Provider requirements

Everything a machine needs before it hosts Community sessions, and how to check it yourself. Plan about 15 minutes, plus one reboot.

The machine

Check it in one go

On the machine, in a terminal:

lsb_release -ds; uname -r; nvidia-smi --query-gpu=name,memory.total --format=csv,noheader; ls /dev/kvm; ls /sys/kernel/iommu_groups | wc -l; df -h /

A kernel below 6.12: install the HWE kernel, then restart.

sudo apt install linux-generic-hwe-24.04 && sudo reboot

nvidia-smi answers only while the NVIDIA driver is loaded. Once the GPU is reserved for sessions (see VFIO below), it no longer shows there: that is expected.

Firmware (BIOS or UEFI)

Open the firmware setup at boot (usually Del, F2 or F10) and set:

Names differ between boards. Look under Advanced, CPU Configuration, Chipset or North Bridge. Save, then boot Ubuntu.

IOMMU on the kernel command line

Intel boards need the IOMMU turned on in the kernel too. On AMD it is on by default; the same line does no harm.

1. Edit /etc/default/grub and add to GRUB_CMDLINE_LINUX_DEFAULT:
   • Intel: intel_iommu=on iommu=pt
   • AMD: amd_iommu=on iommu=pt
2. Run sudo update-grub, then reboot.
3. Check: cat /proc/cmdline shows the options, and ls /sys/kernel/iommu_groups lists groups.

Never add pcie_acs_override. It fakes the isolation between devices, and a machine that uses it fails review.

Each GPU in its own IOMMU group

A rented GPU is attached to one VM, with every device that shares its IOMMU group. List the groups:

for g in /sys/kernel/iommu_groups/*; do echo "Group ${g##*/}"; for d in "$g"/devices/*; do echo "  $(lspci -nns "${d##*/}")"; done; done

Each NVIDIA GPU should share its group only with its own functions (its audio device, sometimes a USB controller on the card) and PCI bridges. If a GPU shares a group with a disk controller or a network card, move it to another PCIe slot, preferably one wired to the CPU.

VFIO

Sessions reach the GPU through vfio-pci, the Linux driver that hands a device to a VM. The Community host install reserves the rented GPUs for you. To check it, or to set it up by hand:

1. Find the IDs of the GPU and its audio function: lspci -nn -d 10de: (for example [10de:2684] and [10de:22ba]).

2. Create /etc/modprobe.d/vfio.conf with:

   options vfio-pci ids=10de:2684,10de:22ba
   softdep nvidia pre: vfio-pci
   softdep nouveau pre: vfio-pci
   

3. Run sudo update-initramfs -u, then reboot.

4. Check: lspci -nnk -d 10de: shows Kernel driver in use: vfio-pci for each rented function.

Two cards of the same model share their IDs: both are reserved. Keep a different card, or the integrated graphics, for the console.

The self-test

The Node Lodge agent runs the same checks and tells you what to fix. On the machine, in the agent folder:

node compute-host.mjs doctor --community

It reads the machine and changes nothing. Each line is pass, warn or fail:

Fix every fail before you pair the machine. Next: Pair a machine.