ganler commented on issue #11696: URL: https://github.com/apache/tvm/issues/11696#issuecomment-1154657453
Thanks for the bug report! However, I think using invalid numerics (i.e., "nan" or "inf") as inputs should be considered undefined behavior. "Fixing" may require extra checking overhead. According to my experience, there will be lots of such "bugs" in fast math. For example, many PyTorch [fast math](https://github.com/pytorch/pytorch/blob/429a80dded856cd06bd12bd9118efce5ac13dba3/torch/csrc/jit/tensorexpr/expr.cpp#L150) impl does not consider invalid inputs. I guess that is because generally, fast math targets best performance with acceptable approximation over valid domains. Personally, I will avoid fuzzing bugs triggered by NaN/Inf inputs as they easily lead to false positives. For example, IEEE 754 even "bans" trichotomy for NaN so that `nan == nan` returns `False` in PyTorch/Numpy, etc. (You can email me if you are interested in more false-positive stories. :-) Let me also cc @junrushao1994 for more professional opinions. -- 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]
