llama.cpp 0.2.0 release notes

Overview

This version marks the beginning of consistent semantic versioning for llama.cpp. The main goal is to make the job of downstream projects and users simpler when deciding which version of llama.cpp/ggml to distribute and use. The new version tags of the format vX.Y.Z will mark "stable" versions of llama.cpp. The usual b[NUM] tags remain as before - created on almost every commit to the master branch. The b tags can be considered as "nightly" or "dev" versions with the latest functionality available, though these could be more unstable.

To summarize:

tag vX.Y.Z - stable, slower release cadence, recommended for downstream distribution and casual users
tag b[NUM] - bleeding edge, faster release cadence, recommended for developers and technical users

More info: dist : releases and versioning of ggml-org projects

Assets

Nightly build: b10566

Web UI: the nightly-tag.txt asset contains the tag of the corresponding nightly release

Changelog since v0.1.2

bb4caa754 llama.cpp : bump version to 0.2.0 (#27498) c4b0225d8 scripts : add release.sh for release preparation (#27497) 5de25a748 sync : ggml 01ff204fb ggml : bump version to 0.21.0 (ggml/1597) 353b32d8b ci : remove duplicate flag (#27488) 7a0e42fd0 Revert "sycl : add Q2_K reordered MMVQ and ESIMD kernels (#26336)" (#27486) 5b6ddc967 ui: Settings navigation cleanup (#27241) e467c2ff6 ci : add nightly-tag.txt to make-release (#27485) 171974745 ci : release clean-up (#27477) 62b226906 kleidiai : add SME2 F32 GEMV kernel support (#26891) ff14356e0 sycl : add Q2_K reordered MMVQ and ESIMD kernels (#26336) 5fff12845 test : make the FA V-is-view-of-K case a test case parameter (#27394) 9e89a196b sycl : Add Q5_K ESIMD kernel (#26376) cd26896c1 opencl: keep the vocab-scale K-quant lm_head on the CPU for Adreno A7X (compiler issue workaround) (#26440) 1cb3f5eb4 sycl: Update gate logic for Alchemist GPUs regarding OneDNN features. (#26635) 6602dd338 sycl: fix multiple warnings in compiling sycl backend (#26713) 9e96cf77f sycl : fix load model with mlock issue (#27250) b2e5e9b28 TP: enable tensor split for LFM2/LFM2MOE (#26993) a298422da docs: fix typos in ET.md (#27457) 749f688fc ggml: support ggml_rope_set_offset on opencl, sycl, wgpu, hexagon (#27345) 0e1d9185c ci: use shell script to check cmake pkg (#27414) a30273376 metal : clamp K extent in tensor API mat-mat kernel for K not a multiple of 32 (#27450) 6503355df opencl: fix q6_K flat mul_mat for Adreno A6x/A7x GPUs with older E031 compilers (#26476) 6b4fa88a6 opencl: fix local size for norm (#27339) 521a64cd0 ui: Stores split refactor (#27240) 681c29d36 mtmd: add --mmproj-device argument (#23255) 07822bddf model : support DSpark for LFM2 models (#27383) 78ec4c378 vulkan: FA MMQ should use fp32 for Q quantization calculations (#27413) 63b64a50a metal : dequant kv cache only for large batches (#27438) bf0040e15 CI: Use LLVM's OpenMP over MSVC_DEBUG_non_redist on Windows (#26678) 9855ad69d server: (router) lazy-load startup_models after main setup (#27424) 8a832e4bf server : fix --docker-repo being treated as router mode (#27416) 2b5621094 CUDA: adding switch points per HW and quant type to tune the mvq->MMQ decode crossover (#26079) dc64a1620 common : gracefully fallback on unsupported regex patterns in JSON schema (#26939) 70aff2525 metal : dequantize quantized KV to F16 before flash attention (#27390) f20395dae Revert "tensor-split meta backend fixes (#26502)" (#27433) 849798132 ggml: fix backend split scheduler race condition (#26040) a3b1effcd convert: fix get block count error for Nemotron 3 Ultra (#27101) d9b6be07d ggml-cuda: provide static workspace for cuBLAS handles (#26574) 929d47a39 graph : create V as a view of K in the k_iswa build_attn (#27392) f466cfa38 spec : avoid binding reference to null pointer (#27404) 2cfdb5fc0 vulkan : add source groups for shaders (#26666) 9ee9fc04c opencl: make the MoE expert scatter deterministic (#26464) d59d455fd tensor-split meta backend fixes (#26502) 990e3bfee hexagon: fix FA HMX queue ordering and pack the rescale D matrices (#27042) b062ba735 opencl: port fused ssm_scan kernel (Mamba-2, d_state in {128, 256}) to GPU (#26439) cd644c395 ggml-cpu: gate __fp16 on __ARM_FP16_FORMAT_IEEE (#26860) 947fd9bb2 server: refactor sleep handling, allow access /metrics during sleep (#27376) ee0ea03ad server : make models endpoints private when authentication is enabled (#26347) dc72703fc vulkan : dequant q8_0 KV once in coopmat1 (#25494) b95502ba9 vulkan: add null checks in ggml_vk_queue_command_pools_cleanup (#27353) 3e7344670 Revert "common: share thread pools when n_threads differ (#27138)" (#27337) 7221e24f5 model : GraniteSWAForCausalLM / GraniteMoeSWAForCausalLM (#25505) 6cc504a2e sycl: report zero devices instead of aborting when the host has none (#27291) 01ac3ad76 ci : add release attestation url (#27389) 2e92ecd02 models : remove duplicate metadata load (#27378) 645ca2834 ci : re-enable release dependency for sycl (#27385) fe8156f78 ggml: add ggml_rope_set_offset (+ metal support) (#27120) 77acca437 ui: read persisted settings before the API key probe (#27365) 95c409c13 mtmd: add mtmd_bitmap_set_mergeable (#27348) 8ef78e644 metal : dequantize q8_0 using packed types (#27370) ee4c505a4 server: add dedup-cache-models preset option (#27346) 98d1e92c2 vulkan: tiled transpose for 0<->2 permuted CONT (#26585) 5112b9738 ggml-webgpu: add mulmat with overlapping src0/src1 (e.g., for minimax-01) (#27321) 0adcc3bb5 ci : add attestation for signed release artifacts (#25933) 79fe79919 tests: skip test-unicode build on win32/BUILD_SHARED_LIBS (#27336) 0329fcdac gguf-py : add size guards to GGUFReader (#27188) 6d0549831 server: (cosmetic) do not print cmd_child_to_router messages [no release] (#27347) 60addddf3 opencl: fix WAR race in the generic FA tile kernels when the WG spans subgroups (#26434) 9731ad3f2 mtmd: deepseek-ocr SAM ggml_conv_2d with the im2col kept in F32 (#26727) af5172627 RPC: populate use_count to enable fusion inside backends (#27142) 3dc7285b4 ui: Services consolidation refactor (#27239) fdf4c6460 ui: Stores consolidation refactor (#27238) 04b569142 common: share thread pools when n_threads differ (#27138) 059670428 quant : Optimise memory usage by evicting weights after processing each layer (#22877) 0882c7bc8 sycl: honor GGML_HINT_SRC0_IS_HADAMARD (#27298) afd439df1 unicode : include '~' in collapsed symbol class (#26972) 169e4a7ff readme : update status badges + regen AUTHORS (#27317) 9d77fa172 ci : Update OpenVINO to 2026.3, skip nemotron-h rollback test (#27292) c0296022f ci: add Windows ARM64 CUDA support to the manual workflow (#27300) 7acdbb1f1 mtmd: fix LFM2 image tiling threshold (#27057)