Lunderberg commented on PR #16687: URL: https://github.com/apache/tvm/pull/16687#issuecomment-2005455165
> However, when this subject was raised in design discussions before, the suggestion was turned down due to a preference of having only one pass (normalization) handle StructInfo, of which purity is a part. We could reopen the discussion, of course. I'd be interested in hearing the discussion on it. I wouldn't mind having all `StructInfo` handling done within a single pass, but I don't think the API of the existing normalization would allow it. The `tvm::relax::Normalizer` may receive any partially-constructed fragment of relax IR, not necessarily a full `IRModule`. The fragments are sufficient for performing local analysis, but full struct inference or purity analysis would require a similar inference of the called function. -- 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]
