NVIDIA GPU Passthrough Setup

I need my NVIDIA card passed through cleanly to an Ubuntu VM running on Proxmox VE 9 so the guest can handle machine-learning workloads without any hiccups. The host is already up and stable; the missing piece is a rock-solid passthrough configuration that leaves the GPU fully available inside the VM while keeping the Proxmox node happy after reboots and updates.

Here’s what I’m after:
• Configure the host (GRUB flags, IOMMU/VFIO modules, OVMF, etc.) so the GPU is properly isolated and assigned.
• Tweak or rebuild the Ubuntu VM until nvidia-smi recognises the card and reports its full compute capability.
• Install the matching proprietary driver plus CUDA toolkit so I can immediately start running my AI workloads.
• Supply a concise write-up or script covering every step, making it easy for me to replicate on other nodes.

I’ll provide console or SSH access and handle reboots on demand. The job is done when the VM boots consistently, `nvidia-smi` works, and the host remains stable.