Hi, I wonder whether current autoTVM supports auto tuning for operations with dynamic shape ? Are there any tutorials?
No it’s not supported yet at least in the upstream. Although we’ve been talking this for a while, we haven’t really got a chance to make a concrete plan.
Thanks for answering! By the way, can we schedule operations with variable tensor shapes as we can do with static tensor shapes? Say, bind or split a given axis at certain stage. I am trying to set up scheduling process by hand written scheduling template, but I cannot figure out how to perform those scheduling primitives for operations in “relay.nn” since I am using “Any” to create variable shaped buffers. Thanks a lot!
I find “te.var” could solve my problem. Thanks a lot!
Yeah. We didn’t do auto tuning on symbolic shape yet, but scheduling symbolic shape operators is well supported