This repository contains the training code of MobileLLM introduced in our work: "MobileLLM: Optimizing Sub-billion Parameter Language Models for On-Device Use Cases", published in ICML 2024.

In this work, we comprehensively consider multiple design factors to obtain high-quality LLMs with fewer than a billion parameters. We integrated (1) SwiGLU activation function, (2) deep and thin architectures, (3) embedding sharing, (4) grouped-query attention to build MobileLLM. MobileLLM-125M/350M attains a remarkable 2.7%/4.3% accuracy boost over preceding 125M/350M SoTA models on zero-shot commonsense reasoning tasks. In our updated version, we further demonstrate that our design philosophy scales effectively to larger models, with SoTA results for MobileLLM-600M/1B/1.5B.

If you find our code useful for your research, please consider citing:

@article{liu2024mobilellm,
    title={MobileLLM: Optimizing Sub-billion Parameter Language Models for On-Device Use Cases},
    author={Liu, Zechun and Zhao, Changsheng and Iandola, Forrest and Lai, Chen and Tian, Yuandong and Fedorov, Igor and Xiong, Yunyang and Chang, Ernie and Shi, Yangyang and Krishnamoorthi, Raghuraman and others},
    journal={arXiv preprint arXiv:2402.14905},
    year={2024}
}

• python 3.9, pytorch >= 2.0
• pip install -r requirement.txt

Dividing a tokenized dataset or tokenize your own dataset, and even distribute it across the total number of training nodes, where each node comprises 1x8 GPUs. Next, organize the data into the following structure:

• basepath
  • 1
    • xxx.jsonl
  • 2
    • xxx.jsonl
  • ...
  • #nodes
    • xxx.jsonl

Each line of a jsonl file is a key-value pair of tokenized data {"token_ids": [1,2,3,4,...]}.

Our training code is compatible with the data pre-processing method in https://github.com/LLM360/amber-data-prep.

The script pretrain.sh is provided to initiate training on a 1x8 node setup using torchrun. This script can be modified to adjust the --nnodes parameter and other settings to suit different multi-node configurations, such as those using slurm or torchx. The learning rate in the script is for 1x8 node with a batch size of 32. If you increase the number of nodes or the batch size, you need to increase the learning rate linearly.

Steps to run:

• In pretrain.sh file, specify the --train_data_local_path to the pre-processed data in Step 2 and --input_model_filename to ./configs/{model_size}/.
• Run bash pretrain.sh

The model weights is still under legal review. If you have any questions, feel free to email (zechunliu at meta dot com) and (cszhao at meta dot com)

It takes the following number of days to train MobileLLM on 1T tokens using 32 NVIDIA A100 80G GPUs.

This code is partially based on Hugging Face transformer repo.

Zechun Liu, Meta Inc (zechunliu at meta dot com)

Changsheng Zhao, Meta Inc (cszhao at meta dot com)

BiT is CC-BY-NC 4.0 licensed as of now.