Commit Graph

  • 591072fcc8 opencl: refactor some ops, concat, repeat, tanh and scale (llama/19226) lhez 2026-02-02 15:54:43 -08:00
  • 871063016d ggml-cpu: FA split across kv for faster TG (llama/19209) Aman Gupta 2026-02-03 01:19:55 +08:00
  • c4003da2b8 Remove support for Nvidia & AMD GPU, because the oneAPI plugin for Nvidia & AMD GPU is unavailable: download/installation channels are out of work. (llama/19246) Neo Zhang 2026-02-02 21:06:21 +08:00
  • 74353e90a1 sycl: implement GGML_OP_TOP_K (llama/19242) Tamar 2026-02-02 15:05:51 +02:00
  • 73e04555eb metal : support virtual devices (llama/18919) Georgi Gerganov 2026-02-02 14:29:44 +02:00
  • 625c8d863e ggml-backend: fix async set/get fallback sync (llama/19179) Johannes Gäßler 2026-02-02 10:00:05 +01:00
  • 0e219ebf89 docs : Minor cleanups (llama/19252) Christian Kastner 2026-02-02 07:38:55 +01:00
  • a0256b8159 Remove pipeline cache mutexes (llama/19195) Nikhil Jain 2026-02-01 18:47:29 -08:00
  • aca5953d8d Bump cmake max version (needed for Windows on Snapdragon builds) (llama/19188) Max Krasnyansky 2026-02-01 14:13:38 -08:00
  • 9b927dd849 ggml-hexagon: flash-attention and reduce-sum optimizations (llama/19141) nullname 2026-01-31 13:14:20 +08:00
  • db9c88744d opencl: add optimized q8_0 mm kernel for adreno (llama/18871) shaofeiqi 2026-01-30 10:19:27 -08:00
  • efd6344939 Correctly fetch q8_1 quantize pipeline in test as needed by 8a3519b (llama/19194) Simon Redman 2026-01-30 11:27:16 -05:00
  • 06e3750407 ggml : bump version to 0.9.6 (ggml/1423) Georgi Gerganov 2026-02-07 09:58:02 +02:00
  • fc1a3e579e cmake : remove unused file (ggml/1419) Georgi Gerganov 2026-01-30 16:29:51 +02:00
  • 941bdabbe4 ruby : add Whisper::Context::Params, fix token memory management (#3647) KITAITI Makoto 2026-02-04 20:33:09 +09:00
  • aa1bc0d1a6 ruby : add VAD::Context#segments_from_samples, allow Pathname, etc. (#3633) KITAITI Makoto 2026-01-30 22:59:36 +09:00
  • bf422cb704 scripts : Fix dSYMs path case for macOS xcframework build (#3630) Frieder Bluemle 2026-01-30 05:57:26 -08:00
  • acbace0571 cuda : fix compile warnings (#0) Georgi Gerganov 2026-01-30 15:56:15 +02:00
  • 953e503fd9 talk-llama : sync llama.cpp Georgi Gerganov 2026-01-30 13:50:58 +02:00
  • b529c0610f sync : ggml Georgi Gerganov 2026-01-30 13:50:43 +02:00
  • 5dca0db99c add tensor type checking as part of cuda graph properties (llama/19186) bssrdf 2026-01-29 23:57:52 -05:00
  • 2a16e7a67f sycl: implement GGML_UNARY_OP_SOFTPLUS (llama/19114) s8322 2026-01-30 06:01:38 +02:00
  • 1b3c27efae sycl: implement GGML_OP_TRI (llama/19089) RachelMantel 2026-01-30 06:00:49 +02:00
  • 829e70044b ggml-webgpu: improve flastAttention performance by software pipelining (llama/19151) Zheyuan Chen 2026-01-29 14:05:30 -08:00
  • 2a89a3f35c hexagon: enable offloading to Hexagon on Windows on Snapdragon (llama/19150) Todor Boinovski 2026-01-29 12:33:21 -08:00
