Thanks @manupa-arm for slogging through. > This might enable bring us to down to get the cost estimation but it will > still explore the combinatorial partitioning possibilities of ops x backends, > right ? (lmk, if I am wrong here).
Any sensible implementation of CostEstimator should cache by (IRModule, Target) to exploit sharing of partition measurements in the same model, or minor variations of an already explored model. I have no idea what kind of cache hit rate we'd see in the wild though. But I think that matches your suggestion. > I suppose having the abstraction might help a user to play around with what > Collage generates as a base Yeah, agree. I'll make the 'search' vs 'rewrite' steps more officially separate. After all, the paper uses the optimal partitioning from the DP search as a starting state for the evolutionary search, so might as well embrace that possibility. I'll piing back after going through your detailed comments. Thanks again. -- Reply to this email directly or view it on GitHub: https://github.com/apache/tvm-rfcs/pull/62#issuecomment-1079333681 You are receiving this because you are subscribed to this thread. Message ID: <apache/tvm-rfcs/pull/62/c1079333...@github.com>
