Lecture 001

High Level Description

Machine learning compilation (MLC): process of transforming and optimizing machine learning execution from its development form to its deployment form.

Goal of MLC

dependency minimization: make code smaller, automatic fusion
leverage native hardware acceleration: support heterogeneous device
optimization

Difference from normal compilation

might not have code generation: can be model description and engine that execute it
it helps both training and inference

Reason for study:

deployment is not easy
in-depth understanding of existing framework
building software stack for emerging hardware

Lower Level Description

Tensor: object

Tensor function: operators

Fusion: fusing linear layer with relu layer

Above is one possible optimization

Abstraction: we can represent code with graph

Abstraction allow us to optimize code.

Table of Content