Homelab with 3 mixed-GPU machines: how do you track local LLM models and VRAM?

pipecanon · reddit · 2026-11-06

A homelab user describes losing track of a local LLM setup spanning three machines with mixed GPUs plus a laptop, each running its own model runner—checking VRAM means SSHing into every box and finding stray quants loaded. They mention GPUStack for centralized cross-node deployment (while noting it can't pool physical VRAM across workers) and ask how others calculate RAM/VRAM needs for weights, KV cache, and loading, and whether per-box SSH is just the honest homelab answer.

Original post →

Homelab with 3 mixed-GPU machines: how do you track local LLM models and VRAM?

More from Infra