Recently, we have open sourced matx: an ahead-of-time compiler for Python. Github:

The overall infrastructure of matx borrows from TVM. Matx is widely used in Bytedance. The magnificent software architecture design of TVM significantly faciliates the development of matx!

Currently, matx aims to integrate with PyTorch/TensorFlow/TVM to server end-to-end applications.

In the future, we hope to add numpy support via IR/Codegen. Everyone is welcome to contribute if interested.

Thanks again!


In addition, thanks to @ziheng @Laurawly for their great help in the early stage.

Thanks @xiandi for sharing, looks like a great usage of the FFI and object type systems.

We have a much stronger emphasize of FFI-object runtime in the TVM unity push – relax will be built with object as first class and I can see many interesting things like Dict, Unicode come natively to tvm unity runtime.

Would love to learn lessons and bring that back to TVM unity so we can better support applications like matx in the common runtime and enable a lot more interpolations together.

1 Like

Thanks to @xiandi, we have drafted an RFC and bring some of the lessons back to the tvm community [RFC] Further Unify Packed and Object in TVM Runtime - #2 by tqchen

Thanks to @tqchen! Due to the time difference, my reply is delayed.

Hope to have the opportunity to write code together, thanks!