Commit Graph

  • 719a05c665 vulkan: fix mmv subgroup16 selection (llama/15775) Ruben Ortlam 2025-09-03 22:55:10 +02:00
  • 4a702a867c vulkan: don't use std::string in load_shaders, to improve compile time (llama/15724) Jeff Bolz 2025-09-03 13:33:15 -05:00
  • 4144ae10e9 vulkan : update ggml_vk_instance_validation_ext_available (llama/15666) Daniel Bevenius 2025-09-03 20:24:50 +02:00
  • 85c7aa3750 ggml vulkan: add hardsigmoid and hardswish operations (llama/15762) Shin-myoung-serp 2025-09-04 03:22:55 +09:00
  • 9eef377330 CUDA: Optimize rms_norm_f32 kernel and its fused variants, giving 1-6% perf E2E (llama/15715) Oliver Simons 2025-09-03 19:59:16 +02:00
  • 51bc843f3a CANN: Add RoPE contiguous check for 310I DUP device (llama/15735) hipudding 2025-09-03 16:46:01 +08:00
  • 75f739c7c8 ggml-cpu : optimize RVV kernels (llama/15720) xctan 2025-09-03 16:16:21 +08:00
  • 91e9e72ecd CANN: Mask unsupported TRANSPOSE_1D operator (llama/15733) hipudding 2025-09-03 14:08:22 +08:00
  • d84b96d9d0 CANN: Fix type float_t to float (llama/15736) Chenguang Li 2025-09-03 10:43:53 +08:00
  • e584edb5ba vulkan: fix shaders gen when no integer dot is available (llama/15740) Ruben Ortlam 2025-09-02 16:02:26 +02:00
  • 5aee53c40f CANN: Resolve soft_max precision issue (llama/15730) hipudding 2025-09-02 17:12:37 +08:00
  • 1e03aa66f7 vulkan: Fix macro parameter order for f32 matmul shaders (llama/15716) Jeff Bolz 2025-09-02 01:37:01 -05:00
  • fb37f91163 opencl: add attn sinks support for FA kernels (llama/15706) rmatif 2025-09-02 08:26:53 +02:00
  • 3db49c1c26 CANN: Support eager execution mode under ACL graph compilation (llama/15712) Chenguang Li 2025-09-02 14:07:48 +08:00
  • 13d3963f71 CANN: Support ext_factor in rope (llama/15710) hipudding 2025-09-02 14:05:23 +08:00
  • f20a7b0e99 ggml-backend: raise GGML_MAX_SPLIT_INPUTS (llama/15722) Johannes Gäßler 2025-09-02 01:14:55 +02:00
  • 9e3600e569 vulkan: use memory budget extension to read memory usage (llama/15545) Gilad S 2025-09-01 22:17:42 +03:00
  • 7a5e7368a3 vulkan: add missing clamps in new mul_mat_id paths (llama/15702) Jeff Bolz 2025-09-01 14:01:10 -05:00
  • d5f80a2982 vulkan: disable large mmv subgroups on older Nvidia GPUs (llama/15717) Ruben Ortlam 2025-09-01 20:58:35 +02:00
  • 8218dc609c ggml: SVE support for exponential functions (llama/15145) s-goto-11 2025-09-02 03:13:49 +09:00
  • 31840a3a56 ggml: aarch64: Implement SVE F16 kernels for vector functions (llama/15115) Prashant Vithule 2025-09-01 23:43:16 +05:30
  • 5e70d901b0 Vulkan: Add Integer Dot Product mul_mat_vec shader for legacy quants (llama/14903) Ruben Ortlam 2025-09-01 16:19:07 +02:00
  • c5f511e697 ggml : WebGPU add TRANSPOSE and RESHAPE to supported ops (llama/15695) Daniel Bevenius 2025-09-01 14:28:49 +02:00
  • 2ba5e0cb47 CUDA: fix build error from ambiguous __half conversions in conv2d (llama/15690) Akarshan Biswas 2025-09-01 06:55:06 +05:30
