Developer's local AI ran on CPU for a week due to silent GPU passthrough failures

A developer running a self-hosted Ollama AI service on Proxmox discovered their setup had been silently falling back to CPU inference for nearly a week, causing roughly ten-times slower performance. The first cause was VRAM overflow: model size and context length share the same GPU memory budget, and when the num_ctx parameter was raised for longer conversations, the combined load exceeded the GPU's capacity, prompting Ollama to quietly offload processing to the CPU without any error or warning. The second cause was a GPU passthrough conflict — a previously shut-down VM still had the graphics card pinned in its config, and on reboot it claimed the card before the Ollama container could, leaving the AI service with no GPU at all. In both cases, Ollama continued running normally from the user's perspective, making the root cause difficult to identify without explicitly checking GPU utilization. The key lessons are to verify actual VRAM usage after changing models or context settings, and to audit GPU passthrough assignments after every host reboot.