iree-org/iree

A retargetable MLIR-based machine learning compiler and runtime toolkit.

What it solves

IREE simplifies the deployment of machine learning models across a wide variety of hardware. It addresses the challenge of scaling ML models to work efficiently on everything from massive data centers to resource-constrained mobile and edge devices.

How it works

IREE is an end-to-end compiler and runtime based on MLIR (Multi-Level Intermediate Representation). It works by lowering ML models into a unified intermediate representation (IR) that can then be optimized and executed on diverse hardware targets.

Who it’s for

It is designed for developers and engineers who need to deploy ML models to diverse hardware environments, including data centers, mobile devices, and edge deployments.

Highlights

  • MLIR-based compiler and runtime architecture.
  • Unified IR for cross-platform deployment.
  • Supports scaling from data center scale to mobile and edge constraints.
  • Part of the LF AI & Data Foundation.

Related

  • Project
    buddy-compiler/buddy-mlir

    Buddy MLIR is an MLIR‑based compiler framework for co‑designing AI models and custom hardware. It lets you compile models like DeepSeek, Whisper, and Qwen3‑VL into a portable `.rax` binary and run them with a tiny CLI runtime on CPUs, RISC‑V, or GPUs.

  • Project
    google/heir

    An MLIR-based compiler toolchain for homomorphic encryption that allows developers to write encrypted computations and deploy them across multiple HE libraries and schemes.

  • Project
    sophgo/tpu-mlir

    An MLIR-based machine-learning compiler that converts pre-trained neural networks from frameworks like PyTorch and ONNX into optimized bmodel files for TPU execution.

  • Project
    NVIDIA/cuda-tile

    CUDA Tile IR is an MLIR‑based compiler infrastructure from NVIDIA that introduces a dedicated dialect for tile‑oriented CUDA kernels. It provides C++ and Python APIs, a binary bytecode format, and tools to translate MLIR to bytecode or cubin, enabling high‑performance kernels that target NVIDIA Tensor‑Core units. The project builds on LLVM/MLIR, aligns with CUDA Toolkit 13.1, and includes a conformance test suite. It is a genuine AI‑related compiler project, not a tutorial or list.

  • Project
    llvm/torch-mlir

    A compiler project that provides a standardized bridge from the PyTorch ecosystem to MLIR, simplifying how hardware vendors deploy PyTorch models to target hardware.