Three Prominent ML Compilers
6 entriesAn ML compiler takes a model expressed as high-level operations such as convolutions and matrix multiplies, optimises the computation graph (for example by fusing operators), and generates low-level code for a specific processor. These three are the most widely cited. Descriptions follow each project's own pages and are not rankings.
Other Compiler Frameworks and Runtimes
10 entriesCompilers, kernel languages and optimising runtimes that sit alongside or beneath the three above.
Papers and Background
3 entriesWhere the ideas come from.
Related Pages
5 entriesRelated directories on this site.
Frequently Asked Questions
What is an ML compiler?
An ML compiler translates a model written as high-level operations (convolutions, matrix multiplications, attention) into low-level instructions for a specific processor. Along the way it optimises the computation graph, for example by fusing operators, choosing data layouts and selecting or generating fast kernels.
How do TensorRT, Apache TVM and XLA differ?
NVIDIA TensorRT targets NVIDIA GPUs and builds an optimised inference engine for a trained network. Apache TVM is an open source compiler framework that targets many kinds of hardware, including CPUs, GPUs and other accelerators. XLA is an open source compiler for GPUs, CPUs and ML accelerators that is used by frameworks such as JAX. These are summaries of the projects' own descriptions, not benchmarks.
What does hardware-specific code generation involve?
The compiler lowers the graph through successive intermediate representations until it reaches code for the target, such as GPU kernels or CPU vector instructions. It may tune tile sizes, memory use and precision for that device, either through built-in rules or by searching over candidate schedules.
How do MLIR and IREE relate to these compilers?
MLIR is compiler infrastructure for building multi-level intermediate representations, and several ML compilers are built on it. IREE is an MLIR-based compiler and runtime. They are listed in the second section as part of the wider compiler stack.
Is this page a ranking?
No. It is a curated directory. Each entry links to the project's own site, and the notes follow the projects' descriptions. Test any compiler on your own model and hardware before choosing.