Hello,
I am interested in using MetaSchedule for my own accelerator. As far as I understand the RFC, it should be possible to follow a similar strategy to AutoTVM with MetaSchedule, i.e. template based scheduling. However, I can’t find any tutorials on MetaSchedule expect. Is there something available that I just can’t find? There are a few examples in the test folder, but they don’t seem to cover template based scheduling.
Hi,
Same question here. I am able to locate the metaschedule folder in the TVM repo, but I cannot find any template based scheduling in the repo , or any internet resource.
For instance, Auto Tuning a Convolutional network has tutorials like : Auto-tuning a Convolutional Network for NVIDIA GPU — tvm 0.15.dev0 documentation (apache.org)
Is there a tutorial similar to this to use metascheduling?
Also @Necrotos , were you able to find any tutorial for this since you posted here? Thanks.
@masahi @jwfromm @tqchen @AndrewZhaoLuo @zxybazh @thierry @junrushao kindly look into this and provide some clarity. Thank you so much.
I found this site here that gives a few examples. It doesn’t cover everything with the detail level I would like, but it’s a good start.
Thanks @Necrotos for pointing to the MLC document, I think for now this tutorial would be a good start to get familiar with MS tuning. Another useful introductory tutorial might be the TVMCon recording for MetaSchedule.
Other than these, another example of tuning for onnx could be found here: https://github.com/apache/tvm/blob/748882aae7be1435f042e22b0fc67cb236705b6c/python/tvm/meta_schedule/testing/tune_onnx.py