Execution

Turning a program into artifacts, and running those artifacts on a device.

Everything up to this chapter describes what you write. This one describes what happens to it: ir.compile and JITFunction.compile produce a CompiledProgram, and ChipWorker dispatches it. Two knobs decide most of what you will care about — where the artifacts go, and which of them the runtime is allowed to reuse between launches.

Page	Covers
Compiling	ir.compile and its parameters, JITFunction.compile, the artifact directory, pass dumps
Running	CompiledProgram's contract, ChipWorker, DeviceTensor, and the RunConfig fields that affect dispatch

Where PyPTO stops

PyPTO produces the artifacts and hands them to the simpler runtime, which owns scheduling, the task rings, and the device lifecycle. This chapter documents the PyPTO side of that boundary only: the API you call, what it produces, and which of its fields the runtime reads.

For the mechanisms behind the boundary — how tasks are scheduled, how dependencies are resolved, what the rings do — see the runtime documentation. Memory covers the part of that machinery you can tune from here.

See Also

• Getting started — the shortest path from a kernel to a result.
• Tools — what to reach for when the result is not what you expected.
• Performance — measuring and tuning what this chapter launches.