model : gpt-oss simulate think tags when reasoning

[aldehir]

Problem: the current implementation generates <|channel|>analysis<|message|>... in the output and clients send it back verbatim. This causes an exception in the official gpt-oss jinja chat templates.

This PR parses out the reasoning and does one of the following:

• reasoning_format = auto - send it in the reasoning_content field.
• reasoning_format = none - wrap it in <think></think>, sent to the content field to avoid exceptions.

Additionally, it parses out any final channels. It does not yet support tool use. More comprehensive parsing is being worked on #15181.

@ggerganov @ngxson

[ngxson]

Are we sure about removing this? It will prevent rendering these token without --special

[aldehir]

I am admittedly new to the code base, however for the web server it seems placing those tokens in preserved_tokens is sufficient to make them render.

I tested it with llama-cli and I see now it does omit them there. I will revert it.

[ngxson]

In addition to this, I think <|constrain|> and <|end|> should also be added in the condition

[ggerganov]

After 6d75412, the llama-cli now crashes after the first message.

[ngxson]

After 6d75412, the llama-cli now crashes after the first message.

I remembered the new jinja template has a check to prevent certain tags from being in the input text. Not sure what's the best way to fix, maybe we need a patch/hotfix for that.

[aldehir]

Yes, I am seeing that as well. It's the exception thrown from the jinja template:

You have passed a message containing <|channel|> tags in the content field. Instead of doing this, you should pass analysis messages...

I can revert the last commit as a temporary workaround. It seems to work with --special in f058384 as well.

[aldehir]

At least for now, I'm going to keep it as I originally had it since it is somewhat usable. I haven't explored the CLI enough to have any good input.

[ngxson]

IIRC chat.cpp also allow patching the jinja template, see an example in common_chat_params_init_deepseek_r1

We can temporary patch the jinja version of harmony, so that it doesn't throw an error. Then later on we can spend more time to do a proper fix

[ngxson]

Also for visiblity, I think probably we don't need to replace the reasoning tags to <think></think> as introduced in this PR. With the migration of the new frontend to Svelte, we will eventually support reasoning_content field which is be much cleaner. We plan to release this version this weekend, so let's hold off this PR a bit.

In the meantime, having tool call support (on your other PR) is a very good feature.

[ggerganov]

@ngxson I would like to have a the WebUI chat experience fixed quickly to the state that was working before the jinja template updates. Do you have a suggestion for a fix? I can also revery the GGUF models back to the old template?

[ngxson]

@ggerganov yes I can try to patch out the exception in jinja template