  • b997e690ef cuda : fix nkvo, offload and cuda graph node properties matching (llama/19165) Georgi Gerganov 2026-01-29 18:45:30 +02:00
  • 34a3e28a08 HIP: add mmf for CDNA (llama/18896) yulo 2026-01-29 18:10:53 +08:00
  • e0a2182970 ggml-zendnn : resolve ZenDNN backend cross-module symbol dependency (llama/19159) Vishal Singh 2026-01-29 09:58:57 +05:30
  • 62ba8b537f CUDA: refactor topk-moe to enable more models (GLM 4.7, Nemotron etc.) (llama/19126) Aman Gupta 2026-01-29 10:31:28 +08:00
  • f0e85bb142 sycl: fix norm kernels: l2_norm, group_norm, rms_norm by remove assert to support more cases (llama/19154) Neo Zhang 2026-01-29 09:20:22 +08:00
  • 33148bb523 Vulkan Flash Attention Coopmat1 Refactor (llama/19075) Ruben Ortlam 2026-01-28 18:52:45 +01:00
  • cc0c103b5d ggml-sycl: remove unused syclcompat header (llama/19140) Patryk Kaminski 2026-01-28 16:33:54 +01:00
  • dda7d9cd1c vulkan: handle device dedup on MacOS + Vega II Duo cards (llama/19058) Oleksandr Kuvshynov 2026-01-28 06:35:54 -05:00
  • 531d7b6781 ggml: new backend for Virglrenderer API Remoting acceleration (v2) (llama/18718) Kevin Pouget 2026-01-28 10:49:40 +01:00
  • 3701413a71 ggml-cpu: arm64: Q4_K scale unroll and vectorization (llama/19108) Alberto Cabrera Pérez 2026-01-28 07:15:56 +00:00
  • 7fb0f823de cuda : fix "V is K view" check for non-unified KV cache (llama/19145) Georgi Gerganov 2026-01-28 09:15:27 +02:00
  • f28a733025 CUDA: tune GLM 4.7 Flash FA kernel selection logic (DGX Spark) (llama/19142) Georgi Gerganov 2026-01-28 09:15:11 +02:00
  • dfdd2fee83 ggml webgpu: Split shared state (webgpu_context) into global state and per-thread state (llama/18976) Nikhil Jain 2026-01-27 20:53:36 -08:00
  • 9c75c793a6 ggml-zendnn : update ZenDNN git tag to main branch (llama/19133) Vishal Singh 2026-01-28 03:51:36 +05:30
  • 9d94d0f782 CUDA: tune GLM 4.7 Flash FA kernel selection logic (llama/19097) Johannes Gäßler 2026-01-27 14:28:56 +01:00
  • 00885e08e2 ggml-cpu: aarm64: q6_K repack gemm and gemv (and generic) implementations (i8mm) #18860 (llama/18888) Alberto Cabrera Pérez 2026-01-27 09:08:10 +00:00
  • 5fcbbdc0dd Reduce CPU-side stalls due to the CUDA command buffer being full (llama/19042) Gaurav Garg 2026-01-27 06:52:44 +00:00
  • b2e2032856 ggml-cpu: Enable FP16 MMA kernels on PPC (llama/19060) shalinib-ibm 2026-01-27 09:22:34 +05:30
  • 56f82a9f33 opencl: add flattened q6_K mv (llama/19054) lhez 2026-01-30 10:34:38 +02:00
  • 41d5d7bb0e CUDA: fix padding of GQA to power of 2 in FA (llama/19115) Johannes Gäßler 2026-01-26 23:24:58 +01:00
  • f63848eada CUDA: faster FA for GQA > 1 but not power of 2 (llama/19092) Johannes Gäßler 2026-01-25 21:19:47 +01:00
  • 4372b87b8e metal : fix recommendedMaxWorkingSetSize availability on legacy iOS/macOS (llama/19088) ccbinn 2026-01-26 02:07:19 +08:00
  • 1642a4fb60 ggml-cpu: Use tiled FA for prompt-processing (llama/19012) Aman Gupta 2026-01-25 23:25:58 +08:00
  • d2b51404e4 kv-cache : support V-less cache (llama/19067) Georgi Gerganov 2026-01-25 15:48:56 +02:00
