AMD published a guide to running passed-through Radeon GPUs in Proxmox VE VMs with ROCm and PyTorch for isolated, API-managed workloads.

AMD has published a technical guide showing how Proxmox VE can be used to run Radeon GPUs as managed virtual infrastructure rather than standalone workstations, including ROCm and PyTorch inside passed-through virtual machines. The article argues that the approach is suited to isolated development, CI runners and air-gapped environments, while noting that VFIO passthrough assigns a GPU exclusively to a VM.

In the guide, published by AMD on 19 October 2026, the company says a hypervisor layer turns AMD hardware into a private GPU cloud that can be managed through templates, snapshots, cloning and REST APIs. AMD says the same passthrough method applies to both integrated Radeon graphics in Ryzen AI Max systems and discrete Radeon AI PRO Graphics, with the guest loading the AMD ROCm stack against the physical device.

AMD’s tested configurations include Proxmox VE 9.1 on an AMD Ryzen AI Max+ 395 system with a Radeon 8060S integrated GPU and on an AMD Ryzen Threadripper PRO 9965WX host with four Radeon AI PRO R9700 discrete GPUs. The guest operating system in both cases was Ubuntu 24.04 HWE, with ROCm 7.2 and PyTorch 2.9.1+rocm7.2.

The article also highlights platform-specific memory ceilings. AMD says Ryzen AI Max+ 395 systems support up to 128 GB of unified LPDDR5X memory, which it says can enable models up to 128 billion parameters at 4-bit quantization. It adds that the Ryzen AI Max+ PRO 495 platform raises the ceiling to 192 GB and can support models exceeding 300 billion parameters at 4-bit quantization.

Source link