Pipelines run Jobs as stages: a stage starts when the stages it depends on have succeeded and reads their
declared outputs at /nodus/inputs/<name>; the whole Pipeline shares one maxCostUSD. See
Pipelines.
Sweeps run one Job once per combination of GPU types, region classes and parameters, each a Job named
<sweep>-<index>, and report every cell’s cost, wall time and throughput with the cheapest and the fastest
cell. See Sweeps.
Indexed Jobs (--completions N) run N indexes in parallel, each with NODUS_INDEX; nodus logs --all-indexes reads every index and nodus cp job/x:outputs/NAME --index I downloads one index’s output.
Outputs written under /nodus/outputs are committed when the run succeeds, listed file by file in
status.outputs and downloadable with nodus cp, by a declared name or by path. Another Job’s output, a file
or a whole directory, is an input (inputs[].output). See Outputs.
A catalog image without a tag (nodus/python, nodus/pytorch) now means its default tag.