Code

CUDA performance optimizations#1530

Code

Merged

ggerganov merged 8 commits intoggerganov:masterggerganov/llama.cpp:masterfrom JohannesGaessler:dfyz-xor-hackJohannesGaessler/llama.cpp:dfyz-xor-hackCopy head branch name to clipboardMay 25, 2023

+111-65

Commits

Commits on May 23, 2023

• xor hack

  JohannesGaessler

  committed
  fbf5588

• block y dim

  JohannesGaessler

  committed
  1a78710

• loop unrolling

  JohannesGaessler

  committed
  82cf01f

• Fixed cmake LLAMA_CUDA_BY option

  JohannesGaessler

  committed
  17dc4c5

• Removed hipblas compatibility code

  JohannesGaessler

  committed
  5d0cf99

• Define GGML_CUDA_DMMV_BLOCK_Y if not defined

  JohannesGaessler

  committed
  e199938

• Fewer iters, more ops per iter

  JohannesGaessler

  committed
  98bfee0

• Renamed DMMV X/Y compilation options

  JohannesGaessler

  committed
  d45df1b