Hello, I am building out an AI server for AVQA (Audio-Visual Question Answering), inference and RAG for continued learning. After doing the usual pass through steps, the VM would not start at all. In fact, when I passed a GPU through in an effort to assist, they disappeared completely and it was back to the drawing board.
Reliable PCIe Passthrough for Hailo-8/8L — use bus reset (not just FLR)
TL;DR
If your Hailo card doesn’t cleanly re-initialize after VM reboots/shutdowns, stop relying solely on FLR (Function-Level Reset). Add a bus-level reset of the upstream bridge before handing the device back to VFIO/QEMU. This has been the most reliable fix in our setup.
The hardware stack:
- Tyan FT77C server
- 2x E5-2697a V4 cpu
- 256gb ddr4 ecc ram
- 3x Nvidia Tesla p100,
- 1x RTX 3060 12gb
- 1x Hailo 8
- 1x Hailo 8L
Software Stack:
- Proxmox
- Ubuntu VM
- Ollama LXC
- Open Web UI LXC
- Key Cloak VM
This is the link to the Github post
I am setting up a