Hi zowoq and Jonas Chevalier! I was wondering what has been your impression of the cuda jobset so far: whether you find it sustainable to build as is, whether it'd be sustainable to scale up? With the experience you've had so far, do you think nix-community should commit to keeping it alive and what'd it take to call it "stable" and announce it to the public?

About nix-community/infra in general, I understand you zowoq are "the official maintainer" and afaiu the infra is funded by the foundation. I wonder what is the general idea for nix-community's sustainability? Does infra require more labour that currently available? Who's taking over if you need to move to other things? Are there plans to get off the "cloud needle"?

Thanks

I was wondering what has been your impression of the cuda jobset so far: whether you find it sustainable to build as is, whether it'd be sustainable to scale up?

Yes, it is sustainable. I don't understand what scale up would mean in this context?

With the experience you've had so far, do you think nix-community should commit to keeping it alive and what'd it take to call it "stable" and announce it to the public?

Yes, I think we can commit to it. It is stable enough that we added it to our docs a couple of days ago. https://nix-community.org/package-sets/

afaiu the infra is funded by the foundation.

No, we don't get any funding from the foundation. We have an open collective and some services are sponsored. https://opencollective.com/nix-community, https://nix-community.org/sponsors/

I wonder what is the general idea for nix-community's sustainability?

I'd say that basically it just depends on the open collective. @zimbatm may have a more nuanced answer for this.

Does infra require more labour that currently available?

No, not at the moment.

Who's taking over if you need to move to other things?

https://nix-community.org/administrators/ Nix community has five admins.

Are there plans to get off the "cloud needle"?

I'm assuming that this means getting our own hardware and finding somewhere to put it? Has been mentioned once or twice but nothing beyond that.

I don't understand what scale up would mean in this context?

Building more "configs": e.g. the current aarch64 jobset only builds the variant for normal plug-in pcie GPUs not for jetson boards but we could extend it (which is probably the most common aarch64 user in fact...). The jobset also only chooses the "fat" variants which include code for all available gpu architectures at once. We could spawn variants for individual architectures, which is particularly relevant for embedded systems. I'd say we don't need to enable any of these at the moment, but if/when we discover this is needed it'd mean multiplying the cost. Similarly, adding the stable branch would add to another multiplicative factor, innocuous on its own but possibly significant when coupled with other toggles

the current aarch64 jobset only builds the variant for normal plug-in pcie GPUs not for jetson boards but we could extend it (which is probably the most common aarch64 user in fact...).

If jetson is the more common use case should we switch to building that instead of the current aarch64 jobset?

The jobset also only chooses the "fat" variants which include code for all available gpu architectures at once. We could spawn variants for individual architectures, which is particularly relevant for embedded systems.

If we built the individual architectures would we still need to build the fat variant? Is building an individual architecture quicker than the "fat" variant?

the current aarch64 jobset only builds the variant for normal plug-in pcie GPUs not for jetson boards but we could extend it (which is probably the most common aarch64 user in fact...).

If jetson is the more common use case should we switch to building that instead of the current aarch64 jobset?

The jobset also only chooses the "fat" variants which include code for all available gpu architectures at once. We could spawn variants for individual architectures, which is particularly relevant for embedded systems.

If we built the individual architectures would we still need to build the "fat" variant? Is building an individual architecture quicker than the "fat" variant?