What is the best approach to convert a pytorch model to TVM?

I remember seeing this PR, which is not merged yet, mentioning torch 1.6. Maybe @masahi can comment here.