The chat template takes quadratic time in the number of messages. A long tool-calling conversation spends seconds in template rendering on every request, before the model does any work.
Not something I ever expected to cost a significant amount of time
