Luchang Li

I found it's because the envs are not set properly.

you need modify the aimet_common/bin/envsetup.sh to set cuda env and set LD_LIBRARY_PATH and PYTHONPATH properly source /root/anaconda3/envs/aimet/lib/python3.6/site-packages/aimet_common/bin/envsetup.sh export LD_LIBRARY_PATH=/root/anaconda3/envs/aimet/lib/python3.6/site-packages/aimet_common/x86_64-linux-gnu:/root/anaconda3/envs/aimet/lib/python3.6/site-packages/aimet_common/:/root/anaconda3/envs/aimet/lib/:$LD_LIBRARY_PATH export PYTHONPATH=/root/anaconda3/envs/aimet/lib/python3.6/site-packages/aimet_common/x86_64-linux-gnu:$PYTHONPATH

> ArmNN Version：armnn-22.05.01 TfLite version: 2.15 System OS: android Model information：I can not directly provide the model, but many transformer based NLP models contain gather op.

Is there a way to disable conversion of specific type of node? Currently I bypass this bug by replacing the softmax by other node type, such as softplus and recovering...

> Hi @luchangli03, > > It's a known limitation: #4101. However, if the input shape for that Softmax node does present and you still see this error, there might be...

> Hi @luchangli03 , > > What version of ArmNN/Compute Library are you using? > > If you are using v22.05 or v22.08, > > Could try this option: >...

I met the same problem when export an other model, and finally I found I set wrong input_names and dynamic_axes, thus the order input_names and dynamic_axes mismatch. After I solve...

你好，能否只根据vector数组大小进行展开传递呢？因为有时候func是个指针，无法根据func来获取函数参数个数，函数返回值为void。

> 你好，单卡a6000 50g上导出13b baichuan会oom，正常读取显存30g足够了，onnx导出会占用更多显存么，有无解决办法，谢谢 fp16也不行？

> 您好，请问下脚本可以用来转llama2和其他llama家族的模型吗？ 应该没太大问题，可以先直接转下试试，有问题再fix