Code

sync : llama.cpp#1122

Code

Merged

ggerganov merged 29 commits intomasterggml-org/ggml:masterfrom sync-llama.cpp-25-02-25ggml-org/ggml:sync-llama.cpp-25-02-25Copy head branch name to clipboardFeb 25, 2025

+7,111-2,004

Commits

Commits on Feb 25, 2025

• HIP: Remove GCN from list of devices that avoid MMQ (llama/11831)

  

  IMbackK

  authored and

  ggerganov

  committed
  bedefd2

  Copy full SHA for bedefd2

• ggml : x2 speed for WASM by optimizing SIMD (llama/11453)

  Show description for c931f5a
  

  2 peopleauthored and

  ggerganov

  committed
  c931f5a

  Copy full SHA for c931f5a

• ggml-cpu : add chunking support to mul_mat_id (llama/11666)

  Show description for a338c1b
  

  slaren

  authored and

  ggerganov

  committed
  a338c1b

  Copy full SHA for a338c1b

• musa: bump MUSA SDK version to rc3.1.1 (llama/11822)

  Show description for 8d75e7b
  

  yeahdongcn

  authored and

  ggerganov

  committed
  8d75e7b

  Copy full SHA for 8d75e7b

• llamafile: use member variable instead of constant for iq4nlt (llama/11780)

  

  jmorganca

  authored and

  ggerganov

  committed
  602b296

  Copy full SHA for 602b296

• vulkan: linux builds + small subgroup size fixes (llama/11767)

  Show description for f8cb11a
  

  netrunnereve

  authored and

  ggerganov

  committed
  f8cb11a

  Copy full SHA for f8cb11a

• ggml: optimize some vec dot functions for LoongArch ASX (llama/11842)

  Show description for 8a8e070
  

  MQ-mengqing

  authored and

  ggerganov

  committed
  8a8e070

  Copy full SHA for 8a8e070

• cuda : add ampere to the list of default architectures (llama/11870)

  

  slaren

  authored and

  ggerganov

  committed
  63d1630

  Copy full SHA for 63d1630

• opencl: Fix rope and softmax (llama/11833)

  Show description for 7388906
  

  lhez

  authored and

  ggerganov

  committed
  7388906

  Copy full SHA for 7388906

• vulkan: initial support for IQ1_S and IQ1_M quantizations (llama/11528)

  Show description for c9d2746
  

  remyoudompheng

  authored and

  ggerganov

  committed
  c9d2746

  Copy full SHA for c9d2746

• repo : update links to new url (llama/11886)

  Show description for 284a156
  ggerganov

  committed
  284a156

  Copy full SHA for 284a156

• metal : optimize dequant q6_K kernel (llama/11892)

  

  akretz

  authored and

  ggerganov

  committed
  551157a

  Copy full SHA for 551157a

• metal : fix the crash caused by the lack of residency set support on Intel Macs. (llama/11904)

  

  halechan

  authored and

  ggerganov

  committed
  fc7c65f

  Copy full SHA for fc7c65f

• vulkan: support multi/vision rope, and noncontiguous rope (llama/11902)

  

  jeffbolznv

  authored and

  ggerganov

  committed
  73b8cf3

  Copy full SHA for 73b8cf3

• vulkan: implement several ops relevant for ggml_opt (llama/11769)

  Show description for 4a790c6
  

  remyoudompheng

  authored and

  ggerganov

  committed
  4a790c6

  Copy full SHA for 4a790c6

• CUDA: use async data loading for FlashAttention (llama/11894)

  Show description for de4401e
  

  2 peopleauthored and

  ggerganov

  committed
  de4401e

  Copy full SHA for de4401e

• ggml: aarch64: implement SVE kernels for q3_K_q8_K vector dot (llama/11917)

  Show description for f156da3
  

  3 peoplecommitted
  f156da3

  Copy full SHA for f156da3

• ggml-cpu: Add CPU backend support for KleidiAI library (llama/11390)

  Show description for c32a87b
  

  chaxu01

  authored and

  ggerganov

  committed
  c32a87b

  Copy full SHA for c32a87b

• MUSA: support ARM64 and enable dp4a .etc (llama/11843)

  Show description for ce7334a
  

  2 peopleauthored and

  ggerganov

  committed
  ce7334a

  Copy full SHA for ce7334a

• CUDA: correct the lowest Maxwell supported by CUDA 12 (llama/11984)

  Show description for d09fbd0
  

  2 peopleauthored and

  ggerganov

  committed
  d09fbd0

  Copy full SHA for d09fbd0

• cuda: Add Q5_1, Q5_0, Q4_1 and Q4_0 to F32 conversion support. (llama/12000)

  

  gcp

  authored and

  ggerganov

  committed
  4ce12b2

  Copy full SHA for 4ce12b2

• CUDA: optimize FA for GQA + large batches (llama/12014)

  

  JohannesGaessler

  authored and

  ggerganov

  committed
  59a7d33

  Copy full SHA for 59a7d33

• CUDA: app option to compile without FlashAttention (llama/12025)

  

  JohannesGaessler

  authored and

  ggerganov

  committed
  ae06ada

  Copy full SHA for ae06ada

• ggml-cpu: Support s390x SIMD Instruction Set (llama/12019)

  Show description for 9b56573
  

  3 peopleauthored and

  ggerganov

  committed
  9b56573

  Copy full SHA for 9b56573

• SYCL: Fix GGML_SYCL_DEBUG macro (llama/11995)

  

  qnixsynapse

  authored and

  ggerganov

  committed
  9de4c68

  Copy full SHA for 9de4c68

• Optimize mul_mat for Q4_0 on Intel GPU (llama/12035)

  Show description for 0e96656
  

  2 peopleauthored and

  ggerganov

  committed
  0e96656

  Copy full SHA for 0e96656

• opencl: fix for small models (llama/11950)

  Show description for 92ac3ac
  

  3 peopleauthored and

  ggerganov

  committed
  92ac3ac

  Copy full SHA for 92ac3ac

• metal : copy kernels for quant to F32/F16 conversions (llama/12017)

  Show description for ffcb1ee
  

  gcp

  and

  ggerganov

  committed
  ffcb1ee

  Copy full SHA for ffcb1ee

• sync : llama.cpp

  Show description for 41c52d2
  ggerganov

  committed
  41c52d2

  Copy full SHA for 41c52d2