The default Python implementation is very slow, using PyPy may accelerate applications that is built on top of TVM python interface. But it seems TVM has no support for PyPy. Is there any way to use TVM with PyPy?
The first thing I would recommend trying is to build with cython enabled, this usually gives us around 1ms level of FFI overhead and bring it close to numpy.
If you have particular part of code in mind, we can also discuss how to migrate some of the code to C++. I am not too sure if PyPI will speedup execution, but would be an interesting topic to discuss. It would nice to talk about what are lacking(perhaps the FFI part?)
Moving my code to C++ is favorable. It seems PyPy is not well accepted in the area of ML. For example, I can’t install PyTorch in PyPy.