For conversation pipeline hf/transformers moved from tokenizer-specific prompt implementation to configurable templates (https://github.com/huggingface/transformers/pull/25323) and since the templates use Jinja, we can't reasonably use them (at least right now). So for now users need to use the text generation serving with the right prompt themselves. FWIW the serving only supported a few specific models (for which we had the templating logic), so I don't think it's been used much.
For conversation pipeline hf/transformers moved from tokenizer-specific prompt implementation to configurable templates (https://github.com/huggingface/transformers/pull/25323) and since the templates use Jinja, we can't reasonably use them (at least right now). So for now users need to use the text generation serving with the right prompt themselves. FWIW the serving only supported a few specific models (for which we had the templating logic), so I don't think it's been used much.