Code

sync : llama.cpp#1098

Code

Merged

ggerganov merged 13 commits intomasterggml-org/ggml:masterfrom sync-llama.cpp-25-02-03ggml-org/ggml:sync-llama.cpp-25-02-03Copy head branch name to clipboardFeb 3, 2025

Commits

Commits on Feb 3, 2025

• vulkan: Catch pipeline creation failure and print an error message (llama/11436)

  Show description for f250ce8
  

  jeffbolznv

  authored and

  ggerganov

  committed
  f250ce8

  Copy full SHA for f250ce8

• vulkan: implement initial support for IQ2 and IQ3 quantizations (llama/11360)

  Show description for ddd688f
  

  2 peopleauthored and

  ggerganov

  committed
  ddd688f

  Copy full SHA for ddd688f

• CUDA/HIP: add warp_size to cuda_device_info

  

  IMbackK

  authored and

  ggerganov

  committed
  b3e9c4c

  Copy full SHA for b3e9c4c

• HIP: Prepare reduction operators for wave 64

  

  IMbackK

  authored and

  ggerganov

  committed
  0533622

  Copy full SHA for 0533622

• HIP: require at least HIP 5.5

  

  IMbackK

  authored and

  ggerganov

  committed
  b850e64

  Copy full SHA for b850e64

• ci: use sccache on windows instead of ccache (llama/11545)

  Show description for 20ea67d
  

  ochafik

  authored and

  ggerganov

  committed
  20ea67d

  Copy full SHA for 20ea67d

• CUDA: use mma PTX instructions for FlashAttention (llama/11583)

  Show description for 242f4f0
  

  2 peopleauthored and

  ggerganov

  committed
  242f4f0

  Copy full SHA for 242f4f0

• HIP: add GGML_CUDA_CC_IS_* for amd familys as increasing cc archtectures for amd gpus are not supersets of eatch other (llama/11601)

  Show description for 63b95b3
  

  IMbackK

  authored and

  ggerganov

  committed
  63b95b3

  Copy full SHA for 63b95b3

• CUDA/HIP: add support for selectable warp size to mmv (llama/11519)

  Show description for b8b4131
  

  IMbackK

  authored and

  ggerganov

  committed
  b8b4131

  Copy full SHA for b8b4131

• HIP: fix flash_attn_stream_k_fixup warning (llama/11604)

  

  JohannesGaessler

  authored and

  ggerganov

  committed
  9db387e

  Copy full SHA for 9db387e

• sync : llama.cpp

  Show description for 5de9270
  ggerganov

  committed
  5de9270

  Copy full SHA for 5de9270

• CUDA: fix Volta FlashAttention logic (llama/11615)

  

  JohannesGaessler

  authored and

  ggerganov

  committed
  21494b8

  Copy full SHA for 21494b8

• sync : llama.cpp

  Show description for c4033f3
  ggerganov

  committed
  c4033f3

  Copy full SHA for c4033f3