neuron-monitor: a period such as "1.0s" is silently replaced by the 5 s default #1419

Summary

neuron-monitor -c <config> accepts "period": "1s" but treats "1.0s", "0.5s" and "5.0s" as if no period were set and delivers the 5 s default. No error or warning is printed, and "500ms" produces no samples at all. The user guide's examples use "1s" and "2s" and say the default is 5 seconds, but do not say which formats are valid, so a period built from a float (f"{seconds}s") fails silently.

What we measured

Environment: inf2.xlarge, Deep Learning AMI Neuron (Ubuntu 22.04), aws-neuronx-tools 2.28.23.0 (kaena-tools/2.28), Neuron runtime 2.30.51, 2026-09-11. Each sample's own period field and the number of samples in 16 s:

Config shape (system metrics group omitted for brevity):

{"period": "1.0s",
 "neuron_runtimes": [{"tag_filter": ".*", "metrics": [{"type": "neuroncore_counters"}]}],
 "system_metrics": [{"type": "neuron_hw_counters"}]}

Run with neuron-monitor -c config.json while a Neuron workload is running, and read period from each report.

Why it matters

A tool that builds the period from a float gets 5 s samples while believing it asked for 1 s, and nothing reports the gap. We hit this ourselves and misattributed it to the tool ("neuron-monitor floors at ~2 s") before comparing the requested period with the delivered one.

Suggested change

Either accept "1.0s" and similar, or fail with a message that names the accepted format; and state the accepted format (whole seconds, Ns) and the behaviour for an unparsed value in the neuron-monitor user guide.

Not checked: tools releases newer than 2.28.23.0. If the behaviour has changed, a one-line note in the release notes would be enough for us.

Our write-up with the code that stopped sending the bad string: https://github.com/pantheongpu/pantheonneuron (docs/checks_that_pass_by_accident.md, "neuron-monitor accepts whole seconds only").