Fix code typo in llama-cli

[ngxson]

Fix a small typo that breaks chat template support on llama-cli -cnv

---

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low

[slaren]

I don't think it is completely right yet, there are still extra new lines added to the assistant messages randomly. I suspect that at least one issue is that the chat template expects a new line after the <end_of_turn> of the assistant, but it is not being added.

[ngxson]

Hmm, that's strange. My result is pretty consistent (with -t 0)

> say 1
1

> say 2
2

> say 3
3

> say 4
4

> say 5
5

> say 6
6

> 

[slaren]

That case is also fixed for me, but I still see many messages ending in double or triple lines during random chat.

[ngxson]

Probably there is some thing more specific for gemma template (or the model itself).

In any case, I'll merge this PR now and have a deeper look on gemma later.

[ngxson]

I don't think it is completely right yet, there are still extra new lines added to the assistant messages randomly. I suspect that at least one issue is that the chat template expects a new line after the <end_of_turn> of the assistant, but it is not being added.

I see what you mean. In this case, we must either patch the template behavior with add_ass, or either patch llama_chat_format_single to be aware of the trailing new line. I'll see which way is better.