I think you mean "run as part of the sampler," the tokenizer (and tokenization) is fixed for a given model. The sampler blob would basically: 1. Modify the out

I think you mean "run as part of the sampler," the tokenizer (and tokenization) is fixed for a given model. The sampler blob would basically: 1. Modify the output token probabilities to fit any arbitrary use case 2. Perhaps do trigger some sort of backtracking / beam-search (I'm not Grant but we've chatted on twitter and built similar things)