Code

llama : refactor graph build code#3837

Code

Merged

ggerganov merged 21 commits intomasterggml-org/llama.cpp:masterfrom llama-refactorggml-org/llama.cpp:llama-refactorCopy head branch name to clipboardNov 1, 2023

+1,520-2,234

Commits

Commits on Oct 28, 2023

• llama : factor out ggml-alloc from graph graph build functions

  Show description for 8b2420d
  ggerganov

  committed
  8b2420d

  Copy full SHA for 8b2420d

• metal : disable kernel load log

  ggerganov

  committed
  5946d98

  Copy full SHA for 5946d98

• llama : factor out tensor offloading outside the build call (wip)

  Show description for 38aca9e
  ggerganov

  committed
  38aca9e

  Copy full SHA for 38aca9e

• llama : offload rest of the models

  Show description for 83d2c43
  ggerganov

  committed
  83d2c43

  Copy full SHA for 83d2c43

• llama : update offload log messages to print node index

  ggerganov

  committed
  3af8771

  Copy full SHA for 3af8771

• llama : comments

  ggerganov

  committed
  51c4f9e

  Copy full SHA for 51c4f9e

Commits on Oct 29, 2023

• llama : support offloading result_norm + comments

  ggerganov

  committed
  4e98897

  Copy full SHA for 4e98897

• llama : factor graph input into a function

  ggerganov

  committed
  0dc05b8

  Copy full SHA for 0dc05b8

• llama : do tensor offload only with CUDA

  ggerganov

  committed
  e14aa46

  Copy full SHA for e14aa46

• llama : fix res_norm offloading

  ggerganov

  committed
  7961790

  Copy full SHA for 7961790

• llama : try to optimize offloading code

  ggerganov

  committed
  b4ad03b

  Copy full SHA for b4ad03b

• llama : fix non-CUDA build

  ggerganov

  committed
  25cfbf6

  Copy full SHA for 25cfbf6

• llama : try to fix build

  ggerganov

  committed
  739b85c

  Copy full SHA for 739b85c

• llama : move refact in correct place + optimize graph input

  ggerganov

  committed
  da93618

  Copy full SHA for da93618

• llama : refactor tensor offloading as callback

  ggerganov

  committed
  1e9c544

  Copy full SHA for 1e9c544

• llama : add layer index to all tensor names

  ggerganov

  committed
  8925cf9

  Copy full SHA for 8925cf9

• llama : add functional header

  ggerganov

  committed
  7610879

  Copy full SHA for 7610879

• llama : comment

  Show description for 79ad734
  ggerganov

  committed
  79ad734

  Copy full SHA for 79ad734

• llama : remove obsolete map for layer counting

  ggerganov

  committed
  210e6e5

  Copy full SHA for 210e6e5

Commits on Oct 31, 2023

• llama : add llm_build helper functions (#3848)

  Show description for 5baefef
  ggerganov

  authored
  5baefef

  Copy full SHA for 5baefef

• Merge branch 'master' into llama-refactor

  ggerganov

  committed
  afb3929

  Copy full SHA for afb3929