VM mode on a multi-GPU server: from BIOS to the first client

Virtual-machine mode opens a client segment that container rentals never see: the people who need their own kernel, their own driver, nested virtualization or Windows. For a machine without VM support those clients simply do not exist — it never appears in their listings. Not «more expensive», not «worse»: invisible.

Setting up a multi-GPU server for VMs is harder than a single-card one, and most of the failures along the way report something other than what actually happened. The enabling script talks about IOMMU groups when the cause is a driver parameter. The apt error points at package pins when the cause is the marketplace's own preferences file. The red bus-bandwidth warning appears because a test image was evicted from the cache. And the VM support flag is wiped not by a marketplace update but by the marketplace's own hourly job — on a four-card build that coincidence is so rare that weeks pass between episodes.

Below is the whole procedure in working order: what to check before going into the BIOS, where the isolation setting hides on server boards, how to harden the cards against the power-saving trap, how to confirm that hardening with a controlled ladder of 1 → 2 → 3 → 4 cards in virtual machines before the first client ever sees the machine, and how to close the dropping flag with a watchdog that disturbs neither clients nor the reliability score.