Code

sync : ggml#2868

Code

Merged

ggerganov merged 38 commits intomasterggerganov/whisper.cpp:masterfrom sync-ggml-25-03-08ggerganov/whisper.cpp:sync-ggml-25-03-08Copy head branch name to clipboardMar 8, 2025

+4,828-2,518

Commits

Commits on Mar 8, 2025

• Told cmake to install ggml-cpp.h as a public header file. (ggml/1126)

  Show description for e1df78f
  

  2 peopleauthored and

  ggerganov

  committed
  e1df78f

  Copy full SHA for e1df78f

• cuda/cpu: Increase support for fp16 unary operations (ggml/1125)

  Show description for 0ad15e3
  

  cmdr2

  authored and

  ggerganov

  committed
  0ad15e3

  Copy full SHA for 0ad15e3

• cuda/vulkan: specify fp32-only support for some operations in supports_op (ggml/1129)

  Show description for a34489c
  

  cmdr2

  authored and

  ggerganov

  committed
  a34489c

  Copy full SHA for a34489c

• cuda: unary ops as float + de-duplicate (ggml/1130)

  

  cmdr2

  authored and

  ggerganov

  committed
  0c02f62

  Copy full SHA for 0c02f62

• ggml-cpu: Fix build with sve (llama/12059)

  Show description for 0894863
  

  MollySophia

  authored and

  ggerganov

  committed
  0894863

  Copy full SHA for 0894863

• vulkan: fix assertion when qy_needs_dequant (llama/12068)

  Show description for efe4017
  

  jeffbolznv

  authored and

  ggerganov

  committed
  efe4017

  Copy full SHA for efe4017

• cmake: Fix ggml backend dependencies and installation (llama/11818)

  Show description for 6e7f839
  

  vvuksanovic

  authored and

  ggerganov

  committed
  6e7f839

  Copy full SHA for 6e7f839

• vulkan: improve im2col (llama/11826)

  Show description for c270cea
  

  daniandtheweb

  authored and

  ggerganov

  committed
  c270cea

  Copy full SHA for c270cea

• vulkan: matmul dequantization improvements (llama/12015)

  Show description for cd143b8
  

  netrunnereve

  authored and

  ggerganov

  committed
  cd143b8

  Copy full SHA for cd143b8

• CANN: Fix build error with GCC 13 (llama/11990)

  Show description for a252113
  

  hipudding

  authored and

  ggerganov

  committed
  a252113

  Copy full SHA for a252113

• ggml: aarch64: implement SVE kernels for q2_k_q8_k vector dot (llama/12064)

  Show description for be928d3
  

  2 peopleauthored and

  ggerganov

  committed
  be928d3

  Copy full SHA for be928d3

• CUDA: fix logic for V100 + GGML_CUDA_FORCE_MMQ (llama/12098)

  

  JohannesGaessler

  authored and

  ggerganov

  committed
  00b059e

  Copy full SHA for 00b059e

• vulkan: add specific MMV kernels for IQ2 and IQ3 quants + optimizations (llama/11595)

  Show description for 4f45b43
  

  remyoudompheng

  authored and

  ggerganov

  committed
  4f45b43

  Copy full SHA for 4f45b43

• ggml : upgrade init_tensor API to return a ggml_status (llama/11854)

  Show description for 2724346
  

  2 peopleauthored and

  ggerganov

  committed
  2724346

  Copy full SHA for 2724346

• CUDA: compress mode option and default to size (llama/12029)

  Show description for 85e0461
  

  Green-Sky

  authored and

  ggerganov

  committed
  85e0461

  Copy full SHA for 85e0461

• ggml-backend : keep paths in native string type when possible (llama/12144)

  

  slaren

  authored and

  ggerganov

  committed
  aa27e01

  Copy full SHA for aa27e01

• SYCL: Move CPY kernels to a separate file and add few missing kernels (llama/12133)

  Show description for 5167979
  

  qnixsynapse

  authored and

  ggerganov

  committed
  5167979

  Copy full SHA for 5167979

• ggml : fix kleidiai build (llama/12159)

