使用qwen3-vl-2b模型构建qnn模型报错

[rossipang]

根据官方教程https://mnn-docs.readthedocs.io/en/latest/transformers/llm.html 运行如下大模型转换命令

python3 llmexport.py --path Qwen3-VL-2B-Instruct --export mnn --smooth --act_bit=16 --quant_block=0 --lm_quant_bit=16 --seperate_embed --quant_bit=4 --sym --act_sym

成功后生成如下模型文件 -rwxrwxrwx 1 root root 342 Nov 10 11:57 config.json* -rwxrwxrwx 1 root root 230 Nov 10 13:47 config_qnn.json* -rwxrwxrwx 1 root root 622329856 Nov 10 11:57 embeddings_bf16.bin* -rwxrwxrwx 1 root root 776 Nov 10 11:57 export_args.json* -rwxrwxrwx 1 root root 489120 Nov 10 11:57 llm.mnn* -rwxrwxrwx 1 root root 1931280 Nov 10 11:57 llm.mnn.json* -rwxrwxrwx 1 root root 1330262232 Nov 10 11:57 llm.mnn.weight* -rwxrwxrwx 1 root root 6335 Nov 10 11:57 llm_config.json* drwxrwxrwx 1 root root 512 Nov 10 13:47 qnn/ -rwxrwxrwx 1 root root 3193555 Nov 10 11:57 tokenizer.txt* -rwxrwxrwx 1 root root 502512 Nov 10 11:57 visual.mnn* -rwxrwxrwx 1 root root 238226780 Nov 10 11:57 visual.mnn.weight*

执行使用 npu/generate_llm_qnn.py 构建 qnn 模型时候报错:

python3 npu/generate_llm_qnn.py --model model --soc_id=532 --dsp_arch=v73

Step1: Make IO Segmentation fault (core dumped)

Cost: 0.2031416893005371 s Step2: Seperate Model model: /mnt/e/projects/MNN-new/transformers/llm/export/model/llm.mnn Segmentation fault (core dumped) Cost: 0.24718451499938965 s Step3: Compile to QNN /opt/qairt/qnn_sdk Traceback (most recent call last): File "/mnt/e/projects/MNN-new/transformers/llm/export/../../../build/../source/backend/qnn/npu_convert.py", line 10, in with open(sys.argv[1]) as f: ^^^^^^^^^^^^^^^^^ FileNotFoundError: [Errno 2] No such file or directory: 'npu_postreat.json' Cost: 0.11310124397277832 s Step4: Move result file to model End

[Nr-rN]

您好，请问问题解决了吗？我也没有找到有关npu_postreat.json任何存储或者生成的地方

[Qxinyu]

这个第一步Make IO失败了，现在工具转换Qwen3-VL还有点问题，我们后面会更新修复。