Overview
New version has been released.
Nightly build: b10566
Web UI: the nightly-tag.txt asset contains the tag of the corresponding nightly release
More info: dist : releases and versioning of ggml-org projects
Changelog since v0.1.2
bb4caa7 llama.cpp : bump version to 0.2.0 (#27498)
c4b0225 scripts : add release.sh for release preparation (#27497)
5de25a7 sync : ggml
01ff204 ggml : bump version to 0.21.0 (ggml/1597)
353b32d ci : remove duplicate flag (#27488)
7a0e42f Revert "sycl : add Q2_K reordered MMVQ and ESIMD kernels (#26336)" (#27486)
5b6ddc9 ui: Settings navigation cleanup (#27241)
e467c2f ci : add nightly-tag.txt to make-release (#27485)
1719747 ci : release clean-up (#27477)
62b2269 kleidiai : add SME2 F32 GEMV kernel support (#26891)
ff14356 sycl : add Q2_K reordered MMVQ and ESIMD kernels (#26336)
5fff128 test : make the FA V-is-view-of-K case a test case parameter (#27394)
9e89a19 sycl : Add Q5_K ESIMD kernel (#26376)
cd26896 opencl: keep the vocab-scale K-quant lm_head on the CPU for Adreno A7X (compiler issue workaround) (#26440)
1cb3f5e sycl: Update gate logic for Alchemist GPUs regarding OneDNN features. (#26635)
6602dd3 sycl: fix multiple warnings in compiling sycl backend (#26713)
9e96cf7 sycl : fix load model with mlock issue (#27250)
b2e5e9b TP: enable tensor split for LFM2/LFM2MOE (#26993)
a298422 docs: fix typos in ET.md (#27457)
749f688 ggml: support ggml_rope_set_offset on opencl, sycl, wgpu, hexagon (#27345)
0e1d918 ci: use shell script to check cmake pkg (#27414)
a302733 metal : clamp K extent in tensor API mat-mat kernel for K not a multiple of 32 (#27450)
6503355 opencl: fix q6_K flat mul_mat for Adreno A6x/A7x GPUs with older E031 compilers (#26476)
6b4fa88 opencl: fix local size for norm (#27339)
521a64c ui: Stores split refactor (#27240)
681c29d mtmd: add --mmproj-device argument (#23255)
07822bd model : support DSpark for LFM2 models (#27383)
78ec4c3 vulkan: FA MMQ should use fp32 for Q quantization calculations (#27413)
63b64a5 metal : dequant kv cache only for large batches (#27438)
bf0040e CI: Use LLVM's OpenMP over MSVC_DEBUG_non_redist on Windows (#26678)
9855ad6 server: (router) lazy-load startup_models after main setup (#27424)
8a832e4 server : fix --docker-repo being treated as router mode (#27416)
2b56210 CUDA: adding switch points per HW and quant type to tune the mvq->MMQ decode crossover (#26079)
dc64a16 common : gracefully fallback on unsupported regex patterns in JSON schema (#26939)
70aff25 metal : dequantize quantized KV to F16 before flash attention (#27390)
f20395d Revert "tensor-split meta backend fixes (#26502)" (#27433)
8497981 ggml: fix backend split scheduler race condition (#26040)
a3b1eff convert: fix get block count error for Nemotron 3 Ultra (#27101)
d9b6be0 ggml-cuda: provide static workspace for cuBLAS handles (#26574)
929d47a graph : create V as a view of K in the k_iswa build_attn (#27392)
f466cfa spec : avoid binding reference to null pointer (#27404)
2cfdb5f vulkan : add source groups for shaders (#26666)
9ee9fc0 opencl: make the MoE expert scatter deterministic (#26464)
d59d455 tensor-split meta backend fixes (#26502)
990e3bf hexagon: fix FA HMX queue ordering and pack the rescale D matrices (#27042)
b062ba7 opencl: port fused ssm_scan kernel (Mamba-2, d_state in {128, 256}) to GPU (#26439)
cd644c3 ggml-cpu: gate __fp16 on __ARM_FP16_FORMAT_IEEE (#26860)
947fd9b server: refactor sleep handling, allow access /metrics during sleep (#27376)
ee0ea03 server : make models endpoints private when authentication is enabled (#26347)
dc72703 vulkan : dequant q8_0 KV once in coopmat1 (#25494)
b95502b vulkan: add null checks in ggml_vk_queue_command_pools_cleanup (#27353)
3e73446 Revert "common: share thread pools when n_threads differ (#27138)" (#27337)
7221e24 model : GraniteSWAForCausalLM / GraniteMoeSWAForCausalLM (#25505)
6cc504a sycl: report zero devices instead of aborting when the host has none (#27291)
01ac3ad ci : add release attestation url (#27389)
2e92ecd models : remove duplicate metadata load (#27378)
645ca28 ci : re-enable release dependency for sycl (#27385)
fe8156f ggml: add ggml_rope_set_offset (+ metal support) (#27120)
77acca4 ui: read persisted settings before the API key probe (#27365)
95c409c mtmd: add mtmd_bitmap_set_mergeable (#27348)
8ef78e6 metal : dequantize q8_0 using packed types (#27370)
ee4c505 server: add dedup-cache-models preset option (#27346)
98d1e92 vulkan: tiled transpose for 0<->2 permuted CONT (#26585)
5112b97 ggml-webgpu: add mulmat with overlapping src0/src1 (e.g., for minimax-01) (#27321)
0adcc3b ci : add attestation for signed release artifacts (#25933)
79fe799 tests: skip test-unicode build on win32/BUILD_SHARED_LIBS (#27336)
0329fcd gguf-py : add size guards to GGUFReader (#27188)
6d05498 server: (cosmetic) do not print cmd_child_to_router messages [no release] (#27347)
60adddd opencl: fix WAR race in the generic FA tile kernels when the WG spans subgroups (#26434)
9731ad3 mtmd: deepseek-ocr SAM ggml_conv_2d with the im2col kept in F32 (#26727)
af51726 RPC: populate use_count to enable fusion inside backends (#27142)
3dc7285 ui: Services consolidation refactor (#27239)
fdf4c64 ui: Stores consolidation refactor (#27238)
04b5691 common: share thread pools when n_threads differ (#27138)
0596704 quant : Optimise memory usage by evicting weights after processing each layer (#22877)
0882c7b sycl: honor GGML_HINT_SRC0_IS_HADAMARD (#27298)
afd439d unicode : include '~' in collapsed symbol class (#26972)
169e4a7 readme : update status badges + regen AUTHORS (#27317)
9d77fa1 ci : Update OpenVINO to 2026.3, skip nemotron-h rollback test (#27292)
c029602 ci: add Windows ARM64 CUDA support to the manual workflow (#27300)
7acdbb1 mtmd: fix LFM2 image tiling threshold (#27057)