Libtorch built from source is 10 times slower than the official binary

I tried both and use cmake myself. But the lib built from source is very slow.
I use the default cmake config.
Is there anyone know the official config ?

Depending on what you are doing, you might be missing accelerator libraries like MKL or CuDNN.
You should get relatively decent diagnostics what is included from cmake and compare that with what you get when you follow the official instructions for building PyTorch from source.
Indeed, the official libtorch used to indeed just build with python and then extracted from the Python build (but I am not sure if that is still the case).

Best regards