mache.parallel.system.ParallelSystem.get_gpu_device_names

ParallelSystem.get_gpu_device_names() → List[str]

Name each of a node’s GPUs, in index order, as the machine does.

A placement’s gpu_ids are indices from 0 to gpus_per_node - 1, but a machine whose devices are not simply numbered from zero names them otherwise – Aurora addresses a tile as 0.1. Such a machine lists its devices in order in its gpu_bind = list:... binding list, so that is used where it is present. This is also what a launch’s ranks will see in the vendor’s visible-devices variable, so a caller checking what a launch got can compare against these names.

Returns:

names (list of str) – The name of each GPU on a node, indexed as gpu_ids are.