AI BriefingBerth v0.3.4 adds experimental GPU passthrough to Windows VMs via VFIO
·2026.10.22 10:02
Key point
The update introduces a command to drain engines and pass the GPU to a VM, alongside gateway fallbacks and security fixes.
Details
- Berth v0.3.4 introduces experimental GPU passthrough to Windows VMs on Linux using VFIO, allowing users to hand over the GPU via a single command while the tool manages engine draining and driver switching.
- The gateway now supports fallback to CPU engines during slow cold starts to prevent request timeouts, and a new minimum turn setting stops the card from constantly flipping between models.
- Security improvements include preventing the gateway from passing node API keys to engines and ensuring escaped child processes remain on their engine lease via systemd scopes.
- Windows-specific fixes correct per-process VRAM counting to only include Berth's card, assign proper names to processes like dwm.exe, and stop Epic's tray app from grabbing the GPU.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
llm-d 1.0 releases production-ready LLM inference stack with multi-tenant QoS and stable API