Users of MLIR

This page collects publicly documented projects that use MLIR as compiler infrastructure or define MLIR dialects and transformations. Entries are alphabetized by project name. Archived or superseded projects are labeled.

Know of a project that is missing, or one whose description has changed? Please submit an update to this page.

Accera

Accera is a research compiler and Python library from Microsoft Research for scheduling compute-intensive loop nests. It translates programs through MLIR pipelines to produce optimized binaries for target platforms.

Allo

Allo is a Python-embedded language and compiler for accelerator design. It uses MLIR to represent modular kernels, compose schedules, and lower designs to hardware and software backends.

AscendNPU IR

AscendNPU IR is an MLIR-based intermediate-representation system for Huawei Ascend NPUs. It provides dialects, transformations, and interfaces for describing and optimizing AI workloads.

Beaver

Beaver provides MLIR and LLVM tooling for Elixir and Zig. It uses Elixir’s metaprogramming facilities to make defining and manipulating MLIR dialects and operations more accessible.

BTOR2MLIR

BTOR2MLIR is an MLIR dialect and toolchain for the BTOR2 hardware-verification format. It enables BTOR2 models to reuse compiler transformations and software verification backends.

Buddy MLIR

Buddy MLIR is a compiler framework that connects domain-specific languages to domain-specific architectures. It provides MLIR dialects, transformations, runtime components, and examples for heterogeneous systems.

ByteIR

ByteIR is an MLIR-based compiler for CPUs, GPUs, and ASICs. It provides compiler passes, frontends, and runtime components for lowering machine-learning workloads across heterogeneous targets.

Catalyst

Catalyst is an experimental just-in-time compiler for hybrid quantum-classical PennyLane programs. Its MLIR-based compiler represents quantum operations in dedicated dialects and lowers programs through LLVM and QIR while supporting control flow and automatic differentiation.

CIRCT

CIRCT is an application of MLIR and LLVM methodology to hardware design. It provides reusable dialects, transformations, and tools for hardware compilers and electronic-design-automation workflows.

ClangIR

ClangIR (CIR) introduces a high-level MLIR representation between Clang’s AST and LLVM IR. It preserves more C and C++ semantics for analysis and transformation before lowering to LLVM IR.

Concrete

Concrete is an open-source compiler for programs that operate on fully-homomorphic-encryption data. Its Python interface lowers encrypted programs through MLIR and LLVM to implementations based on TFHE.

CUDA-Q

CUDA-Q is a heterogeneous quantum-classical programming platform. Its compiler uses MLIR dialects for quantum and classical computation before lowering to QIR, LLVM IR, or target-specific representations.

DSP-MLIR

DSP-MLIR is an experimental MLIR dialect and compiler for digital signal processing. It represents common DSP operations and applies domain-specific rewrites before lowering them through MLIR.

Enzyme

Enzyme provides automatic differentiation across multiple compiler levels. EnzymeMLIR exposes first-class forward- and reverse-mode differentiation in MLIR. Projects such as Enzyme-JAX use it in MLIR-based compiler pipelines.

EUDSL

EUDSL is an experimental LLVM project for building embedded domain-specific languages that target MLIR. It includes language bindings, code generators, and utilities for constructing MLIR-based frontends and tools.

Firefly (archived)

Firefly was a compiler and runtime for Erlang and Elixir. It used MLIR for its final optimization and code-generation stages, targeting native code and WebAssembly. The project was archived in June 2024.

Flang

Flang is LLVM’s Fortran frontend. Its FIR and HLFIR dialects use MLIR to represent, analyze, optimize, and lower Fortran programs before LLVM code generation.

HEIR

HEIR is an open-source MLIR toolchain for homomorphic encryption. It defines dialects and transformations spanning high-level tensor programs, encryption-scheme representations, and target cryptographic libraries.

