During the community meeting today there was some discussion around the road to TVM 1.0 which made me think of this topic. I was wondering if anyone has considered having schemas for output artifacts from TVM, such as the Relay serialization and the AutoTVM tuning logs? I believe thus far the stance has been in a pre-1.0 world no guarantees are made, and I think while practically the “schemas” for these seldom change there are no guarantees from revision to revision. However for folks to build systems around TVM, especially long-lived systems where storing these records in long-term storage is useful, having a schema and a schema evolution story for these outputs will be immensely useful (and similarly not having one will be/is immensely painful).
Curious to get the community’s thoughts!