Code

IBM Granite Architecture#9412

Code

Merged

ggerganov merged 9 commits intoggml-org:masterggml-org/llama.cpp:masterfrom gabe-l-hart:GraniteLMgabe-l-hart/llama.cpp:GraniteLMCopy head branch name to clipboardSep 17, 2024

+135-1

Commits

Commits on Sep 16, 2024

• feat(gguf-py): Add Granite model and params to gguf-py

  Show description for 5ebc5ef
  gabe-l-hart

  committed
  5ebc5ef

  Copy full SHA for 5ebc5ef

• feat(convert_hf_to_gguf): Add registration and param setup for Granite

  Show description for 406833d
  gabe-l-hart

  committed
  406833d

  Copy full SHA for 406833d

• feat(llama.cpp): Add config parsing for Granite multiplier params

  Show description for 383065a
  gabe-l-hart

  committed
  383065a

  Copy full SHA for 383065a

• feat(llama.cpp): First pass at full port of granite deviations from llama

  Show description for ec13f29
  gabe-l-hart

  committed
  ec13f29

  Copy full SHA for ec13f29

• fix(llama.cpp): Determine granite language 3b instruct by vocab size

  Show description for e73d795
  gabe-l-hart

  committed
  e73d795

  Copy full SHA for e73d795

• fix(convert_hf_to_gguf): Use LlamaModel as base for GraniteModel

  Show description for 8086380
  gabe-l-hart

  committed
  8086380

  Copy full SHA for 8086380

• fix(llama.cpp): Switch Granite param names to use _scale for consistency

  Show description for 0bdf04e
  gabe-l-hart

  committed
  0bdf04e

  Copy full SHA for 0bdf04e

• fix(convert_hf_to_gguf/gguf-py): _multiplier -> _scale

  Show description for 65c5bb9
  gabe-l-hart

  committed
  65c5bb9

  Copy full SHA for 65c5bb9

• fix(llama.cpp): Use separate switch clause for granite in llm_load_hparams

  Show description for 5d054a4
  gabe-l-hart

  committed
  5d054a4

  Copy full SHA for 5d054a4