Code

Add LoRA support#820

Code

Merged

slaren merged 15 commits intoggerganov:masterggerganov/llama.cpp:masterfrom slaren:loraslaren/llama.cpp:loraCopy head branch name to clipboardApr 17, 2023

+753-41

Commits

Commits on Apr 16, 2023

• Add lora support

  slaren

  committed
  f52101e

  Copy full SHA for f52101e

• Export lora A matrix pre-transposed

  slaren

  committed
  ac3fbe4

  Copy full SHA for ac3fbe4

• Add support for quantized models

  slaren

  committed
  7136ada

  Copy full SHA for 7136ada

• Use the work buffer instead to fix MSVC build

  slaren

  committed
  dc65707

  Copy full SHA for dc65707

• Update exporter and support scaling

  slaren

  committed
  87c518b

  Copy full SHA for 87c518b

• Add compatibility with ggerganov#801

  slaren

  committed
  c920f00

  Copy full SHA for c920f00

• Support more layer types, fix memory and generation issues

  slaren

  committed
  c45868b

  Copy full SHA for c45868b

• Rebase to master

  slaren

  committed
  57627f0

  Copy full SHA for 57627f0

• Add support for using a different base model

  slaren

  committed
  c150e1b

  Copy full SHA for c150e1b

• Fix windows build

  slaren

  committed
  fc89916

  Copy full SHA for fc89916

• Show warning when using a quantized base model

  slaren

  committed
  14858ba

  Copy full SHA for 14858ba

• ggml_cpy: use the work buffer instead of alloca when quantizing

  slaren

  committed
  3df343b

  Copy full SHA for 3df343b

• Only attempt to use mmap for the lora base model if it is supported

  slaren

  committed
  63da54e

  Copy full SHA for 63da54e

• Reuse definitions from convert.py

  slaren

  committed
  0a6d5ad

  Copy full SHA for 0a6d5ad

• ggml_add: Add more checks

  slaren

  committed
  8d37db3

  Copy full SHA for 8d37db3