Hexagon-MLIR

Hexagon-MLIR is an open-source compiler stack for Qualcomm Hexagon NPUs. It uses MLIR to lower Triton kernels and PyTorch models through Hexagon-specific dialects, transformations, and runtime interfaces.

Intel Extension for MLIR (IMEX)

IMEX develops MLIR dialects, transformations, and runtime integrations for Intel CPUs and GPUs. It also serves as a staging ground for components intended for upstream MLIR.

IREE

IREE is an end-to-end MLIR-based compiler and runtime for machine-learning models. Its retargetable stack scales from data-center systems to mobile and edge devices across a range of accelerator backends.

JSIR

JSIR is an MLIR-based high-level representation for JavaScript analysis and lossless source-to-source transformation. Google uses it for tasks including decompilation and deobfuscation.

LAPIS

LAPIS is an MLIR-based compiler for linear-algebra workloads that targets Kokkos and other performance-portability programming models. It can integrate with Torch-MLIR to compile models originating in PyTorch.

LingoDB

LingoDB is an MLIR-based query compiler for relational and other data-intensive workloads. It uses multiple dialects and transformation stages to optimize and JIT-compile database queries.

MARCO

MARCO is an experimental Modelica compiler. It represents Base Modelica programs with an MLIR dialect and lowers them to executable code and supporting runtime components.

MLIR-AIE (IRON)

MLIR-AIE provides an MLIR toolchain and the IRON Python API for AMD Ryzen AI NPUs and Versal AI Engines. It models compute tiles, data movement, and device configuration and lowers programs to deployable device artifacts.

MLIR-AIR

MLIR-AIR provides dialects, tools, and libraries for asynchronous, hierarchical accelerator programming. It models data movement and execution on spatial accelerators and integrates with MLIR-AIE for AMD AI Engine targets.

MLIR-DaCe

MLIR-DaCe connects MLIR with DaCe’s Stateful DataFlow Graph representation. Its data-centric dialect and conversion infrastructure enable optimization across the two ecosystems.

MLIR-EmitC (archived and superseded)

MLIR-EmitC developed conversions from MLIR dialects to C and C++. The project was archived in December 2024 and superseded by the upstream EmitC dialect and emitter.

MLIR-TensorRT

MLIR-TensorRT is a compiler and runtime for executing MLIR programs with TensorRT and GPU code-generation backends. It accepts StableHLO and other MLIR dialects, partitions supported work for TensorRT, and lowers remaining operations through fallback pipelines.

Mojo

Mojo is an open-source, Pythonic systems programming language for high-performance heterogeneous computing. Its compiler is built with MLIR and supports interoperability with the Python ecosystem.

MQT Compiler Collection

The MQT Compiler Collection is an MLIR-based framework in MQT Core for quantum and classical compilation. It provides quantum dialects, transformations, and interoperability with formats and tools across the quantum software ecosystem.

Numba-MLIR

Numba-MLIR is a proof-of-concept Numba backend that uses MLIR for CPU and GPU code generation. It retains Numba-compatible Python decorators and frontend behavior while replacing the downstream compiler pipeline.

ONNX-MLIR

ONNX-MLIR imports ONNX models into MLIR and lowers them to optimized native binaries or libraries. Its dialects and passes support targets ranging from general-purpose processors to specialized accelerators.

OpenXLA

OpenXLA is an open machine-learning compiler ecosystem whose projects include XLA, StableHLO, and Shardy. These projects use MLIR to represent and transform models from frameworks such as JAX, TensorFlow, and PyTorch for CPUs, GPUs, and accelerators.

PhoebeDB

PhoebeDB is a PostgreSQL-compatible HTAP database system. Its query compiler uses a custom MLIR pipeline to optimize relational operations and JIT-compile them through LLVM.

PlaidML (archived)

PlaidML was a portable tensor compiler and runtime. Its experimental v1 architecture adopted MLIR for graph and kernel compilation. The project was archived in March 2025.