IREE: Intermediate Representation Execution Environment
IREE, pronounced as "eerie," stands for Intermediate Representation Execution Environment. It is an advanced compiler and runtime framework designed for machine learning models. IREE uses MLIR (Multi-Level Intermediate Representation) technology to translate ML models into a common intermediate representation (IR), which can be efficiently executed in various environments. This adaptability makes IREE suitable for large-scale data centers as well as for mobile and edge devices with more constrained resources.
Project Status
Currently, IREE is in the early stages of development. While the basic infrastructure of the project is established, active work continues on enhancing different software components and improving the project's logistics. It is not yet considered ready for everyday use. However, the team encourages feedback through various communication channels to help steer the project forward.
Release Status
IREE offers several packages with different release statuses:
- GitHub Release (stable): The stable release is the most reliable, but it may not include the latest features.
- GitHub Release (nightly): This release includes the latest changes but may be less stable.
- Python iree-compiler: Available on PyPI, this package facilitates compiling ML models using IREE.
- Python iree-runtime: Also available on PyPI, this package handles runtime execution of compiled models.
Build Status
IREE is actively built and tested across multiple platforms to ensure broad compatibility. These platforms include Linux (x64 and ARM64), macOS, and Windows. Each platform has its own build workflows to allow independent updates and efficiency checks.
Communication Channels
The IREE team engages with the community through various channels:
- GitHub Issues: For tracking feature requests, bugs, and other work-related tasks.
- IREE Discord Server: A platform for daily development discussions with the core team and collaborators.
- Email Lists: For announcements and technical discussions, including
iree-announce,iree-technical-discussion, and the legacyiree-discuss.
Related Project Channels
Since IREE heavily relies on MLIR, the MLIR community within the LLVM project is also a valuable resource. Discussions around MLIR can provide additional insights into IREE's capabilities and future directions.
Architecture Overview
IREE’s architecture is designed to efficiently handle the transitioning of ML models from high-level abstractions to low-level executable code. This includes optimizations specific to different deployment environments.
Presentations and Talks
The project team has shared various presentations and talks related to IREE's development and design. These resources are available for viewing on the IREE YouTube channel. Notable presentations include discussions on the design of the IREE runtime and insights into its Intermediate Representation.
License
IREE is released under the Apache 2.0 License with LLVM Exceptions. This open-source license allows for broad usage and modification, which encourages community participation and innovation.
For more details, user guides, and instructions for building from source, visit the IREE website.