How to finetune Qwen3 in non-thinking mode?

[chingf]

I'd like to finetune a Qwen3 model in non-thinking mode. Is there a way to do this?
In another setting I'd usually do this by passing thinking=False as an argument into tokenizer.apply_chat_template. It seems like this was briefly supported after this pull request: #2694, but the change made by that pull request to src/axolotl/utils/schemas/config.py has since been deleted. Is there another suggested way to make sure the model is in non-thinking mode?

[ehartford]

you need to generate fake but plausible thinking blocks and mask them out so you don't override the models thinking skills but you still need the responses to respect the thinking blocks

[NanoCode012]

@chingf , re: the linked PR, it was accidentally removed in a refactor and I've made a PR to add it back here #2837

To answer the non-thinking mode, I haven't looked too much into it, so maybe Eric may know more about it.