When auto-tuning is interrupted by some errors, Can I start tuning from interrupted tasks?
It’s a waste of time to auto-tuning from scratch.
1 Like
Unfortunately no at this moment, but I also have this concern so I’m planning to add a feature for it.
You’re also welcome to file a PR if you have an implementation already.
You could load the .tmp log file again and start the task you interrupted. For example, we have 20 tasks but interrupted in 9th task. You could set tasks = tasks[8:] and tune again. But make sure your script doesn’t delete .tmp log file.
In this case you still need to re-tune 9th task from scratch if it was interrupted, but yes this could be a quick solution to reduce the lost.
Yes. But it is much better than from restart from 0th task and only just retune one single task.
Thank you for your reply, I will try it.
I recommend taking a look at [solved]Can we resume an autotuning session?, which has a similar discussion and some more detailed solutions.
1 Like
I’ve been thinking about a more robust solution to this for a while and have some scripts that implement a sort of AutoTVM ‘cache’ that I use when autotuning. I’ll try and turn these into a PR and open an RFC to see if anyone would be interested.
2 Likes
Please do! And feel free to link the RFC in this thread too.
I’d love to see an RFC for that as well.