Mixed Blackwell (RTX Pro 6000 WE) + RTX A6000 vGPU on Proxmox: Open kernel module forced, GA102 GSP/RM init fails (RmInitAdapter 0x62:0x56:1993)

Graphics / Linux Linux
1

Hi,

I have installed my RTX Pro 6000 into my server which I run 4 VMs from. I have (or had) an A6000 in vGPU mode split across the 4 machines. My aim was keep an RTX Pro 6000 Blackwell Workstation Edition in the host (for passthrough to a VM via vfio-pci), and run vGPU on an RTX A6000 (GA102GL) on the same Proxmox node (Linux KVM).

I’m not trying to vGPU the Blackwell card — only the A6000 should be the pGPU for vGPU.

Proxmox VE 8.x (Debian 12 “bookworm”)
Kernel: 6.8.12-13-pve

  1. vGPU Manager: R580 (package: NVIDIA-GRID-Linux-KVM-580.65.05-580.65.06-580.88.zip)
    Installed NVIDIA-Linux-x86_64-580.65.05-vgpu-kvm.run
  2. Boot mode: CSM (legacy) (not UEFI)
  3. Platform: AMD (IOMMU enabled)
    Kernel cmdline: iommu=pt amd_iommu=on
  4. GPUs in the same host:

    • RTX Pro 6000 Blackwell WE (PCIe 01:00.x)

    • RTX A6000 (GA102GL) (PCIe e1:00.0)

  5. nvidia-smi reports version 580.65.05 when the driver is up.

What happens

  1. The vGPU Manager installer shows this warning when both GPUs are present:

This system requires the open GPU kernel modules, but contains GPUs which are not supported by the open GPU kernel modules. The unsupported GPUs will be ignored.

  1. After install/reboot, the Open GPU kernel module is loaded (as expected with Blackwell present). dmesg confirms:
    confirms:

NVRM: loading NVIDIA UNIX Open Kernel Module ... 580.65.05

  1. With the A6000 bound to nvidia, RM/GSP firmware init fails for GA102 under the open module. Repeating pattern in dmesg:

NVRM: nvAssertFailedNoLog: Assertion failed: pBinArchive != NULL @ kernel_gsp_booter.c:487 NVRM: nvCheckOkFailedNoLog: ... NV_ERR_NOT_SUPPORTED (0x00000056) ... kgspAllocateScrubberUcodeImage(...) NVRM: RmInitAdapter: Cannot initialize GSP firmware RM NVRM: GPU 0000:e1:00.0: RmInitAdapter failed! (0x62:0x56:1993) NVRM: GPU 0000:e1:00.0: rm_init_adapter failed, device minor number 0

Firmware files exist on disk:

/lib/firmware/nvidia/580.65.05/gsp_ga10x.bin /lib/firmware/nvidia/580.65.05/gsp_tu10x.bin

  1. vGPU Manager can’t attach to the A6000 and exits/fails:

nvidia-vgpud: error: failed to attach device: 59 nvidia-vgpud: error: failed to read pGPU information: 9 nvidia-vgpu-mgr: error: vmiop_env_log: Failed to attach device: 0x59 (gpuId 0xe100)

  1. nvidia-smi vgpu -q shows no vGPU types or instances with this open‑module path:

No devices were found # (or just "Active vGPUs: 0" and no Supported vGPU section)

So far I have tried:

Bound the Blackwell functions (01:00.0/01:00.1) to vfio‑pci (via driver_override and via early boot vfio-pci.ids=10de:2bb1,10de:22e8) so that NVIDIA’s driver owns only the A6000.
Even then, the installer still forces the Open kernel modules due to Blackwell being present in the system, and the A6000 continues to fail RM/GSP init under the open path.

Cleaned out any preloads/drop‑ins, restarted services, etc. The failure pattern remains the same.

Verified IOMMU is on and device nodes exist; verified the GSP firmware files are present.

My working assumption is the following

The presence of Blackwell (WE) on the host causes the vGPU installer to require the Open kernel module path.

On R580, the Open kernel module path does not initialise GA102GL (RTX A6000) in a way that vGPU Manager can use (GSP/RM init fails, vGPU attach fails).

Therefore, running RTX Pro 6000 Blackwell (host) and RTX A6000 vGPU on the same node is effectively unsupported today, because Blackwell forces Open RM and GA102 vGPU wants the proprietary module path.

Which brings me to the following questions…

Is A6000 (GA102GL) vGPU supported on R580 when the Open GPU kernel module is in use on the host? (If yes, what’s missing on my system to fix the GSP/RM init failures above?)

On a mixed‑GPU host that includes a Blackwell Workstation Edition, is there a supported way to override the installer’s “Open module required” decision and use the proprietary kernel module just for the A6000 vGPU, while keeping the Blackwell bound to vfio‑pci (no NVIDIA driver attached)?

Does CSM vs UEFI boot mode matter for Open RM / GSP on GA102 here? I’m currently in CSM; should I switch to UEFI to avoid the pBinArchive / RmInitAdapter failures?

I have been trying to make this work for two days going round and round in circles. It was working fine till I installed the new GPU.

Thank you,

N

2

Can anyone from NVIDIA confirm whether from what have read so far, which is, if I have vGPU A6000 running in my workstation and I install the RTX pro 6000 workstation edition with the view to pass through the Pro 6000 to my a VM that it just effectively blows up my vGPUs? So if you are buying this for your workstation and you have a couple of VMs running on it, this will stop you from using the VMs and the licences you have bought from them?!

Surely you should be able to let this run on the vGPU drivers but only as a passthrough. If you want to use MIG instances then you would have to use the open drivers?! You say nothing about this and neither did your sales team.

3

Bump.

I’m in the same boat here, there needs to be a flag to force proprietary drivers only. I have an A5000 with the same issue, there is a black listed older 1660 on the same machine. As a result I can’t install the 580 or 570 driver!

4

In case it helps someone else. This hard to find flag is what you need!

Force proprietary modules on r57x/r58x

NVIDIA added an explicit switch in recent drivers:

--kernel-module-type=proprietary    # (or =open)

Add this to the run command!

Solved