As part of torch::deploy, we’ve been prototyping a way of loading multiple Python interpreters into a single process using a custom shared library loader as a way to get around the global interpreter lock. With this approach many C extension libraries such as Numpy work out of the box.
4 Likes