hexir

Hexir

Hexir is a small compiler for neural networks. You give it a graph of operations, it decides which ones run on the CPU and which on the GPU, and it turns them into code you can run.

It is built on MLIR, and it is small enough to read. That is the point: the gap between MLIR’s Toy tutorial and a production compiler like IREE is enormous, and Hexir sits in the middle.

flowchart LR
    A["your program<br/>tensors"] --> B["decide<br/>where each op runs"]
    B --> C["turn each op<br/>into a kernel"]
    C --> E["write a file<br/>.hxb"]
    E --> F["run it with<br/>hexir-run"]

To run a program, you compile it to a .hxb file and run it with hexir-run, a small program that contains no compiler at all. A GPU kernel is compiled to a CUBIN and embedded, so the runtime launches it with no compiler present.

Start here

If you want to use it, read Getting started.

If you want to understand how it works, read How it works first. It is a picture and eight paragraphs. Then pick whichever of the detailed pages you need.

Where it stands

Hexir is research software. Some parts are finished and some are scaffolding, and the docs say which is which rather than leaving you to find out.

Works	Partly	Not yet
CPU path, end to end	CPU kernels in .hxb are descriptions, not code	Fusion
GPU path, end to end		Memory planning
Per-operation placement		More operations, and a frontend

:hidden:
:maxdepth: 2
:caption: Using Hexir

getting-started

:hidden:
:maxdepth: 2
:caption: How it works

overview
ir-levels
pipeline
runtime

:hidden:
:maxdepth: 2
:caption: Optimizations

optimizations/overview
optimizations/graph
optimizations/kernel
optimizations/memory
optimizations/execution
optimizations/llm

:hidden:
:maxdepth: 2
:caption: Reference

reference/cli
reference/dialects
reference/layout
cuda-server