Code

Q6_K AVX improvements#10118

Code

Merged

slaren merged 10 commits intoggerganov:masterggerganov/llama.cpp:masterfrom netrunnereve:q6_knetrunnereve/llama.cpp:q6_kCopy head branch name to clipboardNov 4, 2024

+38-51

Commits

Commits on Oct 31, 2024

• q6_k instruction reordering attempt

  netrunnereve

  committed
  499e9f2

  Copy full SHA for 499e9f2

• better subtract method

  netrunnereve

  committed
  e3a3432

  Copy full SHA for e3a3432

• should be theoretically faster

  Show description for 0b75215
  netrunnereve

  committed
  0b75215

  Copy full SHA for 0b75215

• optimize bit fiddling

  netrunnereve

  committed
  a420e4c

  Copy full SHA for a420e4c

• handle -32 offset separately. bsums exists for a reason!

  netrunnereve

  committed
  35255d6

  Copy full SHA for 35255d6

Commits on Nov 1, 2024

• use shift

  netrunnereve

  committed
  5b36715

  Copy full SHA for 5b36715

• Merge branch 'ggerganov:master' into q6_k

  netrunnereve

  authored
  ed6f845

  Copy full SHA for ed6f845

• Update ggml-quants.c

  netrunnereve

  authored
  d84c372

  Copy full SHA for d84c372

Commits on Nov 4, 2024

• Merge branch 'ggerganov:master' into q6_k

  netrunnereve

  authored
  4ec3e4a

  Copy full SHA for 4ec3e4a

• have to update ci macos version to 13 as 12 doesnt work now. 13 is still x86

  netrunnereve

  authored
  f85336e

  Copy full SHA for f85336e