qwen2.5 0.5b inference with llama.py is normal, but qwen3 0.6b repeats

Using llama.py to convert qwen2.5 0.5b produces correct responses. The conversion command is as follows:

python examples/qualcomm/oss_scripts/llama/llama.py \
  -b build-android \
  -s 4c576cc \
  -m SM8650 \
  --temperature 0 \
  --model_mode hybrid \
  --max_seq_len 1024 \
  --prefill_ar_len 128 \
  --decoder_model qwen2_5-0_5b \
  --prompt "I would like to learn python, could you teach me with a simple example?" \
  --tasks wikitext \
  --limit 1 \
  --compile_only

Then use the following command for inference:

python examples/qualcomm/oss_scripts/llama/llama.py \
  -b build-android \
  -s 4c576cc \
  -m SM8650 \
  --decoder_model qwen2_5-0_5b \
  --model_mode hybrid \
  --prefill_ar_len 128 \
  --max_seq_len 1024 \
  --prompt "Write a bubble sort in C++." \
  --pre_gen_pte /executorch/llama_qnn

The response is correct as follow:

Using the same method, converting qwen3 0.6b with llama.py:

python examples/qualcomm/oss_scripts/llama/llama.py \
  -b build-android \
  -s 4c576cc \
  -m SM8650 \
  --temperature 0 \
  --model_mode hybrid \
  --max_seq_len 1024 \
  --prefill_ar_len 128 \
  --decoder_model qwen3-0_6b \
  --prompt "I would like to learn python, could you teach me with a simple example?" \
  --tasks wikitext \
  --limit 1 \
  --compile_only

Then perform inference using the following command:

python examples/qualcomm/oss_scripts/llama/llama.py \
  -b build-android \
  -s 4c576cc \
  -m SM8650 \
  --decoder_model qwen3-0_6b \
  --model_mode hybrid \
  --prefill_ar_len 128 \
  --max_seq_len 1024 \
  --prompt "Write a bubble sort in C++." \
  --pre_gen_pte /executorch/llama_qnn

The response is repetitive as follows:

how to slove the problem? thanks !

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

Start with examples/qualcomm/oss_scripts/llama/llama.py and reproduce the qwen2_5-0_5b and qwen3-0_6b conversion and inference commands on SM8650. Compare the generated responses and conversion settings; done means qwen3-0_6b produces a non-repetitive response for the supplied prompt.