Thunderbolt eGPU passthrough possible?

[ruthbd]

Good evening,

I have a 2018 Mac Mini running proxmox 8.3.2. I've recently attached a thunderbolt eGPU with an NVIDIA 3070 GPU. I've followed countless guides, forum posts, etc. to try to get this working in a Debian 12 VM, to no avail. I can get the passthrough "working" as far as the GPU device showing up in the VM, but no matter what I've tried, I get the following after I install the NVIDIA cuda drivers (from the nvidia site).

[    6.500485] NVRM: This PCI I/O region assigned to your NVIDIA device is invalid:
               NVRM: BAR1 is 0M @ 0x0 (PCI:0000:00:10.0)
[    6.515742] NVRM: This PCI I/O region assigned to your NVIDIA device is invalid:
               NVRM: BAR2 is 0M @ 0x0 (PCI:0000:00:10.0)
[    6.515744] NVRM: This PCI I/O region assigned to your NVIDIA device is invalid:
               NVRM: BAR3 is 0M @ 0x0 (PCI:0000:00:10.0)
[    6.515745] NVRM: This PCI I/O region assigned to your NVIDIA device is invalid:
               NVRM: BAR4 is 0M @ 0x0 (PCI:0000:00:10.0)

(pci=realloc doesn't do anything)

So, at this point, after many hours, I have to ask, is this combination even possible? Intel Mac Mini, Thunderbolt, eGPU, NVIDIA 3070, proxmox 8.x, Debian VM?

If not, does any virtualization software work with this hardware combination? Goal: VM that can be used to run LLMs for experimentation, leveraging the hardware GPU connected via eGPU.

Thanks in advance if anyone can assist.

 

[gfngfn256]

I don't do any passthrough per se, but skimming your post it appears you are passing through the eGPU. Have you considered passing through the TB port itself? I'm guessing it will only work if that port is not linked to others & the obvious IOMMU grouping is separate. Also that port/controller must support passthrough.

I can't really help you much more, so good luck.