Commit Graph

  • 3a1ea96373 HIP: adjust RDNA3.5 MMQ kernel selction logic (llama/18666) Johannes Gäßler 2026-01-10 17:19:01 +01:00
  • 484b17053a cmake : update blas logic (llama/18205) Perry Naseck 2026-01-10 11:00:54 -05:00
  • 45be2cd27a Corrected: changed s13 = src1->nb[3] instead of nb[2] (llama/18724) Michael Wand 2026-01-10 01:16:07 -08:00
  • 4af27bf2da opencl: add EXPM1 op (llama/18704) shaofeiqi 2026-01-09 10:13:13 -08:00
  • 4ac8c3b478 Updates to webgpu get_memory (llama/18707) Reese Levine 2026-01-09 08:17:18 -08:00
  • fff3ebd93d llama: use host memory if device reports 0 memory (llama/18587) Aaron Teo 2026-01-09 05:34:56 +08:00
  • a71127dfd8 ggml-webgpu: Fix GGML_MEM_ALIGN to 8 for emscripten. (llama/18628) Masashi Yoshimura 2026-01-09 01:36:42 +09:00
  • 1bb903f599 ggml webgpu: initial flashattention implementation (llama/18610) Reese Levine 2026-01-08 08:23:39 -08:00
  • 0bc0e5616e vulkan: fix push constant size for quantize_q8_1 (llama/18687) Jeff Bolz 2026-01-08 08:40:58 -06:00
  • 678c660e62 vulkan: optimize ssm_scan (llama/18630) Jeff Bolz 2026-01-08 08:16:54 -06:00
  • f2d8588229 metal : add MoE kernel specialization for ne20=5 (llama/18667) 도로로도로또 2026-01-08 19:37:45 +09:00
  • b9965c89a1 ggml: add env var GGML_OP_OFFLOAD_MIN_BATCH (llama/18535) Doctor Shotgun 2026-01-08 01:03:21 -08:00
  • 85a329cb08 opencl: add FILL op support (llama/18682) shaofeiqi 2026-01-07 22:04:50 -08:00
  • 4f2ca7c163 cuda : fix build on cuda 12.8 (llama/18672) Oliver Walsh 2026-01-07 21:32:44 +00:00
  • a91ab72bd9 vulkan: reject ops when a tensor is too large to allocate (llama/18646) Jeff Bolz 2026-01-07 05:03:32 -06:00
  • 096e7e911a vulkan: Warptile tuning for Intel Xe2/Xe3 (llama/18178) virajwad 2026-01-07 02:59:47 -08:00
  • a576ed944a vulkan: more mul mat optimizations (llama/18533) Eve 2026-01-07 10:13:17 +00:00
  • 5c583f3c02 CANN: Fix rename for get_env (llama/18652) hipudding 2026-01-07 16:11:31 +08:00
  • 47671c81db CANN: Rename get_env to get_env_as_lowercase (llama/18624) Raul Torres 2026-01-07 02:01:25 +00:00
  • a5f51ac75b Hexagon add support for f16/f32 flash attention, scale, set-rows and improve f16/32 matmul (llama/18611) Max Krasnyansky 2026-01-06 17:38:29 -08:00
  • 436f30d05f ggml : optimize cuda ssm_scan using warp-level reduction (llama/18505) Aadeshveer Singh 2026-01-06 23:54:34 +05:30
  • dbec71f6cf vulkan: support buffer_from_host_ptr (llama/18467) Jeff Bolz 2026-01-06 10:37:07 -06:00
  • 575d894603 ggml-cuda: refactor cuda graph usage (llama/18637) Aman Gupta 2026-01-06 23:48:45 +08:00
  • ed674cfc10 mmq.cu: tune mmq/rocblas switching for RDNA (llama/18537) Beinsezii 2026-01-06 07:26:07 -08:00
  • 5520f27363 ggml : fix avx512bf16 build (llama/18623) Adrien Gallouët 2026-01-06 07:54:10 +01:00
  • 9a1a6685ba CANN: Make valid_values variable static const (llama/18627) Raul Torres 2026-01-06 03:53:28 +00:00
