GPU monitoring

A GPU poller auto-detects NVIDIA (nvidia-smi), AMD (amdgpu_top first, then rocm-smi), or Apple Silicon (system_profiler), polls every 10 seconds, caches snapshots, and publishes GPU-change events for live dashboard refresh.

AMD APUs report GTT memory (not BIOS-limited VRAM) for actual usable memory; Apple shows unified memory and core count when available. In Docker, /dev/dri must be mounted into llama-dash for AMD stats — the AMD compose file does this.

Detection order

At startup, llama-dash checks telemetry sources in this order:

1. nvidia-smi
2. amdgpu_top -J -n 1
3. rocm-smi
4. system_profiler SPDisplaysDataType -json on macOS

The first working source becomes the poller driver until restart.

Each command has a 5-second timeout. If detection finds no supported source, GPU telemetry is marked unavailable and no polling interval is started.

Polling behavior

When a driver is detected, llama-dash polls every 10 seconds. Each poll updates the cached GPU snapshot and publishes gpu.updated only when the snapshot changes.

If a later poll fails, the snapshot is marked unavailable with the detected driver preserved. This helps distinguish "no GPU source was found" from "the detected source stopped responding".

Reported fields

NVIDIA can report:

• Memory used and total.
• Memory percentage.
• GPU utilization.
• Temperature.
• Power draw and power limit.

AMD via amdgpu_top can report:

• VRAM or GTT memory depending on device type.
• Utilization from GFX and compute fdinfo.
• Temperature.
• Power.
• Compute unit count when available.

AMD via rocm-smi can report memory, utilization, and temperature from CSV output.

Apple Silicon reports GPU model and core count from system_profiler, plus unified-memory usage from sysctl and vm_stat.

Containers

For AMD telemetry inside Docker, mount /dev/dri into the llama-dash container. Without the device mount, the dashboard may still work but GPU polling will remain unavailable.