cui-host GPU Monitor

Window

Compute SM

Memory Ctrl

GPU Temp

Fan Speed

PCIe Activity

PCIe RX

PCIe TX

System CPU

RAM Used

Swap Used

RAM BW

Load Avg

Best Lagged Correlations

Positive lag means the metric led SM compute by that many seconds. This is association, not proof of causation.

Low vs High Compute Windows

Compares samples below 40% SM with samples above 70% SM in the selected time window.

Current Read

Waiting for samples...

Bottleneck Hints

HRM Training Logs

Waiting for progress...

Reads the live JSONL progress log and summarizes loss, token accuracy, exact accuracy, throughput, and ETA.

No log path yet.

Unknown

Process check pending.

Updated -

Step

Loss

Token Accuracy

Exact Accuracy

Seen

Rate

ETA

Runs

Recent Step Events

Newest entries from the progress log. The chart window follows the latest API sample limit.

Data source: automatic GPU backend, preferring amd-smi for ROCm GPUs and falling back to nvidia-smi only when AMD telemetry is unavailable; training statistics read /api/hrm-progress from the HRM JSONL progress log. RAM bandwidth is sampled from pve3 with AMD data-fabric perf counters when available. PCIe values are MB/s. Samples every ~0.67 seconds; refreshes every 2 seconds.