Is it possible to build a minimal libpytorch.so for c++ inference on mobile device?

Mobile
1

Hi,

For now we plan to follow the NativeApp demo in https://github.com/pytorch/android-demo-app, to support inference on mobile devices. After OPs reduction we still get libpytorch_jni.so with size about 11M for arm64-v8a. So,

  1. Is it possible to reduce it further? I have tried turned off some cmake options including USE_MKLDNN, but it seems no effect.
  2. Is there any other clean way to build a c++ dedicated library like libtorch_cpu.so getting rid of jni stuff, to support c++ native linkage.

Thanks!

2

Yes, the custom build functionality allows you to build a libtorch with just the PyTorch ops you actually need.

Best regards

Thomas

3

Thanks for reply!

Except operator customization, is it possible to allow reducing the binary size further, e.g., via the build options?

Thanks.

4

Most of these should already be turned off for PyTorch mobile.

5

Get it, thanks a lot!

BTW, do we have existing option to support building a native library for dedicated c++ linkage, getting rid of jni stuff.

Thanks!