Commit Graph

  • 7ed570ee94 vulkan: fix shmem overrun in mmq id shader (llama/16873) Ruben Ortlam 2025-10-31 08:14:49 +01:00
  • 486d39c2cb ggml-hexagon: respect input size when getting/setting tensor data (llama/16836) l3utterfly 2025-10-31 12:46:31 +08:00
  • 7fdd53ac0d opencl: fix boundary handling for mul_mm (llama/16875) lhez 2025-10-30 16:00:20 -07:00
  • ffe1c832bd cpu: introduce chunking for repack matmuls and enable matmul-id chunking on ARM64 (llama/16833) Max Krasnyansky 2025-10-30 09:06:13 -07:00
  • e1780b209d model: add support for qwen3vl series (llama/16780) JJJYmmm 2025-10-30 23:19:14 +08:00
  • f1fdb91e95 cpu: introduce chunking for flash attention (llama/16829) Max Krasnyansky 2025-10-30 05:26:05 -07:00
  • f7dfa39104 cuda : fix argsort with 64k+ rows (llama/16849) Sigbjørn Skjæret 2025-10-30 08:56:28 +01:00
  • 887d984558 vulkan: Handle argsort with a large number of rows (llama/16851) Jeff Bolz 2025-10-30 01:27:41 -05:00
  • 41f4daca57 Hide latency of bias and gate-loading (llama/16847) Oliver Simons 2025-10-30 04:34:15 +01:00
  • efe8099268 vulkan: Fuse rope+set_rows (llama/16769) Jeff Bolz 2025-10-29 15:13:10 -05:00
  • 35a3fda240 vulkan: Update topk_moe fusion to handle gpt's late softmax (llama/16656) Jeff Bolz 2025-10-29 08:44:29 -05:00
  • bc944bddc8 Vulkan MMQ Integer Dot Refactor and K-Quant support (llama/16536) Ruben Ortlam 2025-10-29 14:39:03 +01:00
  • 4d74160c9a Hexagon Op queue & dispatch optimizations (llama/16820) Max Krasnyansky 2025-10-29 06:29:12 -07:00
  • 6051c704a0 CUDA: use fastdiv in set-rows (llama/16834) Aman Gupta 2025-10-29 21:11:53 +08:00
  • 82a23ca9c4 vulkan: Call ggml_vk_buffer_write_2d from ggml_vk_buffer_copy (llama/16793) Jeff Bolz 2025-10-29 03:53:04 -05:00
  • 5c316c48f7 CUDA: Fix bug in topk-moe for gpt-oss (llama/16821) Aman Gupta 2025-10-29 15:55:06 +08:00
  • 5850c952e5 sycl: add RMS_NORM_BACK operation support (llama/16808) YaelLogic 2025-10-29 08:14:39 +02:00
  • a983c9219d cuda: add SET operation support (llama/16804) YaelGitAccount 2025-10-28 21:10:28 +02:00
  • f863a42d97 initialise buffer.device in ggml_hexagon_session (llama/16816) l3utterfly 2025-10-28 23:16:20 +08:00
  • cb39359e7f CANN: Improve device ID handling and aclnnArange checks (llama/16752) Chenguang Li 2025-10-28 10:54:53 +08:00
  • 0c8ff48103 CUDA: add unused vars to mmvf and mmvq (llama/16807) Aman Gupta 2025-10-28 10:31:21 +08:00
  • 9664420a54 sycl: add SSM_CONV operation support (llama/16800) tamarPal 2025-10-28 03:50:33 +02:00
  • bcda7c3e58 ggml : fix interpolate with align-corners and ne=1 (llama/16700) Acly 2025-10-27 21:50:22 +01:00
  • 1471b1fda7 HIP: fix AMDGPU_TARGETS, update documentation (llama/16803) Johannes Gäßler 2025-10-27 21:39:49 +01:00
  • 0e1b6c5fc4 sycl: add ROLL operation support (llama/16665) tamarPal 2025-10-27 03:20:24 +02:00
