Model Details

MobileLLM is introduced: "MobileLLM: Optimizing Sub-billion Parameter Language Models for On-Device Use Cases", published in ICML 2024.

Model Developer: Meta

Model Architecture: MobileLLM is an auto-regressive language model leveraging an optimized transformer architecture, specifically engineered for on-device applications with constrained resources. MobileLLM integrated several key techniques including: (1) SwiGLU activation function, (2) deep and thin architectures, (3) embedding sharing, (4) grouped-query attention. 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.

How to use

We are providing 2 ways to run the model:

HuggingFace

MobileLLM codebase

HuggingFace

To load the pretrained model for further finetuning or evaluation:

from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("facebook/MobileLLM-1B", use_fast=False)
model = AutoModelForCausalLM.from_pretrained("facebook/MobileLLM-1B", trust_remote_code=True)

Note that the default tokenizer does not contain special tokens. For example you can use:

tokenizer.add_special_tokens(
    {
        "eos_token": "</s>",
        "bos_token": "<s>",
        "unk_token": "<unk>",
    }
)

MobileLLM codebase

We provide the pretraining code in https://github.com/facebookresearch/MobileLLM

> git clone https://github.com/facebookresearch/MobileLLM
> pip install -r requirement.txt

# data pre-process and specify the data path in pretrain.sh
# run pretraining
> bash pretrain.sh 

We also provide evaluation script for calculating ppl of wikitext-2 test split:

> bash eval.sh

You can find more details in the GitHub repo.

Training cost

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

Evaluation

We evaluate the pretrained MobileLLM models on Zero-shot Common Sense Reasoning tasks

MobileLLM-125M

MobileLLM-350M

MobileLLM-600M

MobileLLM-1B

MobileLLM-1.5B

Acknowledgement

This code is partially based on HuggingFace Transformers repo under Apache License.

Citation

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}
}

License

MobileLLM is FAIR NC licensed as of now