areusch opened a new pull request #7250: URL: https://github.com/apache/tvm/pull/7250
In porting up the µTVM AutoTVM code, seems that `check_correctness` assumes that each AutoTVM task is instantiated on the `llvm` builder with the same input shape as is used on the target. Even in tuning a conv2d for x86, I see conv2d_NCHWc used with AutoTVM trying different combinations of `C` and `c`. I don't think this approach is viable and from chatting with a few people it seems like it was a debug tool that might just be deprecated now. I propose we remove until we have something more dependable via AutoScheduling or AutoTIR. @merrymercy @tqchen @jwfromm @junrushao1994 @tmoreau89 ---------------------------------------------------------------- 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. For queries about this service, please contact Infrastructure at: [email protected]
