Crowdsourcing hardware support

[castrojo]

aka: ramala submit-hardware or whatever.

We've hit a few cases with ublue users where the hardware acceleration wasn't supported. I've had this experience myself until one day it magically worked after an upgrade. I've had a discussion about this with @ericcurtin but it was a while ago.

Anyway, I think I have enough users in ublue to make it feasible to crowdsource this so I thought I'd start a thread. What would you need on the shipping custom image, etc.?

[rhatdan]

@ericcurtin thoughts?

Are you talking about opening issues about hardware that does not work, but should? Usually this is some GPU that we did not know about and could easily be added to a list.

[ericcurtin]

@castrojo like you mean...

ramalama submit-hardware

and it sends some request somewhere so we can track the hardware (especially the broken hardware)? I think this could be a good idea. I wonder could we get it to automatically open a github issue with a certain label to avoid setting up new infrastructure. Open to ideas...

Covering all hardware support is difficult, all AI projects suffer from this. We should improve.

We depend on the community a little, to test, submit fixes, etc. on a given piece of hardware.

[castrojo]

I guess I should back it up a bit and go from a higher level, sorry this isn't my area of expertise.

What needs to happen for hardware to be enabled? The reason I was thinking of a command is for the user to submit the information for that hardware to you so you could enable it.

I was thinking it's:

• New user reports the hw doesn't work
• User gets sent here and files an issue with a specific issue template so it's labelled properly, etc.
• User runs command to gather information and then pastes it into the ticket

Or is it a matter of updating a list of model numbers?

[ericcurtin]

It's a big pile of "it depends", sometimes there's buildtime or runtime flags we can change, sometimes there's enablement/bugs to be fixed in llama.cpp to fix such as the llvmpipe fix recently.

The info can be somewhat helpful.

[ericcurtin]

What's often better since everyone uses llama.cpp is to report the build and runtime flags and open an issue against llama.cpp