FastAPI DLS and current cards

[Hetz Ben Hamo]

Hi
I'm using an RTX 2080 with vgpu unlock and the FastAPI DLS service for license, since it's homelab
I'm thinking to buy a card such as A2000 or A4000 for vgpu. Does anyone knows if the FastAPI DLS license service will work with such cards?

 

[zachj]

My understanding is it’s not whether fastapi dls will work; it’s whether vgpu unlock will work. And I recall reading that it will not…

that said I didn’t bother to validate this information before posting it so take it with a liberal pinch of salt.

 

[ipreferpie]

Just to be certain of the current status of things, is this correct:
1) vgpu_unlock doesn’t work with ESXi
2) vgpu_unlock doesn’t work with Ampere+ cards
3) FastAPI-DLS is working up to v3.1 so basically all cards up to Ampere that has vgpu capability (irrespective of host thus ESXi = ok)
Not sure if I captured it correctly or am I missing anything?

 

[zachj]

Just to be certain of the current status of things, is this correct:
1) vgpu_unlock doesn’t work with ESXi
2) vgpu_unlock doesn’t work with Ampere+ cards
3) FastAPI-DLS is working up to v3.1 so basically all cards up to Ampere that has vgpu capability (irrespective of host thus ESXi = ok)
Not sure if I captured it correctly or am I missing anything?

I believe that is correct.

#1 - I do expect there could probably be an equivalent procedure if someone wanted to spend the time. Like for example using open core to spoof the pci device id of the gpu at boot time to fool the grid host driver into thinking you’ve got a compatible card.

DLS will only be broken by a windows/linux guest os driver update from nvidia that can discern a fake license.

 

[ipreferpie]

I believe that is correct.

#1 - I do expect there could probably be an equivalent procedure if someone wanted to spend the time. Like for example using open core to spoof the pci device id of the gpu at boot time to fool the grid host driver into thinking you’ve got a compatible card.

DLS will only be broken by a windows/linux guest os driver update from nvidia that can discern a fake license.

Thanks for the tip on #1…may give that a try if I have time. As for DLS broken from updates, basically to keep it working, is to not update to the latest drivers and wait until a later DLS version comes out to ensure continued compatability?

As for DLS, would you know if it works with later NVidia architecture such as Lovelace?

 

[zachj]

Thanks for the tip on #1…may give that a try if I have time. As for DLS broken from updates, basically to keep it working, is to not update to the latest drivers and wait until a later DLS version comes out to ensure continued compatability?

As for DLS, would you know if it works with later NVidia architecture such as Lovelace?

I don’t personally use DLS; I’ve got an actual vgpu license.

that said license enforcement is done by the windows/linux driver that’s installed inside of a virtual machine having an attached virtual gpu. The gpu itself isn’t doing license enforcement.

so as you said, the only reason to expect DLS to suddenly stop working is if you install an updated nvidia driver in your VM.

the reason DLS works is that from the perspective of the nvidia windows/linux guest driver it is receiving a valid license from the configured license server. For an nvidia driver update to break that would mean that nvidia is fundamentally changing what constitutes a valid license such that the license given to the vm by DLS is no longer considered valid.

 

[zachj]

I’m a bit confused as to why you’d need DLS if you’re following the vgpu-unlock guides; they all tell you to spoof a Quadro or rtx card so that you can install a consumer windows/linux guest os driver. If you do that then the driver doesn’t even look for a license in the first place.

 

[ipreferpie]

It looks like that Quadro spoofing isn’t that recommended way anymore, but instead to use FastAPI-DLS here: PolloLoco / NVIDIA vGPU Guide · GitLab

 

[juma]

I’m a bit confused as to why you’d need DLS if you’re following the vgpu-unlock guides; they all tell you to spoof a Quadro or rtx card so that you can install a consumer windows/linux guest os driver. If you do that then the driver doesn’t even look for a license in the first place.

CUDA doesn't work with spoofing.