Speaker unverified: belief

Transcript context

…Is there a way for the community to donate compute to you? Is there an interface that you can easily transfer your jobs to a different cluster? It's actually been discussed a lot, and we are thinking about adding the option to run evaluation on inference endpoint, where people would be able to pay for the compute of their evaluation. The thing is, at the moment, we really wanted to use a EleutherAI harness because it's a big stable library that everybody uses, and we think that Elusive is doing a great job at evaluations in general. But we have the functionality to run evaluations on inference endpoint in our own evaluation library, which is called Lightval. So we will have to port this functionality to the harness before being able to give it to the users. It's not been high on our priority list because then we will have to set up possibly another space where evaluation will run, or maybe people will have to duplicate some stuff. It's more engineering, and we've been a bit swamped with things to do. I can imagine. Yeah, so hopefully when that opens up for inference endpoints, the only thing I'll caution is that all the inference providers write their own CUDA kernels and implementations of stuff. So sometimes you won't get one-for-one the same model, even though it's the same weights, but it's not exactly the same performance of the model because they quantize or do whatever they want to do with the shortcuts for attention.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.