zxybazh opened a new pull request, #14624: URL: https://github.com/apache/tvm/pull/14624
Based on feedback for `DefaultGPUSchedule` pass, we noticed that while PrimFuncs can be supported without tuning on GPU devices, the performance is not competitive. Therefore we introduce a new FewShotTuning pass for static shape PrimFuncs that directly samples from the search space and make sure the returned candidate can pass post-processors. By design, this pass enables most of the performance optimizations according to the target, and avoided tuning by: - Disable benchmarking by default. - Only do random sampling. - Limit valid trials to user specified small number, by default 1. - Skip tuning apis and directly do sampling and compilation. - Still guarantee result is compilable. This PR is a follow-up for @psrivas2's work and based on discussion with @sunggg and @psrivas2. CC: @jwfromm @junrushao -- 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]