  • e563e239a7 ggml webgpu: add CEIL operation support (llama/18605) nwyin 2026-01-05 13:38:57 -06:00
  • 9956333361 CUDA: fix FA FP16 accumulator overflow for Granite (llama/18614) Johannes Gäßler 2026-01-05 19:51:13 +01:00
  • 804f545454 ggml-cuda: check for srcs outside the cgraph (llama/18583) Aman Gupta 2026-01-05 22:46:36 +08:00
  • 52ba45e2b8 vulkan: fix topk_moe_sigmoid_norm_bias failures in GLM-4.6 (llama/18582) Jeff Bolz 2026-01-05 04:51:39 -06:00
  • 0a99b4c377 vulkan: handle quantize_q8_1 overflowing the max workgroup count (llama/18515) Jeff Bolz 2026-01-05 04:30:14 -06:00
  • 1d657effe3 CANN: add operator fusion support for ADD + RMS_NORM (llama/17512) Chenguang Li 2026-01-05 15:38:18 +08:00
  • 4d6a3fb00d sampling : add support for backend sampling (llama/17004) Daniel Bevenius 2026-01-04 21:22:16 +01:00
  • f0bf5b8cc3 CUDA: disable cuda graph when using n-cpu-moe (llama/18593) Aman Gupta 2026-01-05 01:37:48 +08:00
  • 88f5765c82 ggml-cuda: remove unused params in ggml_cuda_graph (llama/18579) Aman Gupta 2026-01-05 01:37:09 +08:00
  • 1e725546b0 ggml-cuda: fixes for concurrent streams (llama/18496) Aman Gupta 2026-01-03 23:15:01 +08:00
  • 60d178cee9 CUDA: only allocate FA tmp buffer if needed (llama/18564) Johannes Gäßler 2026-01-03 13:55:53 +01:00
  • 304e780e5f (Bugfix, ggml-cuda) Pool alloc count fix + small size computation type adjustment (llama/18559) pl752 2026-01-03 15:13:40 +05:00
  • c9e9f083c2 ggml-hexagon: optimize activation function (llama/18393) Shouyu 2026-01-03 00:24:24 -05:00
  • 9d83865607 vulkan: Optimize GGML_OP_CUMSUM (llama/18417) Jeff Bolz 2026-01-02 15:32:30 -06:00
  • b7ff521e71 vulkan: Implement mmvq for iq1_s/iq1_m (llama/18450) Jeff Bolz 2026-01-02 13:19:04 -06:00
  • b99c911c49 metal : adjust extra size for FA buffer to avoid reallocations (llama/18545) Georgi Gerganov 2026-01-02 19:02:18 +02:00
  • f328b13d5c rpc : use unordered_map::reserve and emplace (llama/18513) Chris Rohlf 2026-01-02 05:09:36 -05:00
  • fbde389665 cuda : fix copy of large tensors (ggml_nbytes <= INT_MAX assertion) (llama/18433) MeeMin 2026-01-02 04:54:20 +05:30
  • f22c1ccbe4 ggml-cuda: remove unneccesary prints on ggml_cuda_init (llama/18502) Aman Gupta 2026-01-01 19:18:43 +08:00
  • b1f65a4a7e vulkan: extend topk_moe to handle sigmoid w/exp_probs_b for nemotron (llama/18295) Jeff Bolz 2026-01-01 01:58:27 -06:00
  • a96310871a examples : fix executable example targets (#3600) Peter A. 2026-01-13 08:08:18 +01:00
  • 679bdb53db ruby : fix segmentation fault (#3591) KITAITI Makoto 2026-01-05 17:41:22 +09:00
  • e9898ddfb9 sync : ggml Georgi Gerganov 2025-12-31 18:26:42 +02:00
  • ce03f8e759 ggml : bump version to 0.9.5 (ggml/1410) Georgi Gerganov 2025-12-31 18:24:07 +02:00
  • 7359ac94d5 talk-llama : sync llama.cpp Georgi Gerganov 2025-12-31 13:13:57 +02:00
