GPU Passthrough for Proxmox Workloads

Dedicated NVIDIA hardware needed to be made usable inside virtualized Linux workloads, with failures possible across both hardware and software layers.

GPU passthrough is environment-specific. Exact server, firmware, PCIe and IOMMU topology, GPU, guest OS, drivers, and rollback requirements materially affect scope. This is not a claim of enterprise NVIDIA vGPU licensing or SR-IOV expertise, and it does not guarantee arbitrary GPU passthrough.

The work had to account for physical PCIe placement and server constraints as well as hypervisor and guest runtime dependencies.

• Worked through IOMMU and VFIO host configuration and PCI device assignment.\n- Validated GPU availability inside Linux guests.\n- Validated GPU-backed Docker and Ollama workloads.\n- Tested multiple NVIDIA devices and physical PCIe placements.\n- Investigated server riser, slot, and external-power requirements during testing.

Configured and tested the host-to-guest device path, then validated the resulting GPU runtime with in-scope workloads.

The environment was treated as a layered dependency chain rather than assuming every failure was an application problem.

Hardware-aware testing and scoped validation were used before considering a workload usable.

Functioning GPU-enabled virtualized workloads were achieved.

Verification covered GPU availability in Linux guests and GPU-backed Docker and Ollama workloads after host, device-assignment, and physical-placement investigation.

This demonstrates hardware-aware virtualization troubleshooting, IOMMU and VFIO, PCI passthrough, NVIDIA Linux drivers, CUDA and runtime troubleshooting, Docker GPU workloads, and local-AI infrastructure.