tvm: an open machine learning compilation framework for universal model deployment and joint optimization

What it solves

Apache TVM is an open machine learning compilation framework designed to make ML compilers accessible. It solves the problem of deploying models into minimum deployable modules across various hardware, while allowing developers to customize compiler pipelines through a Python-first approach.

How it works

TVM uses a cross-level design featuring TensorIR for tensor-level representation and Relax for graph-level representation. This allows the framework to jointly optimize computational graphs, tensor programs, and libraries. It is designed as a foundation infrastructure for building vertical compilers for specific domains, such as LLMs.

Who it’s for

It is intended for developers and researchers who need to optimize and deploy machine learning models across different hardware targets, as well as those building domain-specific compilers for AI.

Highlights

• Python-first development for quick customization of compiler pipelines.
• Universal deployment capabilities to create minimum deployable modules.
• Cross-level representation (TensorIR and Relax) for joint optimization.
• Serves as a foundation for building vertical compilers for domains like LLMs.