  • 543221d824 sycl: add REPEAT_BACK operation support (llama/16734) shani-f 2025-10-27 03:19:50 +02:00
  • 97c3285cc4 CUDA: support for weight clamp in top-k norm (llama/16702) Aman Gupta 2025-10-27 09:06:16 +08:00
  • bd8734c050 ggml-alloc : make gallocr prefer chunks that allow memory reuse (llama/16788) Acly 2025-10-26 23:19:03 +01:00
  • e6ff2bceed cuda : use fast copy when src and dst are of different type and contiguous (llama/16789) Sigbjørn Skjæret 2025-10-26 21:31:41 +01:00
  • 4f4246dcb4 ggml: fix cuda kernel launch configuration for k_compute_batched_ptrs to support large batch (llama/16744) leejet 2025-10-27 02:13:31 +08:00
  • 9f75cc7eef CUDA: General GEMV fusion (llama/16715) Aman Gupta 2025-10-26 19:28:04 +08:00
  • c00ab7e5e6 vulkan: deduplicate Microsoft Direct3D12 devices (llama/16689) Gilad S 2025-10-26 06:37:38 +02:00
  • d0b544da70 vulkan: delete dead code (llama/16732) Giuseppe Scrivano 2025-10-25 10:59:54 +02:00
  • 070b24f65c vulkan: Optimize SSM_SCAN (llama/16645) Jeff Bolz 2025-10-25 00:04:12 -05:00
  • 5166efa7f0 ggml: fix CUDA grid launch condition for large block_nums.y in binbcast (llama/16742) leejet 2025-10-25 03:39:37 +08:00
  • 524046d4d1 CUDA: use CUB for arbitary size argsort (llama/16754) Aman Gupta 2025-10-24 20:46:19 +08:00
  • 47efc4f115 ggml-cuda: use passed ops instead of hardcoded ops (llama/16712) Aman Gupta 2025-10-23 19:14:06 +08:00
  • 0a5b4c2e9b sycl: use async memory allocation to fix crashes during graph recording (llama/16644) Matthew Michel 2025-10-22 20:05:15 -05:00
  • 8bb12395fe Add experimental ggml-hexagon backend for the Hexagon NPU (llama/16547) Max Krasnyansky 2025-10-22 13:47:09 -07:00
  • a2130ac501 Revert "ggml : Leverage the existing GGML_F32_VEC helpers to vectorize ggml_v…" (#16723) Diego Devesa 2025-10-22 11:20:55 -07:00
  • 773041e336 ggml : Leverage the existing GGML_F32_VEC helpers to vectorize ggml_vec_set_f32 for faster fills (llama/16522) sirus20x6 2025-10-22 05:14:14 -05:00
  • 431aaf56f0 CUDA: fix bug in topk-moe softmax (llama/16711) Aman Gupta 2025-10-22 12:33:08 +08:00
  • ba41a6ca6a CUDA: topk-moe: add optional parameter for gpt-oss (llama/16649) Aman Gupta 2025-10-21 22:40:38 +08:00
  • 99cea274e5 CUDA: better error for FA kernel with 0 occupancy (llama/16643) Johannes Gäßler 2025-10-21 15:27:53 +02:00
  • 999a7e0cbf whisper : enable IGPU (#3492) Oleg Orlov 2025-11-01 15:38:28 +03:00
  • c62adfbd1e ruby : tiny bug fix (#3490) KITAITI Makoto 2025-10-29 03:50:44 +09:00
  • f16c12f3f5 wasm : fix Hebrew ID (#3487) Orel-A 2025-10-27 08:49:32 +02:00
  • 322c2adb75 talk-llama : sync llama.cpp Georgi Gerganov 2025-10-22 08:32:16 +03:00
  • 35ea5ced60 sync : ggml Georgi Gerganov 2025-10-22 08:28:23 +03:00
