apply_chat_template method pessimizes perf at the interface layer, forcing 2 copies #641
Description
The function is defined as taking an Option<String> for tmpl but it only ever references the string as a reference since it immediately creates a CString copy internally for the purposes of calling the apply_chat_template function. Option<&str> would be better avoid forcing a copy by the caller just to give it an owned str.
Option<CStr> would actually be best - that way the caller could be responsible for allocating the CString once & then the chat template could be generated repeatedly using 1 in-memory copy of the template across all completions.
Given that the templates can be quite large, this isn't nothing although within the context of chat completion it's probably an overall minor speedup / memory reduction.