Installation Instructions

As a pre-requisite, make sure you have ducttape and (mini)conda installed.

First, clone this repository.

Then, to create a new conda environment with all the necessary dependencies, run the following command:

export CONDA_HOME="/path/to/(mini)conda3"
bash setup/conda.sh

Training

Data format

Before training, you must preprocess the training data. Before preprocessing, the data should be a json file, with the following format:

{"text": "<instance_0_text>"}
{"text": "<instance_1_text>"}

Note that the preprocessing script will pack observations together in vectors of a specified length, and will separate each instance (json line) by the tokenizer's EOS token.

Then, run the bash scripts in this order:

./preprocess_data.sh [OPTIONS]
./convert2megatron.sh [OPTIONS]
./model_sharding.sh [OPTIONS]
./continue_pretraining.sh [OPTIONS]

NOTE: each of these commands may be run with flag --help, which will inform the user on how to use each argument.

For example, for a continued pretraining run with Llama 2 7B on datasets d1 and d2 and 8 GPUs, run the following:

> ./preprocess_data.sh --dataset_json=<path_to_d1> --dataset_bin=<d1_output_path> --vocab_file=<path_to_hf_model>/tokenizer.model --repo=<path_to_repo>
> ./preprocess_data.sh --dataset_json=<path_to_d2> --dataset_bin=<d2_output_path> --vocab_file=<path_to_hf_model>/tokenizer.model --repo=<path_to_repo>
> ./convert2megatron.sh --megatron_model=<megatron_model_path> --model_path=<path_to_hf_model> --size=7 --repo=<path_to_repo>
> ./model_sharding.sh --megatron_model=<megatron_model_path> --sharded_model=<sharded_model_path> --tp=8 --pp=1 --vocab_size=32000 --repo=<path_to_repo>
> ./continue_pretraining.sh --data_path="1 d1 1 d2" --megatron_model=<sharded_model_path> --model_dir=<checkpoint_save_dir> --tokenizer_path=<path_to_hf_model>/tokenizer.model --tp=8 --pp=1 [TRAINING_ARGS]