  • f53eafd745 CUDA: re-use MLA K data for V in MMA FA (llama/19057) Johannes Gäßler 2026-01-24 10:09:36 +01:00
  • 13577a6ce4 ggml-cuda: enable cuda-graphs for n-cpu-moe (llama/18934) Aman Gupta 2026-01-24 14:25:20 +08:00
  • 79f1bb3d35 ggml-hexagon: flash-attn opt (llama/19025) nullname 2026-01-24 14:02:07 +08:00
  • 0d9dda5a99 use malloc to support both iGPU and dGPU in same time (llama/18992) Neo Zhang 2026-01-23 20:54:10 +08:00
  • e090d91f5e ggml-cpu: aarm64: q5_K repack gemm and gemv (and generic) implementations (i8mm) (llama/18860) Alberto Cabrera Pérez 2026-01-23 07:55:08 +00:00
  • 3f96a1da0e mla : make the V tensor a view of K (llama/18986) Georgi Gerganov 2026-01-22 22:09:01 +02:00
  • f21d0cbb1a CUDA: fix alignment check for FA (llama/19023) Johannes Gäßler 2026-01-22 20:39:25 +01:00
  • 0e030b852a opencl: enable the general fp mm for non-cont input and as a fallback for specialized kqv kernel for adreno (llama/18970) lhez 2026-01-22 10:29:25 -08:00
  • d4fafcfc6f CUDA: add gqa_ratio 4 for GLM 4.7 flash (llama/18953) Aman Gupta 2026-01-22 18:51:53 +08:00
  • 167fec69d5 opencl: add TRI op support (llama/18979) shaofeiqi 2026-01-21 22:05:54 -08:00
  • 55927d42ef ggml-zdnn : mark zDNN buffers as non-host (llama/18967) Aleksei Nikiforov 2026-01-22 01:16:21 +01:00
  • b7e323f40b vulkan: Remove transfer_ctx, do everything in compute_ctx. (llama/18945) Jeff Bolz 2026-01-21 11:01:40 -06:00
  • b2bc4d810b vulkan: support flash attention GQA/split_k with small batches (llama/18938) Jeff Bolz 2026-01-21 10:43:43 -06:00
  • 3bbf4ced47 Revert "vulkan: force full subgroups for flash attention to fix intel subgroup crash (#17356)" (llama/18831) Masato Nakasaka 2026-01-22 01:13:43 +09:00
  • 660d943ff8 vulkan: Use mul_mat_vec_id for small values of n (llama/18918) Jeff Bolz 2026-01-21 09:22:02 -06:00
  • 924a9e292c CUDA: Fix builds for older CCCL versions by ifdefing strided_iterator (llama/18964) Oliver Simons 2026-01-21 02:34:29 +01:00
  • fdc83ee3c0 CUDA: Replace init_offsets kernel with iterators in cub-based argsort (llama/18930) Oliver Simons 2026-01-20 13:11:01 +01:00
  • bf71ffa6b3 ggml : cleanup path_str() (llama/18928) Adrien Gallouët 2026-01-20 11:42:49 +01:00
  • b0517d6912 metal : enable FA for MLA heads (llama/18950) Georgi Gerganov 2026-01-20 12:21:28 +02:00
  • 47f3e3b927 ggml : add ggml_build_forward_select (llama/18550) Georgi Gerganov 2026-01-19 20:03:19 +02:00
  • 62a09b106d opencl: fix q6_K mv for m=1 (llama/18893) lhez 2026-01-17 13:50:32 -08:00
  • 389dafc7c2 ggml webgpu: support for backend sampling (llama/18880) Reese Levine 2026-01-30 10:32:34 +02:00
  • 511ca7a1f4 ggml : extend ggml_pool_1d + metal (llama/16429) Thore Koritzius 2026-01-16 15:59:56 +01:00
  • ecb4b80c35 ggml-blas: hide warnings from included BLAS headers (llama/18818) Perry Naseck 2026-01-16 06:38:25 -05:00
  • 42960b6073 CANN: Remove unused ggml_cann_get_device function (llama/18625) Raul Torres 2026-01-16 08:34:09 +00:00
