Code

cuda : optimize argmax#10441

Code

Merged

slaren merged 6 commits intomasterggml-org/llama.cpp:masterfrom sl/cuda-opt-argmaxggml-org/llama.cpp:sl/cuda-opt-argmaxCopy head branch name to clipboardNov 21, 2024

+104-61

Commits

Commits on Nov 21, 2024

• cuda : optimize argmax

  slaren

  committed
  35386e8

  Copy full SHA for 35386e8

• remove unused parameter

  Show description for 0a737d2
  slaren

  committed
  0a737d2

  Copy full SHA for 0a737d2

• fixup : use full warps

  Show description for 1e9447a
  slaren

  committed
  1e9447a

  Copy full SHA for 1e9447a

• Apply suggestions from code review

  Show description for a734da7
  

  slaren

  and

  JohannesGaessler

  authored
  a734da7

  Copy full SHA for a734da7

• fix ub

  slaren

  committed
  316f3d3

  Copy full SHA for 316f3d3

• ggml : check ne00 <= INT32_MAX in argmax and argsort

  slaren

  committed
  48f94d4

  Copy full SHA for 48f94d4