The pull request https://github.com/dmlc/tvm/pull/1487 introduced a requirement on having access to the internet to download pretrained schedules if they haven’t already been downloaded. I would like to put a try except block around that to allow for it to succeed even if the download fails. Would this cause any problems?
@merrymercy I would especially appreciate your feedback.
Thanks
Thanks for pointing this out. You can add a try except block in autotvm.tophub.check_package, and print a warning if it fails.
Cool, sounds good I’ll submit a pr once its ready.
When I allow for compilation to proceed without downloading the pre-tuned parameters, one of the topi unit tests fails because it can’t find the schedule. Shouldn’t it generate a decent schedule if it can’t find a pre-tuned one so the user can compile their model even if it isn’t optimal?
I added the try except block and fallback support in https://github.com/dmlc/tvm/pull/1566.
But the I only enable fallback in that broken test.