llama.cpp
vulkan: scale caching for k quants + misc fixes
#11081
Merged
Overview
Commits
25
Changes
View On GitHub
Commits
  • q6_k scale caching
    netrunnereve committed 1 year ago
  • 16 bit unpack
    netrunnereve committed 1 year ago
  • q4_k test (slow)
    netrunnereve committed 1 year ago
  • revert it
    netrunnereve committed 1 year ago
  • q3_k
    netrunnereve committed 1 year ago
  • q2_k
    netrunnereve committed 1 year ago
  • little stuff
    netrunnereve committed 1 year ago
  • try precalculating products of a and q2_k scales
    netrunnereve committed 1 year ago
  • Revert "try precalculating products of a and q2_k scales"
    netrunnereve committed 1 year ago
  • unpack should be u16, add vim swap to gitignore (about time)
    netrunnereve committed 1 year ago
  • better q4_k scales
    netrunnereve committed 1 year ago
  • q5_k
    netrunnereve committed 1 year ago
  • better q6_k with separate paths for all threads and partial threads in use, plus some more optimizations
    netrunnereve committed 1 year ago
  • q2_k better dequant
    netrunnereve committed 1 year ago
  • q3_k optimizations
    netrunnereve committed 1 year ago
  • q3_k use hmask simd from cpu avx version
    netrunnereve committed 1 year ago
  • Merge https://github.com/ggerganov/llama.cpp into vulkan
    netrunnereve committed 1 year ago
  • make the caches happy
    netrunnereve committed 1 year ago
  • q3_k separate out calculation
    netrunnereve committed 1 year ago
  • q2_k separate out
    netrunnereve committed 1 year ago
  • little stuff
    netrunnereve committed 1 year ago
  • merge master
    netrunnereve committed 1 year ago
  • use calc_superblock everywhere
    netrunnereve committed 1 year ago
  • q2_k optimize scale calculation
    netrunnereve committed 1 year ago
  • more barriers
    netrunnereve committed 1 year ago
FAQ Terms Privacy Refunds Impressum
Loading