Flags for inference enrich the output with explainability information or suppress specific input token/embedding spaces, as described here.
Motivation
I'm not sure if this is out of scope of TGI, but I figured I might bring it up at least.
Explainability and truthfulness is a very important topic for us and our users, and after watchin GTC and reading this paper (github), I believe there's potential here to significant work here that could be done.
The explainability and suppression methods proposed in that paper would be extremely useful for us, though I can understand if that is too niche.
Your contribution
I'd be willing to put in groundwork if that helps, though I also have to admit that I'm out of my depth at that level of manipulating attention, so I would definitely need some guidance if I were to contribute.
Feature request
Flags for inference enrich the output with explainability information or suppress specific input token/embedding spaces, as described here.
Motivation
I'm not sure if this is out of scope of TGI, but I figured I might bring it up at least.
Explainability and truthfulness is a very important topic for us and our users, and after watchin GTC and reading this paper (github), I believe there's potential here to significant work here that could be done.
The explainability and suppression methods proposed in that paper would be extremely useful for us, though I can understand if that is too niche.
Your contribution
I'd be willing to put in groundwork if that helps, though I also have to admit that I'm out of my depth at that level of manipulating attention, so I would definitely need some guidance if I were to contribute.