  • 54fa8216ea sync : ggml Georgi Gerganov 2025-12-31 13:09:05 +02:00
  • 8189f2cb65 metal : add count_equal op (llama/18314) gatbontonpc 2025-12-31 00:39:48 -08:00
  • 2d250f8049 CUDA: fix KQ max calculation (llama/18487) Johannes Gäßler 2025-12-31 09:37:00 +01:00
  • 5deaf8f2a3 metal : remove BF16 x F16 kernels (llama/18456) Georgi Gerganov 2025-12-31 09:53:48 +02:00
  • 467933199a sycl: add newline at the end of CMakeLists.txt (llama/18503) Aman Gupta 2025-12-31 14:23:44 +08:00
  • a3635494da Work around broken IntelSYCLConfig.cmake in Intel oneAPI 2025.x (llama/18345) Rahul Sathe 2025-12-31 06:38:44 +05:30
  • c9955367d4 kleidiai: add and integrate SVE 256-bit vector-length kernel (llama/18458) Charles Xu 2025-12-30 13:04:53 +01:00
  • 6d4aa96bfa CUDA: add log line when mxfp4 acceleration is used (llama/18483) Aman Gupta 2025-12-30 17:40:46 +08:00
  • 5765c5b04e CUDA: fix replacment of bad archs in CMake (llama/18457) Johannes Gäßler 2025-12-29 17:58:20 +01:00
  • d6cb2407b7 CUDA: Blackwell features for non-native builds (llama/18436) Johannes Gäßler 2025-12-29 09:35:42 +01:00
  • e49e88b2d8 cuda: fix race condition in cumsum (llama/18448) Aman Gupta 2025-12-29 14:07:17 +08:00
  • 20f5729921 HIP: Use mmq on MFMA devices for MUL_MAT_ID in cases where a lot of splits would be generated (llama/18202) uvos 2025-12-28 20:12:55 +01:00
  • b8d209f55c Revert "ggml-cuda: use CMAKE_CUDA_ARCHITECTURES if set when GGML_NATIVE=ON (#18413)" (llama/18426) Aman Gupta 2025-12-28 20:53:36 +08:00
  • 54fe9a645d rpc: fix segfault on invalid endpoint format (llama/18387) o7si 2025-12-28 18:34:41 +08:00
  • b3788ef729 cmake: Added more x86_64 CPU backends when building with GGML_CPU_ALL_VARIANTS=On (llama/18186) Boian Berberov 2025-12-28 07:33:29 +00:00
  • 31fc2c37c8 ggml-cuda: use CMAKE_CUDA_ARCHITECTURES if set when GGML_NATIVE=ON (llama/18413) QDelta 2025-12-27 20:33:14 -05:00
  • a800a3acd1 opencl: allow resizing transpose buffers (llama/18384) lhez 2025-12-27 15:51:14 -08:00
  • 29f8155445 ggml-cuda: Use same regex for GGML_NATIVE=OFF (llama/18407) Aman Gupta 2025-12-27 19:56:27 +08:00
  • 015b618d96 vulkan: preprocess mul_mat_id experts and discard workgroups more quickly (llama/18352) Jeff Bolz 2025-12-26 16:12:58 -06:00
  • e37c8ed94e vulkan: optimize decodeFuncB in coopmat2 mul_mat_id shader (llama/18349) Jeff Bolz 2025-12-26 11:15:50 -06:00
  • 331c6ccd31 vulkan: Use BK=32 for coopmat2 mul_mat_id (llama/18332) Jeff Bolz 2025-12-26 11:15:02 -06:00
  • 35cb4abb67 vulkan: small dequantization improvements (llama/18380) Eve 2025-12-26 17:12:11 +00:00
  • 181e36f194 vulkan: Support UPSCALE w/antialias (llama/18327) Jeff Bolz 2025-12-26 10:00:57 -06:00
  • 67473fef57 vulkan: handle rope with large number of rows (llama/18306) Jeff Bolz 2025-12-26 09:53:46 -06:00
