comaniac commented on pull request #6068: URL: https://github.com/apache/incubator-tvm/pull/6068#issuecomment-659059599
> My initial thought is that putting this into partition_graph looks a bit odd as running optimizations seems outside of the scope of the pass. What's motivated that instead of just having the codegen itself run the optimizations it needs? The most important reason has been demonstrated in #5915. Since the codegen should return a runtime module without mutating the graph (e.g., run transpose), running optimizations inside the codegen results in inconsistency between the module Relay/TVM processed and the module codegen processed. In the ACL case, it cannot use the unified weight serialization mechanism from MetadataModule but has to deal with by itself. This is not only tedious but also increases the binary size, because MetadataModule still maintains the original weights that will never be used by ACL runtime. ---------------------------------------------------------------- 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. For queries about this service, please contact Infrastructure at: [email protected]
