My questions are:
- Does TVM guarantee that the library of operators compiled from it will be correct?
- If so, how is it guaranteed?
- Is there a formal proof of that?
- Will TVM test every release version in large quantities?
- If so, to what extent do these tests cover?
‘correct’ means every inference process will not crash and the result is consistent with the result expected by the model