Mojo dentro del grafo de MAX

Inside MAX, a model is represented as a graph: a network of operations where data flows from one node to the next. 🔗

Each node in the graph is an op, a single operation like a matrix multiply or an activation function applied to its inputs.

MAX ships with many built-in ops covering common layers, so most of a typical model graph works without any extra code.

Sometimes your model needs an operation MAX does not include. That is when you write a custom op to fill the gap.

You write custom ops in Mojo, giving you systems-level speed for the exact computation your model requires.

A custom op is essentially a Mojo function that takes input tensors and produces an output tensor for the graph.

Because the op is Mojo, you can use SIMD and other tricks to make that node run as fast as the hardware allows.

Once defined, your custom op is registered so MAX can place it in the graph just like any built-in operation.

MAX still optimizes the full graph, including your custom node, fusing and scheduling work for the best overall throughput.

This pattern lets you keep a model mostly standard while dropping in a hand-tuned Mojo op exactly where you need speed.

Reach for a custom op when a missing or slow node is your bottleneck, not for parts the built-ins already handle well. ✨

Recall how Mojo connects to the MAX computation graph.

You saw that a MAX model is a graph of ops, and Mojo lets you write fast custom ops that slot right into it. 🎯