mbs-octoml edited a comment on pull request #62: URL: https://github.com/apache/tvm-rfcs/pull/62#issuecomment-1077915271
> Is there a model being developed to estimate latency of the chosen partitioning? No. As far as Collage is concerned it just calls the abstract CostEstimator::Estimate interface for each candidate partition, and can remain ignorant as to where those costs come from. In the prototype it is hard coded to tune, build and run locally to help us get going. Here at OctoML we'll need to instantiate the interface to connect to our production tuning/running/caching systems. In principle the interface could also be instantiated to some analytical model a la the cascading scheduler, though we're not planning to build one. I'm thinking the CostEstimator object can just be passed into the CollagePartitioner to leave this extension point open. > How can a user export the search done by collage ? In an early draft I had support for that. Basically it just needs some materialization of the 'optimal' partitioning as an Array of CandidatePartitions. The method CandidatePartition::ParallelRewrite can take that array and rewrite the whole expression, exactly as the Collage pass would have done. So splitting between search and rewrite is pretty easy. But I ended up dropping it all in favor of just relying on the CostEstimator to cache all measurements (which it needs to do anyway given all the sharing opportunities). Firstly, it's not yet clear if there's any significant compile time advantage to bypassing the collage search if every candidate partition to estimate results in a cache hit. I figured I'd at least measure that before adding a fix. But secondly, if someone (the service, the user) is going to go to the trouble of caching the optimal partitioning for a particular (model, targets) pair, why not just cache the built artifact directly and skip all the bother? However let me know if I over simplified and I can add that part back. > think we would need to follow the template Ok. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
