Ollama 780M passthrough stalls on CPU

This reads more like a virtualization and driver-stack problem than an Ollama core issue. Ollama does support the Radeon 780M on Linux, but only when ROCm sees the device correctly, and VM passthrough can still leave you stuck on CPU fallback.

• –Ollama’s docs list the Radeon 780M as a supported `gfx1103` AMD GPU on Linux via ROCm v7.
• –On an APU/iGPU, “VRAM” is shared system memory, so low reported GPU memory alone does not prove offload failure.
• –In Proxmox, the likely failure mode is incomplete passthrough or missing ROCm-visible device access inside the guest.
• –The post is useful as a real-world edge case for self-hosted local inference on AMD hardware, especially for homelab users.
• –There is no broader product announcement here; this is a troubleshooting thread around deployment friction.