/moe-recipes

Mixtre of Experts Library forked from kotoba-recipes

Primary LanguagePython

Kotoba Recipes

Table of Contents

  1. Installation
  2. Instruction Tuning
  3. LLM Continual Pre-Training
  4. Support Models

Installation

To install the package, run the following command:

pip install -r requirements.txt

If you want to use the library in multi-nodes, you need to install the below packages:

module load openmpi/4.x.x

pip install mpi4py

FlashAttention

To install the FlashAttention, run the following command: (GPU is required)

pip install ninja packaging wheel
pip install flash-attn --no-build-isolation

ABCI

If you use ABCI to run the experiments, install scripts are available in kotoba-recipes/install.sh.

Instruction Tuning

scripts/abci/instruction contains the scripts to run instruction tunings on ABCI.

If you want to use custom instructions, you need to modify the src/llama_recipes/datasets/alpaca_dataset.py.

LLM Continual Pre-Training

scripts/abci/ contains the scripts to run LLM continual pre-training on ABCI. 7B, 13B, 70B directories contain the scripts to run the experiments with the corresponding model size (Llama 2).

Support Models