Command-line reference

The dashboard is launched as kempnerpulse or the alias kp — they are identical.

kempnerpulse [options]
kp [options]

Options

Flag	Default	Description
--source SOURCE	http://localhost:9400/metrics	A dcgm-exporter HTTP /metrics endpoint or a local file. For --backend replay, the path to a saved dcgmi dmon capture.
--backend {dcgm,prometheus,replay}	dcgm	Metric source. See Backends.
--poll POLL	0.1 (dcgm) / 1.0 (prometheus)	Sampling interval in seconds. The dcgm backend is honored down to a 100 ms floor; prometheus requires >= 1.0.
--history HISTORY	120	Samples kept for sparkline history.
--focus-gpu ID	—	Start in the focused single-GPU view.
--once	off	Render one snapshot and exit.
--gpus GPUS	—	Explicit GPU ids or ranges (0,1 or 0-3). Overrides the visibility environment.
--show-all	off	Ignore CUDA_VISIBLE_DEVICES / SLURM_*_GPUS and show every accessible GPU.
--weights W_SM,W_TENSOR,W_DRAM,W_GR	AI preset	Custom Real-Utilization weights (normalized to sum to 1).
--ai-weights	(default)	Preset 0.35,0.35,0.20,0.10.
--hpc-weights	—	Preset 0.45,0.15,0.25,0.15.
--mem-weights	—	Preset 0.35,0.10,0.40,0.15.
--export [COLS]	—	Write CSV to stdout. Bare for the default columns, all for every column, or a comma-separated list. See CSV Export Reference.
--version	—	Print the version and exit.

Interactive commands

While the live dashboard runs, type a :-prefixed command and press Enter:

Command	Action
:focus <id>	focused single-GPU detail view
:plot	line-chart (history) view
:job	running-process table
:q	step back one view level (fleet → exit)
:exit	quit (also Ctrl-C)
Esc	cancel an unfinished : command

When the fleet has more GPU cards than fit the window, scroll the rows with ↑/↓, PgUp/PgDn, or j/k. The layout is responsive: it drops summary and footer fields as the terminal narrows, switches GPU cards between one- and two-column detail, and shows a “terminal too small” notice below a minimum size.

GPU visibility selection

KempnerPulse shows only the GPUs visible to your process, picking the first available source in this order:

1. --gpus

2. CUDA_VISIBLE_DEVICES

3. NVIDIA_VISIBLE_DEVICES

4. SLURM_STEP_GPUS

5. SLURM_JOB_GPUS

If none are set, all GPUs on the node are shown. --show-all ignores the environment entirely. Every selection is filtered against the GPUs actually accessible to the process (as reported by nvidia-smi), so cgroup and container restrictions are always respected.

Examples

kempnerpulse                                   # live, dcgm backend, 100 ms
kp --poll 1.0                                  # slower refresh
kp --focus-gpu 0                               # start focused on GPU 0
kp --hpc-weights                               # HPC weight preset
kp --gpus 2,3                                  # only GPUs 2 and 3
kp --once                                      # single snapshot
kp --backend dcgm --export all --poll 0.1      # high-res CSV capture
kp --backend prometheus --source http://host:9400/metrics --poll 2
kp --backend replay --source capture.txt --once