Intro to MAX Graph
The MAX Graph API allows you to build your entire AI inference pipeline in Mojo. By building the graph with MAX Graph, you can combine the performance and programmability of Mojo with the cutting-edge compiler technologies and hardware portability provided by MAX Engine. This combination allows you to build highly performant graphs, with less code that’s more readable than code from other performance-based graph libraries in C or C++.
We built MAX Graph because, although MAX Engine can execute models from frameworks such as PyTorch and TensorFlow faster than the default runtimes, sometimes the high-level abstractions still leave performance on the table. With the MAX Graph API, we give you full control of the graph at a lower level.
Overview
The Graph API is a low-level Mojo API that allows you to create an acyclic
graph (as a Graph) using ops
provided in max.graph.ops or using
custom ops that you create in Mojo.
All ops are connected to each other with
Symbol values, which define the
edges in the graph. A Symbol represents a value that doesn't actually exist
until execution time. It could be a graph input, a constant, or the output of
an op.
To add ops to your graph, you chain a sequence of op functions (each function
takes a Symbol—an input, a constant, or other op output), and then add that
sequence to the graph by passing the final op Symbol to
Graph.output().
You can think of a complete Graph like a function definition, except it's a
static computation graph, not an eager execution graph. That means you cannot
execute the graph until you compile it with MAX Engine.
The Graph API is not designed to train models. Currently, you must load the weights into the model as constants.
Get started
To learn more, see the MAX Graph tutorial or the Graph API examples in GitHub.
Also see how you can implement custom ops in MAX Graph.