comaniac opened a new pull request #7377: URL: https://github.com/apache/tvm/pull/7377
Per discussion in https://discuss.tvm.apache.org/t/ansor-task-scheduler-spending-too-much-trial-on-dead-end-tasks/8955, this PR improves the early stopping criteria in the task scheduler. Now we have two early stopping factors: - `TuningOptions.early_stopping`: Like other options in this object, the `early_stopping` in TuningOptions affects each task. - `task_scheduler.tune(..., early_stopping=None)`: Since this is only for task_scheduler, this `early_stopping` applies to an entire tuning process. Accordingly, when setting `TuningOptions.early_stopping`, a task will be stopped early if it has no improvement for N measurements, but the task scheduler will keep tuning rest tasks. On the other hand, `task_scheduler.tune(..., early_stopping)` serves as the original functionality that early stops an entire tuning process after no improvement for N measurements (regardless which tasks). cc @merrymercy @masahi ---------------------------------------------------------------- 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]
