RunFleet

RunFleet does the rounds of squeue, sacct, nvidia-smi and tail -f for you. It polls each host over your own SSH setup and shows, in one keyboard-driven view, what is running, what is waiting and why, which GPUs are busy or idle, and what went wrong with the jobs that ended.

• Every host at a glance: your running, pending and failed jobs, the GPUs you hold, free GPUs per partition and the next job in line.
• Live GPU readings inside a job: per-GPU utilisation and memory read on the job’s own node, which catches a GPU sitting idle that the job’s average hides.
• Why jobs end and wait: ended jobs get a cause (OOM-GPU, NCCL, TIMEOUT, …) linked to the log line, and pending jobs show the priority factors that hold them back.

A --demo mode runs it on a simulated fleet, with no cluster or account needed; the picture above is that demo.

Its companion, Run-Skills, lets a coding agent launch those jobs.