i do not want to set "export QNN_SDK_ROOT"

[YeWenxuan64]

Traceback (most recent call last):
  File "/home/ubuntu/Softwares/NPU/ai-engine-direct-helper/samples/linux/python/yolov8_det/import_qnn.py", line 30, in <module>
    import qai_appbuilder
  File "/home/ubuntu/Softwares/python/python_venv/lib/python3.12/site-packages/qai_appbuilder/__init__.py", line 16, in <module>
    from .geniecontext import *
  File "/home/ubuntu/Softwares/python/python_venv/lib/python3.12/site-packages/qai_appbuilder/geniecontext.py", line 15, in <module>
    from qai_appbuilder import geniebuilder
ImportError: libGenie.so: cannot open shared object file: No such file or directory

i do not want to set some env on my shell, which make it hard to run python script
every time when import qai_appbuilder it needs the env in the shell. i tired to use os.environ["QNN_SDK_ROOT"]=我的QAI sdk路径, but it faild.
only set this

export QNN_SDK_ROOT=./QNN_SDK_ROOT(我的QAI sdk路径)
export ADSP_LIBRARY_PATH=$QNN_SDK_ROOT/lib/hexagon-v68/unsigned
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$QNN_SDK_ROOT/lib/aarch64-oe-linux-gcc11.2

to work

ps: i don't want to set any other env in .bashrc

[quic-zhanweiw]

Thanks @YeWenxuan64 for using QAI AppBuilder. We'll try to get a better solution.

[zhondai0129]

Hi @YeWenxuan64
you can add a new python script in ai-engine-direct-helper/samples/linux/python/run.py

import subprocess
import os
import argparse

env = os.environ.copy()
QNN_SDK_ROOT = '/home/ubuntu/code/2.37.1.250807'   # change your QNN SDK ROOT path
env['QNN_SDK_ROOT'] = QNN_SDK_ROOT
env['LD_LIBRARY_PATH'] = f'{QNN_SDK_ROOT}/lib/aarch64-oe-linux-gcc11.2'
env['ADSP_LIBRARY_PATH'] = f'{QNN_SDK_ROOT}/lib/hexagon-v68/unsigned'`
parser = argparse.ArgumentParser(description="model command such as yolov8_det")
parser.add_argument('model', help='Model name to run')

args = parser.parse_args()
script_path = f'{args.model}/{args.model}.py'

subprocess.run(['python', script_path], env=env)

if you want to run model script, such as yolov8_det:
cd ai-engine-direct-helper/samples/linux/python
python run.py yolov8_det

Best Regards

[YeWenxuan64]

Thanks @YeWenxuan64 for using QAI AppBuilder. We'll try to get a better solution.

Can we make it like this: when the user compiles QAI AppBuilder themselves with a parameter like "--with-qnn-sdk", then it will also link and package the SDK together in the wheel.

[quic-zhanweiw]

@YeWenxuan64 Good idea. We'll try it.

[YeWenxuan64]

@quic-zhanweiw can't wait to tase !!!

[YeWenxuan64]

When can a new version be released

[quic-zhanweiw]

@YeWenxuan64 Sorry, we're working on some other features recently. Perhaps we can have "--with-qnn-sdk" ready next week.

[YeWenxuan64]

You're too slow, I've already done it(maybe)🤣.
let me show:
first, pick the libs we need form qnn-sdk, then put them into a folder.

second, write some magic in the init.py of QAI AppBuilder.

third, write some magic in the qnncontext.py of QAI AppBuilder too.

then i ran it sucessfully with the yolo8_det.py

[YeWenxuan64]

快夸夸我ヾ(≧▽≦*)o

[quic-zhanweiw]

Great! Thanks @YeWenxuan64 for your sharing!

[quic-zhanweiw]

@YeWenxuan64
We updated a change for this. Could you please check if it works for you when you have time? Thanks!
ac9d681

[YeWenxuan64]

nearly perfect
here:

first.

second.
we still need some .sos. (about which one? IDK becouse i lived in UTC+8 so i get to sleep zzzzz~)

[quic-zhanweiw]

@YeWenxuan64
first: Could you describe it clearer?
second: Why do you need libCalculator_skel.so, libQnnSaver.so, libSnpeHtpV68Skel.so since QAIRT SDK don't support these libraries currently.

[YeWenxuan64]

@YeWenxuan64 first: Could you describe it clearer? second: Why do you need libCalculator_skel.so, libQnnSaver.so, libSnpeHtpV68Skel.so since QAIRT SDK don't support these libraries currently.

sorry~(´。＿。｀).i don'k know how qnn works so i copid all of them......

i will try them one by one later

[YeWenxuan64]

here is the result
it seems need the libQnnHtpV68Skel.so with a reason that i don't know

without the libQnnHtpV68Skel.so

with the libQnnHtpV68Skel.so

only tested on the yolov8_det example.(what if ohter models or AI functions need more .so libraries)