  • bb5f844ec7 CANN: Optimize MUL_MAT_ID (llama/15658) hipudding 2025-09-01 08:57:23 +08:00
  • ed7ebdc757 CANN: fix RoPE cache issue on multi-device (llama/15629) hipudding 2025-09-01 08:57:00 +08:00
  • 3d470687de metal : fix checks for available FA kernels (llama/15700) Georgi Gerganov 2025-08-31 19:43:30 +03:00
  • b11c972b88 llama : separate compute buffer reserve from fattn check (llama/15696) Diego Devesa 2025-08-31 06:49:03 -07:00
  • db7ecfb61d vulkan: handle large sizes for get_rows (llama/15686) Jeff Bolz 2025-08-31 03:13:27 -05:00
  • 191def71ce vulkan: mul_mat_id coopmat2 optimizations (llama/15546) Jeff Bolz 2025-08-31 02:06:43 -05:00
  • b092e95aaa vulkan : remove unused portability_enumeration_ext variable (llama/15679) Daniel Bevenius 2025-08-31 08:46:42 +02:00
  • 20ce6fcf6a vulkan: Allow fallback to sysmem memory when vidmem is full (llama/15649) Jeff Bolz 2025-08-31 01:30:54 -05:00
  • 71f0ee70bf vulkan: clamp matmul and FA results to the max finite value (llama/15652) Jeff Bolz 2025-08-31 01:27:57 -05:00
  • 74583845b6 ggml: update kleidiai to v1.13.0 (llama/15663) Charles Xu 2025-08-30 18:03:42 +02:00
  • f6ba3949b6 llama: use FA + max. GPU layers by default (llama/15434) Johannes Gäßler 2025-08-30 16:32:10 +02:00
  • b7809c401b CUDA: use FP32 arithmetic for conv2d (llama/15683) Johannes Gäßler 2025-08-30 16:20:32 +02:00
  • a6dec4f49d vulkan: Skip syncing for prealloc_y when it is reused (llama/15544) Jeff Bolz 2025-08-30 04:11:22 -05:00
  • d629af157e CANN: FIx compiler warnings (llama/15661) Chenguang Li 2025-08-30 10:18:35 +08:00
  • 82ce91e7d2 CUDA: fix bug in rms_norm fusion (llama/15660) Aman Gupta 2025-08-29 21:30:06 +08:00
  • 6d7ddaf793 CUDA: fuse adds, fuse add with rms norm (llama/15631) Aman Gupta 2025-08-29 11:35:58 +08:00
  • dc9f55bbb0 CUDA: add conv2d (llama/15635) mnehete32 2025-08-29 00:03:03 +05:30
  • 6287027a2c ggml-cpu: fix invalid hsum build in debug s390x (llama/15634) Aaron Teo 2025-08-28 22:39:27 +08:00
  • 6dffbaa0cb ggml : fix SSM_SCAN for n_groups > 1 (llama/15625) compilade 2025-08-28 10:11:36 -04:00
  • cac6253744 kv-cache : remove LLAMA_SET_ROWS checks (llama/15505) Georgi Gerganov 2025-08-28 12:27:02 +03:00
  • 88c0582b61 cuda: Add cublasLt_static linking when GGML_STATIC is enabled (llama/15622) matiaslin 2025-08-27 17:32:36 -07:00
  • 65fa2c0c1a HIP: Enable support for ggml_backend_cuda_register_host_buffer (llama/15615) uvos 2025-08-27 13:58:54 +02:00
  • 02e8b23137 CANN: refactor mask handling and improve performance in FA (llama/15561) Chenguang Li 2025-08-27 17:21:41 +08:00
  • ece1bdfe7e ggml-cpu : add basic RVV support for vector f32 ops (llama/15057) xctan 2025-08-27 16:44:22 +08:00
  • a6ec224efa OpenCL: add fused group_norm/norm, mul, add (llama/15314) rmatif 2025-08-27 08:36:05 +02:00
  • 94fa9f63b3 SYCL: fix rms_norm_mul_add for tensor dim not a multiple of sg_size (llama/15592) Akarshan Biswas 2025-08-27 00:27:49 +05:30
