Cannot add ONNNX runtime library to final binary.

[trohit920]

I am trying to build a project involving Onnx runtime gpu library on Nvidia Jetson NX Board. I ran the

nvidia@ubuntu:/media/nvidia/Extreme_Pro/train_project$ pyinstaller --onefile --name="osn_tcms" --add-binary /home/nvidia/.local/lib/python3.8/site-packages/onnxruntime/capi/libonnxruntime_providers_cuda.so:lib --upx-dir=/usr/local/cuda-11.4/targets/aarch64-linux/lib/ main_v3.py --runtime-tmpdir /media/nvidia/Extreme_Pro/train_project/tmp -d bootloader

command to successfully build the final binary using Pyinstaller. But on running the final binay from dist folder gives me below error.

nvidia@ubuntu:/media/nvidia/Extreme_Pro/train_project$ ./dist/osn_tcms [283456] PyInstaller Bootloader 6.x [283456] LOADER: executable is /media/nvidia/Extreme_Pro/train_project/dist/osn_tcms [283456] LOADER: _MEIPASS2 is not set [283456] LOADER: archivename is /media/nvidia/Extreme_Pro/train_project/dist/osn_tcms [283456] LOADER: Cookie found at offset 0x23DFCCB9 [283456] LOADER: creating temporary directory... [283456] LOADER: Found runtime-tmpdir /media/nvidia/Extreme_Pro/train_project/tmp [283456] Cannot create temporary directory!

Please suggest what am I doing wrong and how to resolve this issue.

[rokm]

I think the directory pointed to via --runtime-tmpdir must already exist, and we try to create a _MEIXXXX directory inside that. So this is probably failing because /media/nvidia/Extreme_Pro/train_project/tmp does not exist.

But why are you overriding --runtime-tmpdir in the first place?

[trohit920]

Thanks for the quick reply after correcting it I encountered another issue.

./dist/osn_tcms 
[295386] PyInstaller Bootloader 6.x
[295386] LOADER: executable is /media/nvidia/Extreme_Pro/train_project/dist/osn_tcms
[295386] LOADER: _MEIPASS2 is not set
[295386] LOADER: archivename is /media/nvidia/Extreme_Pro/train_project/dist/osn_tcms
[295386] LOADER: Cookie found at offset 0x23DFCCB9
[295386] LOADER: creating temporary directory...
[295386] LOADER: Found runtime-tmpdir /media/nvidia/Extreme_Pro/train_project/tmp
[295386] LOADER: created temporary directory: /media/nvidia/Extreme_Pro/train_project/tmp/_MEIazijTR
[295386] LOADER: Extracting binaries
[295386] Failed to create symbolic link libQt5Core-9e162752.so.5.15.0!
[295386] LOADER: temppath is /media/nvidia/Extreme_Pro/train_project/tmp/_MEIazijTR
[295386] LOADER: Error extracting binaries

[rokm]

Hmmm, we fail to create symlink for libQt5Core-9e162752.so.5.15.0 for some reason. Is /media/nvidia/Extreme_Pro/train_project/tmp on a filesystem that does not support symbolic links?

Can you try without --runtime-tmpdir to have the application unpack itself in the default temporary directory (presumably /tmp)?

As a side note, libQt5Core-9e162752.so.5.15.0 likely comes from python-opencv, and if you are not using OpenCV's HighGUI for GUI, you should probably be using python-opencv-headless instead...

[trohit920]

The /media/nvidia/Extreme_Pro/train_project/tmp is an external hard drive as the Jetson module is not having enough space in /tmp folder. So is there any work around to this.

[rokm]

The /media/nvidia/Extreme_Pro/train_project/tmp is an external hard drive as the Jetson module is not having enough space in /tmp folder. So is there any work around to this.

If the fileystem does not support symbolic links, then no. Except maybe using PyInstaller 5.x, which collects only hard copies.

[bwoodsend]

If your application needs so much space that it doesn't fit in /tmp then you probably shouldn't be using onefile mode.