  • 9a8cfb040c ggml: add ggml_can_fuse_subgraph (llama/16662) Aman Gupta 2025-10-21 16:43:14 +08:00
  • 5c4c477d00 opencl: fix warnings and clean up profiling (llama/16688) lhez 2025-10-20 22:26:17 -07:00
  • 7f16c71068 vulkan: Handle FA with all -inf mask values (llama/16447) Jeff Bolz 2025-10-20 22:16:08 -05:00
  • 55cf00c20a sycl : add PAD_REFLECT_D1 operator support (llama/16145) YehuditE 2025-10-21 01:21:12 +03:00
  • 70b4d22f01 ggml-alloc : fix leak when reusing a tensor with a larger size (llama/16679) Diego Devesa 2025-10-20 05:53:50 -07:00
  • bb76672081 SYCL: Add support for FLOOR,CEIL,ROUND and TRUNC unary operators (llama/16613) safranowith 2025-10-20 11:08:32 +03:00
  • 82bdf31267 ci : fix binaries release failure for s390x (binaries may not work yet) (llama/16664) Aaron Teo 2025-10-20 05:06:39 +08:00
  • 72d98011db HIP: fix GPU_TARGETS (llama/16642) Johannes Gäßler 2025-10-18 14:47:32 +02:00
  • 414901a42c vulkan: Implement topk_moe fused shader, ported from CUDA (llama/16641) Jeff Bolz 2025-10-18 05:22:57 -05:00
  • 08345f15ec CUDA: use registers instead of smem in topk-moe (llama/16647) Aman Gupta 2025-10-18 17:52:53 +08:00
  • 8ffdf4bd96 opencl: transposed gemm/gemv moe kernel with mxfp4,f32 (llama/16602) Shawn Gu 2025-10-17 17:55:32 -07:00
  • 6aa18cccd8 rpc : report actual free memory (llama/16616) Radoslav Gerganov 2025-10-17 18:02:52 +03:00
  • d22008b631 vulkan: Add State Space Model (SSM) Operations Support (llama/16463) Giuseppe Scrivano 2025-10-17 14:23:47 +02:00
  • 328263f8fd ggml : fix SpaceMit IME array out-of-bounds in task assignment (llama/16629) muggle-stack 2025-10-17 18:01:23 +08:00
  • 4a384826a8 vulkan: fix debug build (add_rms_len/data not found) (llama/16624) Jeff Bolz 2025-10-17 02:31:04 -05:00
  • 0ae492641c metal : add CONV_TRANSPOSE_2D (llama/16542) Ilia Ilmer 2025-10-17 02:33:58 -04:00
  • 82332cea27 SYCL SET operator optimized for F32 tensors (llama/16350) GittyBurstein 2025-10-17 05:36:40 +03:00
  • 7bb53032b3 sycl : add ARANGE operator (llama/16362) GittyBurstein 2025-10-16 16:26:21 +03:00
  • fe965613c0 CANN: format code using .clang-format (llama/15863) Chenguang Li 2025-10-16 16:41:11 +08:00
  • 3c136d699a ggml-cpu: replace putenv with setenv for const-correctness (llama/16573) takuya kodama 2025-10-16 13:10:32 +08:00
  • f7b5ecf195 SYCL: Add GGML_OP_MEAN operator support (llama/16009) yael-works 2025-10-16 07:21:28 +03:00
  • 757d51d21d cpu : add FLOOR, CEIL, ROUND and TRUNC unary operators (llama/16083) safranowith 2025-10-15 22:24:51 +03:00
  • bef9f74553 opencl: add q8_0 mm support (llama/16469) lhez 2025-10-15 10:51:04 -07:00
  • 16dab3d122 opencl: fix FA for f32 (llama/16584) lhez 2025-10-15 10:48:28 -07:00
  • d8a146b0f9 metal: optimise GGML_OP_SUM (llama/16559) Sam/Samuel 2025-10-15 23:05:56 +09:00
