Berth v0.3.4 adds GPU VFIO passthrough: lend your 4090 to a Windows VM and get it back
Distinct-Pie2389 · reddit · 2026-10-22
Berth v0.3.4 ships experimental Linux GPU VFIO passthrough: berth vfio run safely drains your inference engines, hands the card (e.g. RTX 4090) to a Windows VM, and returns it to llama-swap once Windows shuts down — without touching drivers or passwords. Also new: gateway fallback to CPU engines during cold starts, minimum-turn-per-engine to stop model flip-flopping, systemd scopes for escaped child processes, and no more passing your node API key to engines. Windows fixes include per-process VRAM counts and proper process naming. Docs and libvirt configs on GitHub.
More from Infra
- Dual RTX PRO 6000 Benchmarks: DeepSeek-V4-Flash vs Qwen3.8-27B Local Serving — dee_hw · 2026-10-22
- Korean neocloud Elice now up 68% from IPO-day low — the panic headlines have gone quiet — firstadopter · 2026-10-22
- Synopsys adopts CUDA-X for EDA after Siemens cuLITHO, expanding Nvidia's GPU reach into chip design — pstAsiatech · 2026-10-22
- Nvidia moves to back GPU cloud provider Firmus after investors balk at its IPO — pstAsiatech · 2026-10-22
- Railway blocks deploys for 16+ hours, forcing theo to take three services offline — theo · 2026-10-22
- Dev urges compute marketplaces as SF Compute shows $4.29/gpu/hr resale prices — arthurcolle · 2026-10-22