Commit Graph

  • b5fb9b9f58 CUDA: faster tile FA, add oob checks, more HSs (llama/16492) Johannes Gäßler 2025-10-11 20:54:32 +02:00
  • a91dd3be72 release : v1.8.1 v1.8.1 Georgi Gerganov 2025-10-12 11:17:59 +03:00
  • ea174c62bc bench : update [no ci] Georgi Gerganov 2025-10-12 08:47:48 +03:00
  • ff4c1a5a53 talk-llama : sync llama.cpp Georgi Gerganov 2025-10-12 08:37:14 +03:00
  • ed6a3063ec sync : ggml Georgi Gerganov 2025-10-12 08:36:34 +03:00
  • d201705e71 metal : fix mul-mm condition + fix mul-mv permuted kernels (llama/16494) Georgi Gerganov 2025-10-11 16:54:10 +03:00
  • 1cc342427b cuda : avoid initializing unused devices (llama/16510) Diego Devesa 2025-10-11 04:02:26 -07:00
  • d8f1aa4e1d cmake : Dont define XOPENSOURCE on AIX (llama/16481) Prajwal B Mehendarkar 2025-10-10 13:45:46 +05:30
  • d83fef35df cpu : optimize the ggml NORM operation (llama/15953) duduta 2025-10-09 22:11:15 +03:00
  • b9eac9419c CANN: Improve ACL graph matching (llama/16166) Chenguang Li 2025-10-09 15:50:25 +08:00
  • c8b2c56fd2 kleidiai: kernel interface refactoring (llama/16460) Charles Xu 2025-10-09 09:29:17 +02:00
  • 7df6766b63 refactor soft_max, add soft_max_back (llama/16472) Neo Zhang Jianyu 2025-10-09 15:25:11 +08:00
  • 21e6e72a2f Disable CUDA host buffers on integrated GPUs (llama/16308) ai-fonsi 2025-10-08 20:21:46 +02:00
  • 7ef78a72e1 metal : mark FA blocks (llama/16372) Georgi Gerganov 2025-10-08 10:57:53 +03:00
  • 4eea3efc49 ggml webgpu: profiling, CI updates, reworking of command submission (llama/16452) Reese Levine 2025-10-07 13:48:56 -07:00
  • 4bce4fa5e9 metal : add support for non-padded FA KV (llama/16148) Georgi Gerganov 2025-10-07 08:23:30 +03:00
  • 6cf0c21b09 tests : add -INF blocks to the KQ mask in the FA tests (llama/16380) Georgi Gerganov 2025-10-07 08:22:35 +03:00
  • 1a4116f942 metal : various optimizations + refactoring (llama/16446) Georgi Gerganov 2025-10-07 08:21:40 +03:00
  • 0e431b3cea ggml : fix unaligned access in AMX code (llama/16315) Georgi Gerganov 2025-10-06 16:05:27 +03:00
  • 0f29d7c3fa ggml-cpu : fix leftover handling in ggml_vec_scale_f32 for SVE (llama/16443) Daniel Bevenius 2025-10-06 14:17:12 +02:00
  • b8bdf06182 ggml webgpu: actually add softmax, fix rms_norm offset (llama/16400) Reese Levine 2025-10-04 20:59:31 -07:00
  • 2ca8fa37fa vulkan: use a more appropriate amount of threads when generating shaders (llama/16418) Eve 2025-10-04 20:04:27 +00:00
  • 93882335a8 rpc : check src buffer when copying tensor (llama/16421) Radoslav Gerganov 2025-10-04 16:22:45 +03:00
  • af51bbab88 rpc : add support for multiple devices (llama/16276) Radoslav Gerganov 2025-10-04 12:49:16 +03:00
  • 49e0a426f3 vulkan : incremental shader builds (llama/16341) Acly 2025-10-11 17:59:36 +03:00
  • 93c1305565 metal : fix loop bound in ggml_mem_ranges (llama/16412) Georgi Gerganov 2025-10-03 19:18:56 +03:00
  • a70144a873 ggml : fix graph reallocation with multiple chunks (llama/16396) Acly 2025-10-03 13:49:08 +02:00
  • 2e6888089f vulkan: Replace uses of maxMemoryAllocationSize and VK_WHOLE_SIZE (llama/16354) Jeff Bolz 2025-10-03 05:50:46 -05:00
  • 90bdcf2ef6 vulkan: Fix FA coopmat1 invalid array indexing (llama/16365) Jeff Bolz 2025-10-03 04:52:46 -05:00
  • fd11cd97ab vulkan: in flash attention, bounds check against nem1 (don't rely on GGML_KQ_MASK_PAD) (llama/16316) Jeff Bolz 2025-10-03 03:33:08 -05:00
  • 27ebde6afd ggml webgpu: add support for soft_max, optimize rms_norm (llama/16357) Reese Levine 2025-10-02 11:00:31 -07:00
  • 33ca8355c4 model : Apertus model implementation (llama/15852) Piotr Wilkin (ilintar) 2025-10-02 19:43:22 +02:00
