This repo contains the model architecture, training scripts, and utilities of 1.5-Pints and 0.12-Pint, developed by Pints.AI. By providing access to the model's codebase and architecture, this initiative seeks to facilitate the replication, experimentation, and further open-source development of Pint.
Join us at Discord: https://discord.com/invite/RSHk22Z29j
@misc{tan202415pintstechnicalreportpretraining,
title={1.5-Pints Technical Report: Pretraining in Days, Not Months -- Your Language Model Thrives on Quality Data},
author={Calvin Tan and Jerome Wang},
year={2024},
eprint={2408.03506},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2408.03506},
}Typically just stick to Ubuntu 22.04 LTS x86-64. Debian 12 has been tested to work as well.
GOTCHA1: Dont use arm64 / aarch64. xformers does not support ARM64 processors.
GOTCHA2: We should not install system-wide CUDA using apt. It is best to constrain the CUDA installation to within the conda environment, so that different projects can use different CUDA versions.
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh && \
sh Miniconda3-latest-Linux-x86_64.shSource just to be sure conda cli will be available:
source ~/.bashrcSometimes if you still face conda: command cannot be found, you can find the installation and source it:
Note: This path assumes you took up the default installation settings. Otherwise, find where you installed it.
source ~/miniconda3/etc/profile.d/conda.shgit clone https://github.com/Pints-AI/1.5-Pints.git && \
cd 1.5-Pintsconda create --prefix ./.conda python=3.10 && \
conda activate ./.condaNote: Stick to Python 3.10. 3.12 breaks a lot of things as of now (23 Feb 2024), and 3.11 has not been tested.
conda install nvidia/label/cuda-12.1.1::cuda-toolkitpip install -r requirements.txt && \
pip install flash-attn --no-build-isolation && \
pip install -r pretrain/requirements.txtNote: The pip install for dropout_layer_norm can take up ~30 minutes to build depending on the machine.
curl -s https://packagecloud.io/install/repositories/github/git-lfs/script.deb.sh | sudo bash
sudo apt-get install git-lfscd /path/to/dataset_dir
git clone https://huggingface.co/datasets/pints-ai/Expository-Prose-V1python -m prepare_dataset.standard_parquet \
--source_path /path/to/dataset_dir \
--train_val_split_ratio 0.9 \
--max_cores 60 \
--destination_path /path/to/output_dirRefer to prepare_dataset folder for the dataset preparation scripts.
max_cores is not required if you don't OOM on high core machines.
fabric run \
--accelerator=cuda \
--devices=8 \
pretrain/main.py \
--data_dir data/output \
--out_dir ../1.5-pints \
--gpus 8 \
--global_batch_size 512 \
--learning_rate 4e-4 \
--micro_batch_size 8 \
--max_step 96180 \
--warmup_steps 2000 \
--weight_decay 0.1 \
--beta1 0.9 \
--beta2 0.95 \
--grad_clip 1.0 \
--min_lr 4e-5 \
--model_name 1.5-Pints-2k \
--wandb_name <run_name> \
--wandb_project <project_name> \
--tokenizer_dir tokenizer/pintsNote1: --devices and --gpus must be the same. See pretrain.py's setup arguments for all parameters that you can adjust.
Note2: Select the architecture (layers/dimensions/heads) configuration using --model_name. This must be in lit_gpt/config.py.
Note3: Select a micro_batch_size to optimize GPU memory. So far once started, it remains stable, even during validation. micro_batch_size need not be a number that batch_size is divisible by. batch_size is derived from global_batch_size / devices.
Note4: Modify TRAIN_DATA_CONFIG in pretrain/main.py to decide on the datasets used for training. Ensure that the dataset is prepared beforehand.
If you are asked for the wandb API key, you can login and get from: https://wandb.ai/authorize
cd finetune && \
pip install -r requirements.txtBefore you start finetuning, you need to convert the pretrain weights:
python convert/convert_pretrained_checkpoint.py --checkpoint_dir path/to/checkpoint --output_dir path/to/outputlightning run \
--accelerator=cuda \
--devices=8 \
finetune/full.py \
--checkpoint_dir <path to lit_model.pth> \
--out_dir ~/1.5-pints-2k/ep2/step-00045000/finetuned \
--model_name 1.5-Pints-2k \
--gpus 8 \
--train.save_interval 6000 \
--train.global_batch_size 512 \
--train.micro_batch_size 8 \
--train.lr_warmup_steps 1125 \
--train.epoch 5 \
--train.learning_rate 2e-5 \
--train.max_seq_length 2048 \
--train.beta1 0.9 \
--train.beta2 0.95 \
--train.weight_decay 0.1 \
--logger_name wandb \
--tokenizer_dir tokenizer/pints \
--known_data_max_seq_length 2048 \
--wandb_project <project name>DPO is opted for use post-finetuning. See here for the execution process.
See here
python convert_lit_to_hf.py \
--checkpoint_name lit_model.pth \
--directory ../models/1.5-pints \
--model_name 1.5-Pints-2k \
--output_config=True \
--safetensors=True \
--delete_pytorch_model=TrueNote: We found better success using the safetensors file. Therefore it's recommended to use it instead of pytorch_model.bin.
from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("/model/path")
model = AutoModelForCausalLM.from_pretrained("/model/path")
prompt = '''<|im_start|>user
Do not go gentle into that good night.<|im_end|>
<|im_start|>assistant
'''
tokenized_input = tokenizer.encode(prompt)
tokenized_output = model.generate(tokenized_input)
print(tokenizer.decode(tokenized_output))This codebase comes with tests. If you need to make modifications, you can run the tests to ensure your modifications did not disrupt the existing code.
Install test requirements:
pip install -r requirements.test.txtRun pytest:
python -m pytest --verbose