Machine-Learning Accelerator System Exploration Tools

Overview

Machine learning accelerators have been used extensively to compute models with high performance and low power. Unfortunately, the development pace of ML models is much faster than the accelerator design cycle, leading to frequent changes in the hardware architecture requirements, rendering many accelerators obsolete. Existing design tools and frameworks can provide quick accelerator prototyping, but only for a limited range of models that fit into a single hardware device. With the emergence of large language models such as GPT-3, there is an increased need for hardware prototyping of large models within a many-accelerator system to ensure the hardware can scale with ever-growing model sizes.

MASE provides an efficient and scalable approach for exploring accelerator systems to compute large ML models by directly mapping onto an efficient streaming accelerator system. Over a set of ML models, MASE can achieve better energy efficiency to GPUs when computing inference for recent transformer models.

MASE Publications

Fast Prototyping Next-Generation Accelerators for New ML Models using MASE: ML Accelerator System Exploration, link

@article{cheng2023fast,
title={Fast prototyping next-generation accelerators for new ml models using mase: Ml accelerator system exploration},
author={Cheng, Jianyi and Zhang, Cheng and Yu, Zhewen and Montgomerie-Corcoran, Alex and Xiao, Can and Bouganis, Christos-Savvas and Zhao, Yiren},
journal={arXiv preprint arXiv:2307.15517},
year={2023}}

MASE: An Efficient Representation for Software-Defined ML Hardware System Exploration, link

@article{zhangmase,
title={MASE: An Efficient Representation for Software-Defined ML Hardware System Exploration},
author={Zhang, Cheng and Cheng, Jianyi and Yu, Zhewen and Zhao, Yiren}}

Repository structure

This repo contains the following directories:

components - Internal hardware library
scripts - Installation scripts
machop - MASE's software stack
hls - HLS component of MASE
mlir-air - MLIR AIR for ACAP devices
docs - Documentation
Docker - Docker container configurations

Getting Started with Docker

First, make sure the repo is up to date:

make sync

Start with the docker container by running the following command under the repo:

make shell

It may take long time to build the docker container for the first time. Once done, you should enter the docker container. To build the tool, run the following command:

cd /workspace
make build

This should also take long time to finish.

If you would like to contribute, please check the wiki for more information.

Getting Started with Conda

First, make sure the repo is up to date:

make sync

Install conda by following the instructions here, then build and activate the environment as follows.

conda env create -f machop/environment.yml
conda activate mase

Optionally, you can verify the CLI utility is running.

./machop/ch --version

Running an example neural network

In this example, we'll use the command-line interface (CLI) to train a toy model on the Jet Substructure (JSC) dataset, quantize it to integer arithmetic and evaluate the quantized model on the test split of the dataset.

First, train the toy model over 10 epochs by running the following command.

./machop/ch train jsc-tiny jsc --max-epochs 10 --batch-size 256

Now, transform the model to integer arithmetic and save the model checkpoint. The toml configuration file specifies the required arguments for the quantization flow.

./machop/ch transform jsc-tiny jsc --config ./machop/configs/examples/jsc_toy_by_type.toml

Finally, evaluate the quantized model performance.

./machop/ch test jsc-tiny jsc --load <path/to/checkpoint>

See the Machop README for a more detailed introduction.

MASE Dev Meetings

Subscribe Mase Weekly Dev Meeting (Wednesday 4:30 UK time). Everyone is welcomed!
Direct Google Meet link
Join the Mase Slack
If you want to discuss anything in future meetings, please add them as comments in the meeting agenda so we can review and add them.

Donation

If you think MASE is helpful, please donate for our work, we appreciate your support!

bakhtiarZ / mase