launch¶

The entrypoint inside a distributed training job.

A static, cgo-free launcher that wraps torchrun: it derives the rendezvous identity from the restart attempt, turns a preemption signal into "checkpoint now", watches for hangs and GPU faults, and records why a run ended so the supervisor outside the job can act on it. Part of the phpboyscout Go toolkit.

Documentation¶

The documentation follows the Diátaxis framework:

• Tutorials: learning-oriented lessons.
• How-to guides: task-oriented recipes for specific goals.
• Explanation: how launch works, and why.
• Reference: the API, configuration and file formats.