v1.5.0

What's new in v1.5.0

The Running Models sensor now tells you which backend is serving each loaded model and how much VRAM it's using.

Added

• Backend per running model — each entry in the Running Models sensor now includes a backend attribute (e.g. llama-cpp, vllm, whisper), so you can see what's actually serving each loaded model.
• VRAM usage per running model — each entry gains a VRAM estimate (vram_estimate_gb / vram_estimate_display) plus context_length. Where a backend supports the live status RPC, real state and memory_gb are included too (best effort).
• Roll-up attributes — backends_in_use (the distinct backends in play) and total_estimated_vram_gb (sum across all running models), handy for at-a-glance dashboards.

To use them: add the backend / vram_estimate_display fields from sensor.localai_monitor_running_models to a table or template card.

Notes

• The per-model VRAM figure is an estimate derived from the model config — LocalAI does not expose measured per-model VRAM for most backends (e.g. llama-cpp). Measured, server-wide GPU/VRAM usage remains on the Resources sensor.
• No configuration changes required; no new minimum LocalAI version.