Code

Refactor lora adapter support#8332

Code

Merged

ngxson merged 42 commits intoggml-org:masterggml-org/llama.cpp:masterfrom ngxson:xsn/fix_lorangxson/llama.cpp:xsn/fix_loraCopy head branch name to clipboardJul 15, 2024

+944-511

Commits

Commits on Jul 6, 2024

• lora: load to devide buft

  ngxson

  committed
  67c5e14

  Copy full SHA for 67c5e14

• add patch tensor function

  ngxson

  committed
  e9d7b6c

  Copy full SHA for e9d7b6c

• correct tensor patch

  ngxson

  committed
  4e28ad4

  Copy full SHA for 4e28ad4

• llama_lora_adapter_apply

  ngxson

  committed
  1b4ffba

  Copy full SHA for 1b4ffba

• correct ggml_backend_tensor_copy

  ngxson

  committed
  b88ce0f

  Copy full SHA for b88ce0f

Commits on Jul 7, 2024

• add llm_build_mm

  ngxson

  committed
  f6d090d

  Copy full SHA for f6d090d

• Merge branch 'master' into xsn/fix_lora

  ngxson

  committed
  a1666aa

  Copy full SHA for a1666aa

• fix auto merge

  ngxson

  committed
  30faf1f

  Copy full SHA for 30faf1f

Commits on Jul 8, 2024

• update based on review comments

  ngxson

  committed
  79e2982

  Copy full SHA for 79e2982

• add convert script

  ngxson

  committed
  847135a

  Copy full SHA for 847135a

• no more transpose A

  ngxson

  committed
  712fecb

  Copy full SHA for 712fecb

• add f16 convert

  ngxson

  committed
  84288ff

  Copy full SHA for 84288ff

• Merge branch 'master' into xsn/fix_lora

  ngxson

  committed
  41ced24

  Copy full SHA for 41ced24

• add metadata check

  ngxson

  committed
  0e16188

  Copy full SHA for 0e16188

• add sanity check

  ngxson

  committed
  6c617e2

  Copy full SHA for 6c617e2

• fix ftype

  ngxson

  committed
  7a83f20

  Copy full SHA for 7a83f20

• add requirements

  ngxson

  committed
  d52455f

  Copy full SHA for d52455f

• fix requirements

  ngxson

  committed
  802565c

  Copy full SHA for 802565c

• fix outfile

  ngxson

  committed
  95b3eb0

  Copy full SHA for 95b3eb0

• Merge pull request #8 from ngxson/xsn/fix_lora_convert

  Show description for 03d24ca
  ngxson

  authored
  03d24ca

  Copy full SHA for 03d24ca

Commits on Jul 9, 2024

• conversion: only allow selected models

  ngxson

  committed
  ee2b35c

  Copy full SHA for ee2b35c

• fix types

  ngxson

  committed
  713665d

  Copy full SHA for 713665d

Commits on Jul 10, 2024

• cuda : do not use dmmv if the tensor does not have enough cols

  slaren

  committed
  f15167a

  Copy full SHA for f15167a

• llama : lora fixes

  slaren

  committed
  9841fbd

  Copy full SHA for 9841fbd

• Merge pull request #9 from ggerganov/sl/fix_fix_lora

  Show description for 4fe0861
  ngxson

  authored
  4fe0861

  Copy full SHA for 4fe0861

• do not disable mmap with lora

  Show description for 1faf7e5
  

  ngxson

  and

  slaren

  committed
  1faf7e5

  Copy full SHA for 1faf7e5

• Merge branch 'master' into xsn/fix_lora

  ngxson

  committed
  e68344c

  Copy full SHA for e68344c

• llm_build_lora_mm_id

  ngxson

  committed
  916e959

  Copy full SHA for 916e959

Commits on Jul 12, 2024

• convert_lora : MoE LoRA conversion support

  Show description for 9d96328
  compilade

  committed
  9d96328

  Copy full SHA for 9d96328

Commits on Jul 15, 2024

• convert_hf : simplify modify_tensors for InternLM2

  Show description for 8956543
  compilade

  committed
  8956543

  Copy full SHA for 8956543

• llama : use llm_build_lora_mm in most model graphs

  compilade

  committed
  87301bd

  Copy full SHA for 87301bd

• Merge branch 'master' into xsn/fix_lora

  ngxson

  committed
  703573f

  Copy full SHA for 703573f

• auto scale

  ngxson

  committed
  42415a4

  Copy full SHA for 42415a4

• Revert "auto scale"

  Show description for 5b18118
  ngxson

  committed
  5b18118

  Copy full SHA for 5b18118

• remove redundant params

  ngxson

  committed
  f68d092

  Copy full SHA for f68d092

• Merge branch 'master' into xsn/fix_lora

  ngxson

  committed
  b704448

  Copy full SHA for b704448

• Apply suggestions from code review

  Show description for 9175f4b
  

  ngxson

  and

  slaren

  authored
  9175f4b

  Copy full SHA for 9175f4b

• change kv metadata

  ngxson

  committed
  0ba23ba

  Copy full SHA for 0ba23ba

• move add_type to __init__

  ngxson

  committed
  b1c4069

  Copy full SHA for b1c4069

• Merge branch 'master' into xsn/fix_lora

  ngxson

  committed
  4d9ac0f

  Copy full SHA for 4d9ac0f

• convert_hf : move add_type to main()

  Show description for d09382f
  compilade

  committed
  d09382f

  Copy full SHA for d09382f

• Merge branch 'master' into xsn/fix_lora

  ngxson

  committed
  383b6bc

  Copy full SHA for 383b6bc