Accelerator Design and Programming Language
==============================================================================
Documentation | Installation | Tutorials
Allo is a Python-embedded, MLIR-based language and compiler designed to facilitate the modular and composable development of large-scale, high-performance machine learning accelerators. It provides a unified abstraction for both accelerator design and programming, enabling developers to express complex architectures in a structured and reusable manner. Allo offers several key features:
- Composable Design and Programming: Allo supports both behavioral and structural composition, allowing users to incrementally build and compose accelerator components into a complete system with minimal overhead.
- End-to-End Deployment: Allo enables automatic accelerator generation from PyTorch models and integrates tightly with a high-performance simulator and a formal verifier, streamlining validation, testing, and deployment workflows.
- Multiple Backend Support: Allo currently targets AMD and Intel FPGAs as well as AMD Ryzen NPUs (AI Engine), with planned support for GPUs and ASICs in future releases.
Getting Started
Please check out the Allo documentation for installation instructions and tutorials. If you encounter any problems, please feel free to open an issue.
IMPORTANT: If you are using a coding agent for our codebase, please import AGENTS.md.
Publications
Please refer to our PLDI'24 paper for more details. If you use Allo in your research, please cite our paper:Hongzheng Chen, Niansong Zhang, Shaojie Xiang, Zhichen Zeng, Mengjia Dai, and Zhiru Zhang, "Allo: A Programming Model for Composable Accelerator Design", Proc. ACM Program. Lang. 8, PLDI, Article 171 (June 2024), 2024.
Please also consider citing the following papers if you utilize specific components of Allo:
- Dataflow programming model: Shihan Fang, Hongzheng Chen, Niansong Zhang, Jiajie Li, Han Meng, Adrian Liu, Zhiru Zhang, "Dato: A Task-Based Programming Model for Dataflow Accelerators", arXiv:2509.06794, 2025.