Change _fix_chat_template in case a template has both endif and endfor

[giuliabaldini]

Fixes #1284.

When running the following code with unsloth at the current commit (4cbebe151d9c8f813e4e69be1d86a5657a44ee60) I

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="OpenMeditron/Meditron3-8B",
    max_seq_length=4096,
    load_in_4bit=True,
)

I get the following error:

RuntimeError: Unsloth: The tokenizer `OpenMeditron/Meditron3-8B`
does not have a {% if add_generation_prompt %} for generation purposes.
Please file a bug report immediately - thanks!

This used to work definitely until the Qwen commit.

The problem is that the template has both an endif and an endfor, so the current code finds the wrong spot in the template with .index(). This is the old_template:

{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>
'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{{ '<|start_header_id|>assistant<|end_header_id|>
' }}

The PR should fix this by always taking what comes last, either endif or endfor.

[danielhanchen]

@giuliabaldini Thanks for the PR and apologies on the delay!

@Erland366 Could you review this thanks :)

[Erland366]

Sorry for the delay

I tested the PR on all Unsloth + meta-llama + Qwen + Mistral (assuming non-Unsloth model by default is not fixed) 7B/8B/9B instruct model and there's no issue. Logic seems good too

[danielhanchen]

Thanks @giuliabaldini - apologies for the delay again!!