junrushao1994 commented on PR #74: URL: https://github.com/apache/tvm-rfcs/pull/74#issuecomment-1159307847
Thanks @areusch for pointing me to the thread! Definitely happy to read the [discussion](https://discuss.tvm.apache.org/t/export-tir-to-json/12329), and glad to see that @vinx13 unblocks the `SaveJSON` method for NDArrays :-) completely agreed that TVMScript could be the usecase which provides readability if this is the case that @SebastianBoblestETAS and @MichaelJKlaiber wanted. > what functionality does JSON provide that TVMScript doesn't do right now as it's designed? Is there a reason to have two formats to begin with? imo it is not worth having two formats here unless there's a clear need for structural analysis of TIR outside of TVM. Yeah JSON doesn't provide any extra functionality, but it's pretty stable and reliable protobuf-like serialization, which could be faster to parse/dump if we only want to save the IRs or send them around between processes without the need of readability and manipulatability. For general users, our suggestion should be like, just use TVMScript to write TIR and forget about JSON. For advanced developers, for example, we want to store TIR in tuning database, cache TIR somewhere for later usage, JSON is definitely the "to-go" approach. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
