comaniac commented on pull request #6323: URL: https://github.com/apache/incubator-tvm/pull/6323#issuecomment-687267906
Although it might be more code to have a customized runner (it's not that mach to be honest), you can keep the customized runner locally so it sounds fine to me. To me this is the flexibility of AutoTVM. You can customize everyone for different purposes, but only upstream the generic one after solid evaluations. The concerns I have for the solution in this PR is that you check and use an attribute `measure_batch` in `Tuner`, but `measure_batch` is not a member of that class. It means you manually add this attribute in runtime for this feature (i.e., `tuner.measure_batch = my_func`) to override the use of local variable. To me, this increases the uncertainty and it's hard to track and maintain in the future. More importantly, `create_measure_batch` is just a wrapper to call the runner, so it seems not necessary to provide the flexibility at this level. cc @merrymercy, @tqchen ---------------------------------------------------------------- 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]
