Is there a way to turn off Qwen3 32B thinking mode and add a thinking budget?

[tjoymeed]

HI Team,

Thanks a lot for your excellent work!

I have two separate questions:

1. Is there a way in VeRL to turn off Qwen3 32B thinking mode?
2. Is there a way in VeRl to pass in the argument of thinking budget?

Thanks!

[mihirp1998]

yeah this will be very useful to have. Having thinking budget for Qwen3 during RL tarining

[DenceChen]

tokenizer_config.json -> chat_template -> {%- if enable_thinking is not defined or enable_thinking is false %}

[Fdioa]

tokenizer_config.json -> chat_template -> {%- if enable_thinking is not defined or enable_thinking is false %}

Will this method work when using vllm_sync for rollout? I couldn't find any implementation of vllm_sync_rollout that uses tokenizer.apply_chat_template.

[Chenruishuo]

HI Team,

Thanks a lot for your excellent work!

I have two separate questions:

1. Is there a way in VeRL to turn off Qwen3 32B thinking mode?
2. Is there a way in VeRl to pass in the argument of thinking budget?

Thanks!

for question 1, set +data.apply_chat_template_kwargs='{enable_thinking:false}' \