  • 2fceb5a80f CANN: fix an issue where get_env was not fully renamed (llama/18796) Chenguang Li 2026-01-16 16:24:04 +08:00
  • 854274a297 CANN: support gated linear attn (llama/18653) hipudding 2026-01-16 16:18:49 +08:00
  • ed6004d051 OpenCL: add SOLVE_TRI op support (llama/18846) shaofeiqi 2026-01-15 11:17:17 -08:00
  • 290ff3d28d cuda : print less debug logs when disabling cuda graphs (llama/18868) Georgi Gerganov 2026-01-15 20:53:01 +02:00
  • f2f0ba0384 CUDA: fix allignment on register spill for FA (llama/18815) Johannes Gäßler 2026-01-15 15:14:50 +01:00
  • 78a23d4830 ggml-cpu: optimize ggml_vec_dot_bf16 for Power9 (llama/18837) shalinib-ibm 2026-01-15 15:01:18 +05:30
  • 50b7ab3d46 hexagon: support for OP_CPY, host buffers now optional (llama/18822) Max Krasnyansky 2026-01-30 10:28:03 +02:00
  • bc09047405 CUDA: Factor out and re-use block_reduce function (llama/18785) Oliver Simons 2026-01-15 03:44:54 +01:00
  • 4b155e9bfb vulkan: Check maxStorageBufferRange in supports_op (llama/18709) Jeff Bolz 2026-01-14 03:59:05 -06:00
  • 25aeb66a4a CUDA : fix typo in clang pragma comment [no ci] (llama/18830) Daniel Bevenius 2026-01-14 10:31:49 +01:00
  • 49762e8fb3 vulkan: work around Intel fp16 bug in mmq (llama/18814) Ruben Ortlam 2026-01-14 09:41:23 +01:00
  • 17656e56dc ggml-metal: do not copy headers for embedded, use current binary dir for embedded (llama/18705) Perry Naseck 2026-01-14 02:22:25 -05:00
  • c6a495ae5d HIP: add fattn-mma-f16 for RDNA4 (llama/18481) yulo 2026-01-13 20:52:16 +08:00
  • 7aa8818647 examples : use -dev/--device and WHISPER_ARG_DEVICE (#3557) Bráulio Oliveira 2026-01-21 04:40:30 -03:00
  • f53dc74843 whisper : Fix UTF-8 character boundary issue in segment wrapping (max_len) (#3592) Yshtola 2026-01-16 20:16:05 +08:00
  • 2eeeba56e9 release : v1.8.3 v1.8.3 Georgi Gerganov 2026-01-15 11:54:31 +02:00
  • 21c1765fcb benches : update Georgi Gerganov 2026-01-15 11:53:09 +02:00
  • 47af2fb70f sync : ggml Georgi Gerganov 2026-01-13 19:11:04 +02:00
  • 6ee0eaf531 CUDA : fix unused argument when USE_CUDA_GRAPH=OFF (llama/18800) Georgi Gerganov 2026-01-13 12:25:53 +02:00
  • ab1828dc1c vulkan: change memory_logger to be controlled by an env var (llama/18769) Jeff Bolz 2026-01-12 06:32:55 -06:00
  • aedf332ec5 vulkan: Use VK_EXT_shader_64bit_indexing to handle large mat_mul(_id) (llama/18678) Jeff Bolz 2026-01-12 05:32:13 -06:00
  • 716d68aca9 vulkan: Disable large coopmat matmul configuration on proprietary AMD driver (llama/18763) Ruben Ortlam 2026-01-12 07:29:35 +01:00
  • c0433783c3 Vulkan: Optimize Matmul parameters for AMD GPUs with Coopmat support (llama/18749) Ruben Ortlam 2026-01-11 17:33:33 +01:00
  • ecfcc65fbf talk-llama : sync llama.cpp Georgi Gerganov 2026-01-12 14:48:26 +02:00
  • 13dc9a912b sync : ggml Georgi Gerganov 2026-01-12 14:44:38 +02:00
  • d4ce2e554f opencl: add SOFTPLUS op support (llama/18726) shaofeiqi 2026-01-10 21:57:44 -08:00