Is this folder created by djl?

[Guibeen]

I noticed this folder in my service environment, and I find it is created during the launching of service.

/home/map is the home directory. I know that DJL downloads and cached pytorch libs by default and .djl.ai/pytorch is the default cache directory, so is this folder the very libraries downloaded? If the answer is yes, which djl class is doing this operation?

The folloing DJL dependencies are added to the pom.xml, and the jars are packaged to the service libs directory. Is the folder unzipped from the existing jar, pytorch-native-cpu-precxx11-1.13.1-linux-aarch64?

[frankfliu]

@Guibeen

Yes, the folder is created by DJL, DJL extract shared libraries from jar file:

1. 1.13.1 is the version, in native/lib/pytorch.properties file
2. 20221220 is the jar creation time
3. cpu-precxx11 is the pytorch flavor
4. linux-aarch64 is the platform

You can find the code here: https://github.com/deepjavalibrary/djl/blob/master/engines/pytorch/pytorch-engine/src/main/java/ai/djl/pytorch/jni/LibUtils.java#L324

[Guibeen]

Thank you for your answer! In fact, I'm facing a problem with libstdc++.so.6 in pytorch-native-cpu-precxx11-1.13.1-linux-aarch64.jar: its GLIBCXX version is up to 3.4.19, but my program requires 3.4.21. How can I solve this, is it possible to organize the .so files to a new directory, and add its path to env var PYTORCH_LIBRARY_PATH?

[frankfliu]

@Guibeen
you can set env var or system properties: PYTORCH_LIBRARY_EXCLUSION=libstdc++.so.6

It will load your system libstdc++.so.6 instead of bundled version.

[Guibeen]

I have added these commands in the install shell script.

export PYTORCH_LIBRARY_EXCLUSION=libstdc++.so.6
export PYTORCH_LIBRARY_PATH=${HOME}/.djl.ai/pytorch/1.13.1-cpu-precxx11-linux-aarch64
export PYTORCH_FLAVOR=cpu-precxx11

but it seems not working for PYTORCH_LIBRARY_EXCLUSION. Is there any difference in the processing of these env vars?

[Guibeen]

it is finally solved by organizing the .so files to a new directory, and adding its path to env var PYTORCH_LIBRARY_PATH