  • 31c7784e09 llamafile: PowerPC Sgemm Optimization (llama/15558) shalinib-ibm 2025-08-26 21:05:25 +05:30
  • 53010199a1 CUDA: return -1 for nonexistent compiled arch (llama/15587) Johannes Gäßler 2025-08-26 16:01:20 +02:00
  • 1c21a850be metal : optimize FA vec for large sequences and BS <= 8 (llama/15566) Georgi Gerganov 2025-08-26 14:22:14 +03:00
  • dc693ca8c9 metal : improve MUL_MAT_ID (llama/15541) Georgi Gerganov 2025-08-26 12:46:15 +03:00
  • 3bb52acb46 metal : remove contiguous assertion for src0 in IM2COL (llama/15577) Sigbjørn Skjæret 2025-08-26 08:51:43 +02:00
  • 9828caafb5 Add a warning for special devices (llama/15563) Yoshi_likes_e4 2025-08-26 13:15:33 +07:00
  • 79e2bd5ea8 vulkan: Remove splitting for mul_mat_id (llama/15568) Jeff Bolz 2025-08-25 23:42:44 -05:00
  • 2468074e91 CUDA: Accelerate MXFP4 table lookup using __byte_perm (llama/15451) Qeeweew 2025-08-26 05:21:22 +08:00
  • 582ef379ab opencl: fix support ops condition for rms_norm (llama/15560) lhez 2025-08-25 14:18:09 -07:00
  • 335d2a5405 vulkan: fix min subgroup 16 condition for mmid subgroup optimization (llama/15565) Ruben Ortlam 2025-08-25 17:56:59 +02:00
  • 8851ef5463 metal: fix regression when no metal devices are present (llama/15531) Ihar Hrachyshka 2025-08-25 11:27:34 -04:00
  • 1e856b2919 CUDA: MoE helper in device code, better tile sizes (llama/15525) Johannes Gäßler 2025-08-25 17:23:40 +02:00
  • 54be54f4ce metal : add FA kernels for HS=40 (llama/15559) Georgi Gerganov 2025-08-25 10:14:48 +03:00
  • 86331f74e0 CANN: ROPE cache sin/cos repeat (llama/15501) Chenguang Li 2025-08-25 10:32:21 +08:00
  • ee11ed42a9 vulkan: apply MUL_MAT_ID subgroup optimization to non-coopmat devices (llama/15524) Ruben Ortlam 2025-08-24 19:36:36 +02:00
  • 85d4d2c875 vulkan: Support FA with any multiple of 8 head sizes (llama/15537) Jeff Bolz 2025-08-24 04:24:25 -05:00
  • 8c7872d6ed vulkan: enable Conv2D for Apple after MoltenVK fixed the bug (llama/15526) Ruben Ortlam 2025-08-24 10:48:53 +02:00
  • 27817867cc vulkan: workaround MoltenVK compile failure in multi_add (llama/15506) Jeff Bolz 2025-08-24 03:48:21 -05:00
  • b0d15e1eb6 CUDA: fix half2 -> half conversion for HIP (llama/15529) Johannes Gäßler 2025-08-23 21:37:06 +02:00
  • 2f6288c33c vulkan: optimize rms_norm, and allow the work to spread across multiple SMs (llama/15281) Jeff Bolz 2025-08-23 13:16:17 -05:00
  • d8eb9f7d67 vulkan: Rewrite synchronization to allow some overlap between nodes (llama/15489) Jeff Bolz 2025-08-23 02:33:36 -05:00
  • 5094171c37 vulkan : support ggml_mean (llama/15393) Acly 2025-08-23 08:35:21 +02:00
  • 485c5c3b3b vulkan: optimize mul_mat_id loading row ids into shared memory (llama/15427) Jeff Bolz 2025-08-23 01:31:54 -05:00
  • bb5d7e2c31 ggml WebGPU: add support for quantization types (llama/15440) Reese Levine 2025-08-22 11:28:03 -07:00