  • 33f75a88ac CANN: implement the SSM_CONV operator (llama/17737) 0Marble 2025-12-26 09:12:04 +08:00
  • 51778354ce ggml-cuda: fix regex for arch list (llama/18371) Aman Gupta 2025-12-26 01:35:14 +08:00
  • 8e02f0919d cuda: optimize cumsum cub path (llama/18362) Aman Gupta 2025-12-25 23:55:38 +08:00
  • ea07c5d3b7 ggml-cuda: fix blackwell native builds (llama/18361) Aman Gupta 2025-12-25 22:12:11 +08:00
  • 5f0488f012 CANN: Add support for CONV_TRANSPOSE_1D when kernel size > 255 (llama/17934) Penglin Cai 2025-12-25 16:46:09 +08:00
  • db75fff539 ggml : optimize cuda cumsum fallback kernel (llama/18343) Aadeshveer Singh 2025-12-25 09:41:13 +05:30
  • 41e578ec8a CUDA: experimental native mxfp4 support for blackwell (llama/17906) Aman Gupta 2025-12-24 22:28:26 +08:00
  • f863735caa vulkan: fix command buffer corruption in ggml_backend_vk_event_wait (llama/18302) Jeff Bolz 2025-12-24 05:36:34 -06:00
  • bab2c02da5 CANN : refactor ACL graph cache (llama/17752) Wang Weixuan 2025-12-24 17:50:24 +08:00
  • 1356600679 vulkan: use fewer FA rows for small cache runs (llama/18280) Ruben Ortlam 2025-12-24 08:59:14 +01:00
  • ec9239d3b7 CANN: Uses yarn_ramp cache in ROPE (llama/17725) TianHao324 2025-12-24 14:55:33 +08:00
  • 9bdd4658f4 rpc : add check for rpc buffer type (llama/18242) Chris Rohlf 2025-12-23 04:56:49 -05:00
  • e4c89612cd ggml-hexagon: create generalized functions for cpu side op (llama/17500) nullname 2025-12-23 15:13:24 +08:00
  • 2f33395197 ggml-hexagon: gelu optimization (llama/18151) Shouyu 2025-12-22 13:56:52 -05:00
  • 5b0c1c1580 llamafile: add rvv support for sgemm kernels (llama/18199) Taimur Ahmad 2025-12-22 23:20:23 +05:00
  • f2fe1e5baf opencl: unpack q4_0 for adreno in get_tensor (llama/18278) lhez 2025-12-22 10:19:01 -08:00
  • dbbe6c11b5 vulkan: Extend rope fusions to allow mrope (llama/18264) Jeff Bolz 2025-12-22 11:03:13 -06:00
  • 98e59a43d1 vulkan: Implement set_tensor_async and the event interfaces (llama/18047) Jeff Bolz 2025-12-21 14:52:09 -06:00
  • b68b12f2d5 llama: fix RPC for -fit on (llama/18233) Johannes Gäßler 2025-12-21 19:33:08 +01:00
  • b893e0813a vulkan: fix im2col overflowing maxworkgroupcount (llama/18180) Jeff Bolz 2025-12-21 03:32:58 -06:00
  • f407c5e562 vulkan/cuda: fix topk_moe with exp_probs_b (llama/18071) Jeff Bolz 2025-12-21 03:27:34 -06:00
  • ad6ee3865d vulkan: support GGML_UNARY_OP_XIELU (llama/18062) Jeff Bolz 2025-12-21 03:17:58 -06:00
  • 3cd141f1a9 vulkan: in graph_optimize, try to group ADD operations (llama/18060) Jeff Bolz 2025-12-21 03:05:08 -06:00
  • 449fc7c024 Vulkan: some improvement on mul_mat_iq2_xs (llama/18031) lovedheart 2025-12-21 09:59:52 +01:00
  • 0983985f06 Added comments explaining thread block size selection logic based on row count and column size, derived from historical commit context (llama/18212) Aadeshveer Singh 2025-12-20 16:58:57 +05:30