Allensmile / PaLM-colossalai

Scalable PaLM implementation of PyTorch

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Pathways Language Model (PaLM) based on PyTorch

A PyTorch implementation of the model architecture of Pathways Language Model (PaLM): Scaling to 540 Billion Parameters for Breakthrough Performance. We take advantage of Colosssal-AI to exploit multiple optimization strategies, e.g. data parallelism, tensor parallelism, mixed precision & ZeRO, to scale the training to multiple GPUs.

You are very welcome to contribute in any way to help us enhance the usability of this project.

Preparation

  1. Install requirements, e.g. Colosssal-AI, which is a Pytorch-based large-scale model training system with various efficient parallelization techniques.
pip install -r requirements.txt
  1. Use HuggingFace datasets to download Wikitext-2 dataset. The placeholder /PATH/TO/DATA is optional and is ./wiki_dataset by default.
python ./tools/download_wiki.py -o </PATH/TO/DATA>
  1. Download tokenizer files by calling the following command. The place holder /PATH/TO/TOKENIZER/ is optional and is ./token by default.
bash ./tools/download_token.py </PATH/TO/TOKENIZER/>

Usage

  1. Configure your settings in CONFIG_FILE.py like below. We also provide some examples in ./configs
SEQ_LENGTH = 512
BATCH_SIZE = 8
NUM_EPOCHS = 10

parallel = dict(
    tensor=dict(mode='1d', size=2),
)

model = dict(type="palm_small")
  1. Set dataset & tokenizer paths
export DATA=</PATH/TO/DATA/>
export TOKENIZER=</PATH/TO/TOKENIZER/>
  1. Run
torchrun --nproc_per_node NUM_GPUS \
    train.py --from_torch --config CONFIG_FILE.py
  1. Run With Docker

    Dockerfile is provided in this repository and you can run PaLM in Docker with the following commands.

# build docker image
docker build -t palm .

# exec training
docker run -ti --gpus all --rm palm \
    torchrun --nproc_per_node NUM_GPUS \
        train.py --from_torch --config CONFIG_FILE.py

About

Scalable PaLM implementation of PyTorch


Languages

Language:Python 98.0%Language:Dockerfile 1.4%Language:Shell 0.6%