  • d7b7498e76 ggml: add conv3d op (llama/15182) rmatif 2025-08-22 15:33:15 +02:00
  • 18ca4e8f63 cuda : add Pad Reflect 1D support (llama/14659) Yavor Ivanov 2025-08-22 14:06:29 +03:00
  • 380d3db216 ggml-cpu: Support Q5_0 and Q5_1 on s390x (llama/15486) Aaron Teo 2025-08-22 16:11:04 +08:00
  • be841c3f6e CANN: Optimize RMS_NORM using cache (llama/15419) Chenguang Li 2025-08-22 14:12:07 +08:00
  • 554f96f385 sched : fix possible use of wrong ids tensor when offloading moe prompt processing (llama/15488) Diego Devesa 2025-08-21 14:09:32 -07:00
  • 9dd5039968 vulkan : support conv_2d_dw with f16 weights (llama/15392) Acly 2025-08-21 17:01:51 +02:00
  • 7eebd498ff vulkan: add exp operation (llama/15456) Dong Won Kim 2025-08-22 00:00:16 +09:00
  • 04d0f9a066 vulkan: Reuse conversion results in prealloc_y (llama/15410) Jeff Bolz 2025-08-21 09:55:00 -05:00
  • c5874bcf42 ggml : fix condition of im2col on Metal backend (llama/15460) Xuan-Son Nguyen 2025-08-21 07:32:26 +02:00
  • 7c077845fd musa: add GGML_UNUSED_VARS (llama/15446) R0CKSTAR 2025-08-21 11:06:05 +08:00
  • 622dec5bf6 sched : copy only the used experts when offloading prompt processing (llama/15346) Diego Devesa 2025-08-20 16:35:28 -07:00
  • 8f0579a33d CUDA: refactor FA support/selection code (llama/15454) Johannes Gäßler 2025-08-20 23:14:14 +02:00
  • 316ed78d68 CUDA: replace GGML_CUDA_F16 with CUDA arch checks (llama/15433) Johannes Gäßler 2025-08-20 16:58:49 +02:00
  • 5907ab3e4a vulkan: shorten pipeline name strings (llama/15431) Jeff Bolz 2025-08-20 09:33:14 -05:00
  • 0eb2d653bd musa: fix build warnings (llama/15258) R0CKSTAR 2025-08-20 10:17:37 +08:00
  • db1d2380a0 opencl: mark argsort unsupported if cols exceed workgroup limit (llama/15375) lhez 2025-08-20 02:25:51 +08:00
  • 2572322bac CANN: optimize rope operator (llama/15335) SHUAI YANG 2025-08-19 21:28:22 +08:00
  • 02b49af98d musa: handle __hgt2_mask, available starting from MUSA SDK rc4.3.0 (llama/15413) R0CKSTAR 2025-08-19 18:33:47 +08:00
  • 2ce5860a62 ggml-cpu: add mxfp4 VSX intrinsics for Power9+ (ppc64le) hardware (llama/15385) Marvin Gießing 2025-08-19 10:54:31 +02:00
  • 80447f7412 cuda : remove obsolete sources (ggml/1332) Georgi Gerganov 2025-08-18 22:01:00 +03:00
  • 44fa2f647c ggml : Fix MKL detection by quoting BLAS_INCLUDE_DIRS (#3426) Carlos Zoido 2025-09-19 05:33:53 +02:00
  • edea8a9c3c whisper : prefer curl over wget in download scripts (#3409) Siva Mahadevan 2025-09-08 00:32:19 -04:00
  • bb0e1fc60f ci : remove brew installation of cmake for macos-latest (#3408) Daniel Bevenius 2025-09-05 15:20:32 +02:00
  • 9bfc535130 tests : use CMake definitions for model/sample paths (#3406) Daniel Bevenius 2025-09-04 15:08:30 +02:00
  • 7745fcf328 Handle negative value in padding (#3389) Treboko 2025-08-24 18:34:23 +02:00
  • c09b0e0c4c models : update./models/download-ggml-model.cmd to allow for tdrz download (#3381) Thea Mukhi 2025-08-24 05:52:33 -04:00