haojin2 commented on issue #11795: [WIP] Fix problematic backward of take & embedding URL: https://github.com/apache/incubator-mxnet/pull/11795#issuecomment-407580355 @KellenSunderland One argument that I would like to make here is that end-to-end performance is what matters most to actual users, since that's what they would experience on their end. On the other hand, as the LargeBatch kernel is exhibiting flaky problematic behavior, I don't think it should be kept (I think one other guy on the dev list thread about this is also agreeing with this point), not to mention here it's not even making any fundamental performance difference. In conclusion, I think the LargeBatch version should be got rid of anyways, but I shall investigate whether the new general take grad kernel or the AddTakeGrad kernel is giving a better performance using some kind of profiler, does that sound good to you?
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
