed down. + * Extended server and API behavior with multimodal token + counting, a new `*/input_tokens` endpoint, reduced unnecessary + checkpoint restores, and disabled on-device speculative + checkpoints. + * Updated build and packaging bits: CMake now skips CPU-only + tools when the CPU backend is disabled, XCFramework header and + module-map generation was fixed, Docker CUDA 13 moved to + 13.3.0, and MSYS CI dependencies were slimmed down. @@ -325,4 +359,4 @@ - * Refreshed the WebUI with single-line reasoning preview, improved - keyboard accessibility, fixed tool selector state and counting, - pinned packages for reproducible installs, and enabled npm - install safeguards with `ignore-scripts=true`. + * Refreshed the WebUI with single-line reasoning preview, + improved keyboard accessibility, fixed tool selector state and + counting, pinned packages for reproducible installs, and + enabled npm install safeguards with `ignore-scripts=true`. @@ -349,3 +383,3 @@ - vision handling, including Gemma 4 unified fixes, - DeepSeek-OCR updates, Hunyuan OCR consolidation, granite-speech, - and MiniCPM-V 4.6 support. + vision handling, including Gemma 4 unified fixes, DeepSeek-OCR + updates, Hunyuan OCR consolidation, granite-speech, and + MiniCPM-V 4.6 support. @@ -387,2 +421,2 @@ - - Hardware Acceleration: Implemented 128-bit RISC-V Vector - (RVV) and SIMD kernels for enhanced CPU performance. + - Hardware Acceleration: Implemented 128-bit RISC-V Vector (RVV) + and SIMD kernels for enhanced CPU performance. @@ -412,3 +446,3 @@ - * 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. + * 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. @@ -419,2 +453,2 @@ - Metal, WebGPU, SYCL, CANN, HIP/ROCm, Hexagon, ZenDNN, and a - new VirtGPU backend for Virglrenderer API remoting. ++++ 410 more lines (skipped) ++++ between /work/SRC/openSUSE:Factory/llamacpp/llamacpp.changes ++++ and /work/SRC/openSUSE:Factory/.llamacpp.new.3698736/llamacpp.changes Old: ---- llama-b11146-ui.tar.gz llamacpp-0.5.0.tar.gz New: ---- llama-b11429-ui.tar.gz llamacpp-0.6.0.tar.gz ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ Other differences: ------------------ ++++++ llamacpp.spec ++++++ --- /var/tmp/diff_new_pack.yZWu7D/_old 2026-10-15 17:08:54.175659198 +0200 +++ /var/tmp/diff_new_pack.yZWu7D/_new 2026-10-15 17:08:54.177659281 +0200 @@ -18,7 +18,7 @@ %global backend_dir %{_libdir}/ggml -%global upstream_build 11146 +%global upstream_build 11429 %global license_dir %{_datadir}/licenses %global llama_sover %{version} @@ -27,7 +27,7 @@ %global mtmd_sover %{llama_sover} %global mtmd_sover_suffix 0 -%global ggml_sover 0.25.1 +%global ggml_sover 0.26.0 %global ggml_sover_suffix 0 %if 0%{?suse_version} == 1500 @@ -43,7 +43,7 @@ %endif Name: llamacpp -Version: 0.5.0 +Version: 0.6.0 Release: 0 Summary: Inference of Meta's LLaMA model (and others) in pure C/C++ License: Apache-2.0 AND MIT AND BSD-2-Clause AND BSD-3-Clause AND ISC AND MPL-2.0 ++++++ llama-b11146-ui.tar.gz -> llama-b11429-ui.tar.gz ++++++ ++++ 7613 lines of diff (skipped) ++++++ llamacpp-0.5.0.tar.gz -> llamacpp-0.6.0.tar.gz ++++++ /work/SRC/openSUSE:Factory/llamacpp/llamacpp-0.5.0.tar.gz /work/SRC/openSUSE:Factory/.llamacpp.new.3698736/llamacpp-0.6.0.tar.gz differ: char 22, line 1