  • e29508be8b musa: update compile flags (llama/16265) R0CKSTAR 2025-10-02 21:29:56 +08:00
  • b73f67d3f6 HIP: Disable ROCWMMA fattn on CDNA when compiled against ROCWMMA 2.0.0 (llama/16221) uvos 2025-10-01 23:09:25 +02:00
  • b0560310aa vulkan: make ggml_vk_default_dispatcher support older vulkan headers (llama/16345) Eve 2025-10-01 07:56:36 +00:00
  • 31bb869929 opencl: support pad_ext (llama/15888) lhez 2025-09-30 10:45:45 -07:00
  • 8208cea829 ggml webgpu: support for rope,div,sub,glu,scale,cont operators (llama/16187) Reese Levine 2025-09-30 09:57:51 -07:00
  • 199626d79e opencl: support ne3 in get_rows (llama/15866) lhez 2025-09-30 09:55:13 -07:00
  • c3b5c4d934 whisper : Support using devices of type iGPU (#3469) Ruben Ortlam 2025-10-11 16:55:16 +02:00
  • 85871a9469 whisper : add support for --carry-initial-prompt (#3395) Andreas Lubbe 2025-10-10 18:51:15 +02:00
  • a0ca50f3b9 cli: Fix assignment for vad_min_silence_duration_ms (#3467) Andreas Lubbe 2025-10-10 15:21:03 +02:00
  • d3a29d7b88 minor : fix code style (#3463) Georgi Gerganov 2025-10-10 11:33:01 +03:00
  • 85d1d3d3dc vad : free vad_segments in whisper_vad (#3463) Silviu Caragea 2025-10-10 04:20:21 +00:00
  • 98930fded1 whisper : clean-up headers Georgi Gerganov 2025-10-09 10:48:40 +03:00
  • 8877dfc11a [skip ci]Bump Ruby bindings' version to 1.3.4 (#3461) KITAITI Makoto 2025-10-08 20:45:20 +09:00
  • c8223a8548 vad : fix memory leaks in VAD implementation (#3453) Daniel Bevenius 2025-10-06 14:57:44 +02:00
  • 7849aff7a2 ruby : Loose RegExp for test (#3448) KITAITI Makoto 2025-10-01 21:33:11 +09:00
  • 2a56869669 bindings-java : disable flash attention by default (#3445) Daniel Bevenius 2025-10-01 09:13:34 +02:00
  • 8c0855fd6b bench : update [no ci] Georgi Gerganov 2025-09-30 21:40:32 +03:00
  • 47fcd7da8b scripts : add -nfa option [no ci] Georgi Gerganov 2025-09-30 21:37:00 +03:00
  • 8a67c55c8a wchess : fix link [no ci] Georgi Gerganov 2025-09-30 21:28:03 +03:00
  • 41fc9dea6a release : v1.8.0 v1.8.0 Georgi Gerganov 2025-09-30 21:25:16 +03:00
  • 5904d00dbb examples : add wchess.wasm to wasm examples build (#3443) Daniel Bevenius 2025-09-30 16:23:01 +02:00
  • 0b3587acdd whisper : enable flash attention by default (#3441) Georgi Gerganov 2025-09-30 15:47:20 +03:00
  • 1e5ad50f8f bench : add rtx 5090 [no ci] Georgi Gerganov 2025-09-30 13:28:50 +03:00
  • 527ff158d0 ggml : bump version to 0.9.4 (ggml/1363) Georgi Gerganov 2025-09-30 13:42:39 +03:00
  • 8a7f3b03c9 ggml : bump version to 0.9.4 (ggml/1363) sync-ggml-25-09-30-2 Georgi Gerganov 2025-09-30 13:42:39 +03:00
  • e4bf87b0e9 bench : update [no ci] Georgi Gerganov 2025-09-30 12:51:25 +03:00
  • b57b9d3a27 sync : ggml Georgi Gerganov 2025-09-30 12:31:08 +03:00
  • 62b3b86e3f cuda : Enable CUDA Graph usage for Nemotron Nano v2 (NemotronH) (llama/16328) anavp-nvidia 2025-09-30 08:13:22 +00:00
  • 78f85f2b92 metal : dynamic simdgroups for MV kernels (llama/16340) Georgi Gerganov 2025-09-30 11:03:23 +03:00
  • 01e86b69ab kleidiai : fix work size and threads sync for fp16 (llama/16246) Charles Xu 2025-09-30 09:07:20 +02:00
  • 35ebdf7304 ggml: riscv: add riscv spacemit backend (llama/15288) alex-spacemit 2025-09-29 22:50:44 +08:00
  • 94fe9bbe2b ggml-backend : add root cause in error message if loading backend library fails (llama/16172) Rafal Lewczuk 2025-09-29 13:17:09 +02:00
  • 32be14f8eb bench : update [no ci] (#3439) Georgi Gerganov 2025-09-29 17:42:38 +03:00
  • a77d11d91e bench : warm-up all kernels (#3438) Georgi Gerganov 2025-09-29 17:27:53 +03:00
  • 22c12ee86d ggml : remove oboslete files (#0) Georgi Gerganov 2025-09-29 16:47:30 +03:00
