Hello,
I am building PyTorch from source on an aarch64 system and have noticed performance differences of Hugging face models(Bert, resnet etc) when running on 64 threads/cores, compared to the pip-installed version (2.4.0). Upon inspecting the shared libraries linked to libtorch_cpu.so, I found that the pip-installed version links to librt.so.1, libdl.so.2, libz.so.1 and libpthread.so.0, but these libraries are missing in the source build.
I used following method to build:
[builder/aarch64_linux/aarch64_wheel_ci_build.py at main · pytorch/builder · GitHub].
Please help me out to build pytorch similar to pip-installed version including all missing libraries.