areusch commented on PR #14665: URL: https://github.com/apache/tvm/pull/14665#issuecomment-1514978253
@Hzfengsy yes it should work for FuseOpsByPattern (I've updated CompositeFunctionAnnotator to make this work). @junrushao I agree there are limitations here when it comes to Spans in deep schedule optimizations. However, it provides considerable introspection into the graph-level optimizations done by the compiler. I'm not attached to the way this information is tracked, but I do think it's reasonable to want to know which ops got fused how. I see this as a way to surface that information. What's the thinking behind building a tracking method that works for Spans (or similar provenance) inside scheduled compute before surfacing graph-level fusion information? At present, this method is essentially confined to Call() in the `main` function. Any such system to track in-compute spans should be a superset of this, and it seems like it'd make more sense to wait for a concrete use case before we design it. The concrete use case for this tracking is: I want to know which Relax ops were generated as part of a model import and which ones were fused together. -- 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]
