Quickstart

Every command works as either kempnerpulse or the shorter kp.

Launch the live dashboard

kempnerpulse

This is equivalent to kempnerpulse --backend dcgm --poll 0.1 — it streams directly from dcgmi dmon at ~100 ms resolution and shows every GPU visible to your process (respecting CUDA_VISIBLE_DEVICES / SLURM_JOB_GPUS). Press Ctrl-C to quit.

One-shot snapshot

kp --once          # render a single frame and exit

Export CSV

kp --export                 # default columns, to stdout
kp --export all > run.csv   # every column
kp --export sm_active_pct,tensor_active_pct,real_util_pct   # a custom set

See CSV Export Reference for the full column reference.

Pick a weight preset

The Real Utilization composite is a weighted blend of SM / Tensor / DRAM / GR counters. Choose a preset for your workload, or supply custom weights:

kp --ai-weights      # AI / LLM training & inference (default): 0.35,0.35,0.20,0.10
kp --hpc-weights     # general HPC:                              0.45,0.15,0.25,0.15
kp --mem-weights     # memory-bound / bandwidth-heavy:           0.35,0.10,0.40,0.15
kp --weights 0.40,0.30,0.20,0.10   # custom (normalized to sum to 1)

Choose which GPUs to show

kp --focus-gpu 0     # start focused on GPU 0
kp --gpus 0,1        # only GPUs 0 and 1 (also accepts ranges like 0-3)
kp --show-all        # ignore SLURM/CUDA env and show every accessible GPU

Switch backends

kp --backend prometheus --source http://localhost:9400/metrics --poll 1.0
kp --backend replay --source capture.txt --once     # no GPU needed

Interactive commands

While the live dashboard is running, type :-prefixed commands:

Command	Action
:focus <id>	focus a single GPU’s detailed view
:plot	line-chart view (history sparklines)
:job	running-process table
:q	step back one view (focus/plot/job → fleet)
:exit	quit

In the fleet view, scroll card rows with ↑/↓, PgUp/PgDn, or j/k when there are more GPUs than fit the window.