any plans in supporting custom sampler like huggingface? Could be nice to have for e.g. grammar constrained decoding or exotique sampling methods. Stuff like Classifier Free Guidance (also supported by hf) or Logit Diff Amplification would also be cool in the future, but I guess they are tricker because they depend on calls to different models. Like you would need 2 sampling process depending on each other (and one of them force the output to the other)
any plans in supporting custom sampler like huggingface? Could be nice to have for e.g. grammar constrained decoding or exotique sampling methods. Stuff like Classifier Free Guidance (also supported by hf) or Logit Diff Amplification would also be cool in the future, but I guess they are tricker because they depend on calls to different models. Like you would need 2 sampling process depending on each other (and one of them force the output to the other)