Disable Headless mode on RTX a6000

[lloydeployd]

I have two RTX a6000 Turing GPUs that we previously used in vGPU headless mode. We were successful but let the project they were used for die out. After a couple years we want to now use them for local LLM not wanting them to be virtualized. The cards are still in headless mode and when I connect them to the Ubuntu machine it is recognized as a 03:00.0 3D controller: NVIDIA Corporation GA102GL [RTX A6000] (rev a1) and is not recognized as a usable GPU. Running nvidia-smi after disabling nouveau and installing the 580 driver it comes with a message about no compatible device found. I’m presuming it’s due to it’s previous use in headless mode. When I run the mode selector tool to set it enabled at 256MB bar1 it errors with this message:
terminate called after throwing an instance of ‘std::runtime_error’
what(): The PCI BAR assignment for the processed device is invalid.
Please check with NVIDIA web site for possible SBIOS Setup setting
to fit with the processed device.

I have set the kernel to “pci=realloc” and then to “pci=realloc=off” and both do not allow me to enble the GPU.

I’m running this on a Dell Precision Tower 7910, single Xeon E5-2643 v4, 64GB DDR4 RAM, and a 1200watt PSU.

What do I need to do to get this GPU running operational? I don’t care if it’s headless or with enabled display ports, the machine it’s in won’t be using a display.

[rs277]

Looking at the manual for the 7910, the BIOS entry, “Memory Map IO above 4GB”, is off by default. It would be worth trying to enable it, if not already.

Also, nitpick, the GA102GL A6000 is an Ampere, not Turing arch.

[lloydeployd]

Hot expletive! you are right on both… it is Ampere. My lazy self didn’t check specifics. I just remembered it’s not the Ada Lovelace edition.

I enabled that “Memory Map IO above 4G” and the driver now sticks and I get nvidia-smi to recognize and show the GPU resources. Thank you!