Iree

IREE, short for Intermediate Representation Execution Environment, is an MLIR based compiler and runtime that lowers machine learning models to a unified intermediate representation. The same toolchain is meant to scale from datacenter workloads down to the constraints of mobile and edge deployments, with binary sizes as small as 30KB on embedded systems. The project documents a four step workflow: import a model from a supported framework, select a deployment configuration, compile the model, and run it with IREE's runtime components. Supported import sources are JAX, ONNX, PyTorch, and TensorFlow with TensorFlow Lite. The compiler tools run on Linux, Windows and macOS, and can target Linux, Windows, macOS, Android, iOS, bare metal and, experimentally, WebAssembly, across ARM, x86 and RISC V architectures. Accelerator and API support covers Vulkan, ROCm and HIP, CUDA and Metal, with AMD AIE and WebGPU marked experimental. Key features include ahead of time compilation in which scheduling and execution logic are compiled together, support for dynamic shapes, flow control and streaming, low overhead pipelined execution, and debugging and profiling support including Tracy. The site also publishes reference material for API bindings, MLIR dialects and passes, and design documents for contributors. It is aimed at ML engineers and systems developers who need to deploy trained models efficiently to diverse hardware. The site is project documentation hosted alongside the iree org repository on GitHub. The vendor does not publish pricing or paid plans.