Skip to main content
Warpscale bundles Grafana with your organization. Sign in to Warpscale, then open https://<your-organization-host>/grafana.

​
Dashboards

The bundled dashboards are in the Warpscale folder.

​
Warpscale · vLLM

Serving behaviour for one vLLM instance, picked from the dropdown. Two rows.
  • vLLM Instance — peak KV cache use, peak running and waiting requests, and the hourly medians for queue wait, prefill throughput, and time per output token
  • vLLM Tenant — the same instance broken down by tenant: prefill and decode tokens, token share, KV residency and occupancy, queue wait, decode excess, contention, contention per request, finished requests
The medians on the first row are the instance’s own typical pace, and the tenant panels are scored against them. Decode excess is decode time past that pace. Contention adds queue wait on top, so a tenant high on contention is the one being slowed down — not the one doing the slowing. Contention per request divides by the tenant’s finished requests, so it does not climb just because a tenant sent more traffic. The tenant row needs the tenant header wired up. See Tenant attribution — without it every request lands under _untagged.

​
Warpscale · Training Features

Named per-GPU features for a training run, grouped the way a run fails: training loop, GPU compute, GPU memory, CUDA sync, data transfer and communication, checkpoint I/O, disk I/O, and hardware health. Use this to see what changed — backward rate, memory pressure, projected time to OOM, NCCL blocked time, thermal headroom.

​
Warpscale · Raw Metrics

The signals underneath the features, before any distillation. NVML hardware counters per host — temperature, power, clocks, PCIe, ECC, throttle flags — and eBPF counters per run for CUDA, NCCL, and Torch. Use this when you want to check a feature against what it was derived from.

​
Warpscale · Metric Explorer

One metric at a time, chosen from a dropdown. Use it for anything the other three do not plot.