Reasoning without using the think function

[imjking]

Hi, i want to use Qwen3_0.6B model in 8255 device, i exported pte model and run it on device successfully. Now i want to disable the "think" function to verify something, how can i achieve it ?
I use the following command and get outputs.txt:
./qnn_llama_runner_ndk27 --decoder_model_version qwen3 --tokenizer_path tokenizer.json --model_path hybrid_llama_qnn.pte --prompt "who are you" --seq_len 512 --eval_mode 1 --temperature 0.8 && cat outputs.txt

cc @cccclai @winskuo-quic @shewu-quic @haowhsu-quic @DannyYuyang-quic @cbilgin

[winskuo-quic]

Hi @imjking,
I have tried using the official HF (https://huggingface.co/Qwen/Qwen3-0.6B) to generate a prompt to verify the difference between user prompt for think and no think. It seems like the only thing you have to do is to append "<think>\n\n</think>\n" after the assistant, as shown below:

In this case, could you go to

executorch/examples/qualcomm/oss_scripts/llama/qnn_llama_runner.cpp

Line 163 in 40a18e7

formatted_prompt.append("<|im_start|>assistant");

and append "<think>\n\n</think>\n" right after this line?
After you added this line, please rerun https://github.com/pytorch/executorch/blob/main/backends/qualcomm/scripts/build.sh.
You can just do build.sh --no_clean --skip_x86_64
Thanks

[imjking]

@winskuo-quic Thanks very much.