Chat template is busted
#28
by FrenzyBiscuit - opened
I'm on the latest version of the quants.
When using the 26B A4B Q8 and using tool calls, it... breaks.
None of that should be showing up as it is. Those should just be showing tool calls only (without ANY text of what its doing).
the 31B is better at this, but also falls on its face.
It outputs the results in the thinking portion and doesn't actually output the results correctly!
This is with the latest llamacpp on chat completion.
In comparision, Qwen 3.5 does not have this issue -.-.
