v0.21.0

Overview

New version has been released.

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

Changelog since v0.20.2

8599e0e sync : llama.cpp
9a538ea Revert "sycl : add Q2_K reordered MMVQ and ESIMD kernels (llama/26336)" (llama/27486)
28d7ab8 ci : extract build job into separate release workflow (#1598)
91147d0 ggml : bump version to 0.21.0 (#1597)
19a3858 scripts : restore release.sh preparation script (#1596)
bb83ce0 ci : replace release workflow with make-release (#1595)
5da4cf1 ci : split self-hosted jobs into separate workflow (#1594)
33be41c sync : llama.cpp
451e766 kleidiai : add SME2 F32 GEMV kernel support (llama/26891)
59bbab9 sycl : add Q2_K reordered MMVQ and ESIMD kernels (llama/26336)
5a232ab test : make the FA V-is-view-of-K case a test case parameter (llama/27394)
14dfda3 sycl : Add Q5_K ESIMD kernel (llama/26376)
df8336e opencl: keep the vocab-scale K-quant lm_head on the CPU for Adreno A7X (compiler issue workaround) (llama/26440)
9969469 sycl: Update gate logic for Alchemist GPUs regarding OneDNN features. (llama/26635)
1d30b1b sycl: fix multiple warnings in compiling sycl backend (llama/26713)
7378ac1 sycl : fix load model with mlock issue (llama/27250)
4979ee2 ggml: support ggml_rope_set_offset on opencl, sycl, wgpu, hexagon (llama/27345)
33c9ea5 metal : clamp K extent in tensor API mat-mat kernel for K not a multiple of 32 (llama/27450)
9e7a4c2 opencl: fix q6_K flat mul_mat for Adreno A6x/A7x GPUs with older E031 compilers (llama/26476)
505842b opencl: fix local size for norm (llama/27339)
b0d45de vulkan: FA MMQ should use fp32 for Q quantization calculations (llama/27413)
3821f6e metal : dequant kv cache only for large batches (llama/27438)
7159fc6 CI: Use LLVM's OpenMP over MSVC_DEBUG_non_redist on Windows (llama/26678)
02a0ab2 CUDA: adding switch points per HW and quant type to tune the mvq->MMQ decode crossover (llama/26079)
e4159dd metal : dequantize quantized KV to F16 before flash attention (llama/27390)
9a1d234 Revert "tensor-split meta backend fixes (#26502)" (llama/27433)
0ab579a ggml: fix backend split scheduler race condition (llama/26040)
ee401c8 ggml-cuda: provide static workspace for cuBLAS handles (llama/26574)
6a6a4c1 vulkan : add source groups for shaders (llama/26666)
e3e8dc1 opencl: make the MoE expert scatter deterministic (llama/26464)
7c6058c tensor-split meta backend fixes (llama/26502)
37c79ff hexagon: fix FA HMX queue ordering and pack the rescale D matrices (llama/27042)
585e463 opencl: port fused ssm_scan kernel (Mamba-2, d_state in {128, 256}) to GPU (llama/26439)
ff4f3e4 ggml-cpu: gate __fp16 on __ARM_FP16_FORMAT_IEEE (llama/26860)
480cc77 vulkan : dequant q8_0 KV once in coopmat1 (llama/25494)
8fc9724 vulkan: add null checks in ggml_vk_queue_command_pools_cleanup (llama/27353)
f673848 sycl: report zero devices instead of aborting when the host has none (llama/27291)
aa9d222 ggml: add ggml_rope_set_offset (+ metal support) (llama/27120)
868403c metal : dequantize q8_0 using packed types (llama/27370)
55e4b3b vulkan: tiled transpose for 0<->2 permuted CONT (llama/26585)
52b66a4 ggml-webgpu: add mulmat with overlapping src0/src1 (e.g., for minimax-01) (llama/27321)
981a41b opencl: fix WAR race in the generic FA tile kernels when the WG spans subgroups (llama/26434)
f16a1a2 RPC: populate use_count to enable fusion inside backends (llama/27142)
b74262d sycl: honor GGML_HINT_SRC0_IS_HADAMARD (llama/27298)