  • 0c9d49927c CUDA: Changing the CUDA scheduling strategy to spin (llama/16585) Julius Tischbein 2025-10-15 13:54:15 +02:00
  • 8ed913da0e metal : avoid using Metal's gpuAddress property (llama/16576) Georgi Gerganov 2025-10-14 20:33:05 +03:00
  • 23c19308d8 server : set no_context == true (#3482) Georgi Gerganov 2025-10-20 15:39:48 +03:00
  • 4979e04f5d release : v1.8.2 v1.8.2 Georgi Gerganov 2025-10-15 10:29:42 +03:00
  • 8ba3c13b0c talk-llama : sync llama.cpp Georgi Gerganov 2025-10-14 22:09:02 +03:00
  • ff2253b08a sync : ggml Georgi Gerganov 2025-10-14 22:08:53 +03:00
  • 499f183e75 vulkan: Add ACC_TYPE_VEC2 implementation (llama/16203) SavicStefan 2025-10-14 19:18:05 +02:00
  • 2eb9119754 CUDA + openCL: fix bug in accessing rms_norm->src while doing fusion (llama/16577) Aman Gupta 2025-10-14 22:48:08 +08:00
  • 393fbbc80b vulkan: Support FA with K/V in F32 (llama/16543) Jeff Bolz 2025-10-14 08:53:37 -05:00
  • 73e200ee85 vulkan: Improve build time for MSVC (llama/16545) Jeff Bolz 2025-10-14 07:51:36 -05:00
  • 1bdd746bc8 CUDA: enable FA for FP32 KV cache (llama/16546) Johannes Gäßler 2025-10-14 14:22:47 +02:00
  • f2075667fa CUDA: use fastdiv + ggml_cuda_mad for mmvf (llama/16557) Aman Gupta 2025-10-14 19:16:21 +08:00
  • b4c5c6f71f CUDA: add fp kernel for larger batch size MoE (llama/16512) Aman Gupta 2025-10-14 19:15:15 +08:00
  • a12848e8e9 cuda : remove legacy copy-op pointer indirection code (llama/16485) Anav Prasad 2025-10-14 09:53:49 +00:00
  • 25ac94a6cb metal : FA support F32 K and V and head size = 32 (llama/16531) Georgi Gerganov 2025-10-13 23:07:57 +03:00
  • 66b0fc2fb7 opencl: fix build targeting CL 2 (llama/16554) lhez 2025-10-13 11:50:37 -07:00
  • 77272fe0df CUDA: fix numerical issues in tile FA kernel (llama/16540) Johannes Gäßler 2025-10-13 16:29:45 +02:00
  • 8a9c2ba6a1 ggml : fix build broken with -march=armv9-a on MacOS (llama/16520) Jie Fu (傅杰) 2025-10-13 20:48:47 +08:00
  • 417ecdddc5 CANN: fix CPU memory leak in CANN backend (llama/16549) Chenguang Li 2025-10-13 17:01:24 +08:00
  • bfd88b8b6e metal: add support for opt_step_sgd (llama/16539) Sam/Samuel 2025-10-13 16:25:02 +08:00
  • ccac1b4772 ggml : fix scalar path for computing norm (llama/16558) Georgi Gerganov 2025-10-13 11:22:27 +03:00
  • 53e21364a6 CANN: Update several operators to support FP16 data format (llama/16251) hipudding 2025-10-13 08:52:22 +08:00
  • 7f22fe5d8f metal : add opt_step_adamw and op_sum (llama/16529) Sam/Samuel 2025-10-13 02:43:14 +08:00
  • be778c992f fix UT fault cases: count-equal, argsort, pad OPs (llama/16521) Neo Zhang Jianyu 2025-10-12 21:53:35 +08:00
  • 70eb30f28e ggml : Fix FP16 ELU positive branch (llama/16519) sirus20x6 2025-10-12 00:25:37 -05:00
  • 53721d6309 ggml: Correct SVE implementation in ggml_vec_dot_f16_unroll (llama/16518) sirus20x6 2025-10-12 00:15:00 -05:00