  Show description for 7914670
  

  ag2s20150909

  authored and

  ggerganov

  committed
  7914670

  Copy full SHA for 7914670

• HIP: implement FlashAttention via rocWMMA for CDNA and RDNA3+ (llama/12032)

  Show description for 9a517d6
  

  hjc4869

  authored and

  ggerganov

  committed
  9a517d6

  Copy full SHA for 9a517d6

• ggml : portability fixes for VS 2017 (llama/12150)

  Show description for e38d29c
  

  2 peopleauthored and

  ggerganov

  committed
  e38d29c

  Copy full SHA for e38d29c

• vulkan : sync (llama/0)

  Show description for 66e42fa
  ggerganov

  committed
  66e42fa

  Copy full SHA for 66e42fa

• ggml : ggml_compute_forward_concat() for arbitrary tensor type (ggml/1118)

  Show description for ea027a4
  

  vmobilis

  authored and

  ggerganov

  committed
  ea027a4

  Copy full SHA for ea027a4

• ggml : fix GGMLMetalClass ODR (llama/12200)

  Show description for 6979ba8
  

  pminev

  authored and

  ggerganov

  committed
  6979ba8

  Copy full SHA for 6979ba8

• SYCL: Disable f16 Unary OPs as not supported by the kernels (llama/12201)

  

  qnixsynapse

  authored and

  ggerganov

  committed
  c19d8eb

  Copy full SHA for c19d8eb

• ggml-cpu: Faster IQ1 mul_mat_vec on AVX2 using BMI2 instructions (llama/12154)

  Show description for 2b2c567
  

  remyoudompheng

  authored and

  ggerganov

  committed
  2b2c567

  Copy full SHA for 2b2c567

• opencl : fix profile-related errors (llama/12095)

  Show description for dccfba4
  

  2 peopleauthored and

  ggerganov

  committed
  dccfba4

  Copy full SHA for dccfba4

• opencl : fix ulong kernel args were set from int variables (llama/12174)

  Show description for 03b31a0
  

  linehill

  authored and

  ggerganov

  committed
  03b31a0

  Copy full SHA for 03b31a0

• opencl : fix buffer alignment (llama/12197)

  Show description for c6e70f7
  

  linehill

  authored and

  ggerganov

  committed
  c6e70f7

  Copy full SHA for c6e70f7

• HIP/CUDA: set the paramerter value in maintain_cuda_graph instead of replaceing it. (llama/12209)

  Show description for 8a3a93e
  

  IMbackK

  authored and

  ggerganov

  committed
  8a3a93e

  Copy full SHA for 8a3a93e

• CUDA: fix FA logic for PTX 7.0 and CC >= 7.5 (llama/12222)

  

  JohannesGaessler

  authored and

  ggerganov

  committed
  425cefd

  Copy full SHA for 425cefd

• cmake : fix undefined reference errors for std::filesystem in ggml (#12092) (llama/12094)

  Show description for f755166
  

  2 peopleauthored and

  ggerganov

  committed
  f755166

  Copy full SHA for f755166

• opencl: Noncontiguous norm, rms_norm, disable fp16 for some ops (llama/12217)

  Show description for dbf9384
  

  lhez

  authored and

  ggerganov

  committed
  dbf9384

  Copy full SHA for dbf9384

• metal : fix default.metallib build (llama/12224)

  Show description for 41079a9
  

  danbev

  authored and

  ggerganov

  committed
  41079a9

  Copy full SHA for 41079a9

• metal : simplify kernel arguments using a struct (ggml/3229) (llama/12194)

  Show description for 5141a2c
  

  2 peopleauthored and

  ggerganov

  committed
  5141a2c

  Copy full SHA for 5141a2c

• ggml-cpu: faster AVX2 variant for IQ1_M (llama/12216)

  

  remyoudompheng

  authored and

  ggerganov

  committed
  16754cd

  Copy full SHA for 16754cd

• sync : ggml

  ggerganov

  committed
  0b4956d

  Copy full SHA for 0b4956d

• cmake : fix ggml-config (ggml/0)

  ggerganov

  committed
  ed5c494

  Copy full SHA for ed5c494

• objc : fix build, tmp remove GPU support, use C++17

  ggerganov

  committed
  209e1f3

  Copy full SHA for 209e1f3