  • d8cdcce884 ci : add self-hosted workflows (#3437) Georgi Gerganov 2025-09-29 16:42:39 +03:00
  • b4909a6c78 whisper : remove ggml_mul_mat padding (#3436) Georgi Gerganov 2025-09-29 16:42:08 +03:00
  • fcf0181ee2 talk-llama : sync llama.cpp Georgi Gerganov 2025-09-29 15:18:41 +03:00
  • 404a93114c sync : ggml Georgi Gerganov 2025-09-29 15:18:18 +03:00
  • 3201382792 cmake : remove metal flag (llama/0) Georgi Gerganov 2025-09-29 12:33:38 +03:00
  • 112e10f2e4 ggml : check cuda and metal argsort limits and add test (llama/16323) Sigbjørn Skjæret 2025-09-29 11:09:00 +02:00
  • 7ce0a7bcd0 ggml : fix dependencies for ggml_set_rows (llama/16318) Georgi Gerganov 2025-09-29 08:41:28 +03:00
  • a375e4c4d2 vulkan: Fix validation failure in quantized flash attention (llama/16292) Jeff Bolz 2025-09-28 23:50:37 -05:00
  • 5c6e795607 ggml : fix GGML_F32_VEC_FMA argument order in ggml_vec_mad1_f32 (llama/16307) Sigbjørn Skjæret 2025-09-28 23:15:03 +02:00
  • 55d45edf6d vulkan: 64-bit im2col (llama/16135) Jeff Bolz 2025-09-28 01:38:37 -05:00
  • 0102733cca metal : extend mat-mat multiplication support (llama/16225) Georgi Gerganov 2025-09-28 09:34:44 +03:00
  • 45976f2857 metal : fuse non-sequential nodes (llama/16102) Georgi Gerganov 2025-09-28 09:34:05 +03:00
  • 91ab93b756 vulkan: handle mat_mul with A matrix > 4GB (llama/16176) Jeff Bolz 2025-09-27 20:36:34 -05:00
  • eb982dd786 vulkan: support arbitrary KV dimension in flash attention (llama/16160) Jeff Bolz 2025-09-27 16:43:39 -04:00
  • bc1ac13c2f vulkan : make the vulkan.hpp dynamic dispatcher instance private (llama/16224) Acly 2025-09-27 22:41:03 +02:00
  • 85e4455cd3 CUDA: mul_mat_id for mmf for bs <= 64 for f16 and bs <= 32 for f32 (llama/16277) Aman Gupta 2025-09-28 00:49:32 +08:00
  • e856483cd6 CUDA: refactor and deduplicate vector FA kernels (llama/16208) Johannes Gäßler 2025-09-27 18:45:07 +02:00
  • 88dd9e0d45 vulkan: throw system error instead of SIGABRT during init on older devices (llama/16156) Dmytro Minochkin 2025-09-27 19:26:46 +03:00
  • 97bd65f90f vulkan: support GET_ROWS for k-quants (llama/16235) Jeff Bolz 2025-09-27 06:36:11 -04:00
  • 23b3598952 devops: add s390x & ppc64le CI (llama/15925) Aaron Teo 2025-09-27 02:03:33 +08:00
  • 670d54ef5d metal : report OOM errors (llama/16274) Georgi Gerganov 2025-09-26 14:14:28 +03:00
  • 9823c5cc51 common : use cpp-httplib as a cURL alternative for downloads (llama/16185) Adrien Gallouët 2025-09-26 13:12:19 +02:00
  • 89a7b4d22c ggml-cpu: implement MXFP4 SIMD for s390x (llama/16193) Aaron Teo 2025-09-26 18:27:25 +08:00
  • 98ac209ae1 musa: fix build warnings (llama/15611) R0CKSTAR 2025-09-26 08:56:10 +08:00
  • d9bf63cfb8 CUDA: add a fused top-K MoE kernel (llama/16130) Aman Gupta 2025-09-25 22:35:05 +08:00
  • 24ea5476de ggml : fix loongarch lsx compilation error (llama/15864) junchao-zhao 2025-09-25 17:22:55 +08:00
  • 611ff19f20 ggml : remove -dev suffix from release version (ggml/1355) Daniel Bevenius 2025-09-26 17:34:42 +02:00
  • 06d7b3d124 ggml : bump version to 0.9.3 (ggml/1353) Daniel Bevenius 2025-09-25 14:39:05 +02:00
  • ac678efb35 metal : fuse NORM + MUL + ADD, support non-multiples of 4 (llama/16220) Georgi Gerganov 2025-09-25 11:30:16 +03:00
  • 268f1c961b metal : relax reorder conditions (llama/16216) Georgi Gerganov 2025-09-25 11:29:42 +03:00
  • 0a5b811f2e metal : restore im2col perf (llama/16219) Georgi Gerganov 2025-09-25 11:29:08 +03:00
  • 0946619662 rpc : use ggml logging facilities Radoslav Gerganov 2025-09-25 10:20:02 +03:00
  • cd431223e0 llama: print memory breakdown on exit (llama/15860) Johannes Gäßler 2025-09-24 16:53:48 +02:00