Lunderberg commented on PR #10907: URL: https://github.com/apache/tvm/pull/10907#issuecomment-1092904174
I like the general division between launcher and session, and that mirrors the general compile-time / run-time split that exists throughout TVM. From what I can tell, there's two distinct use-cases that are pointing toward the different preferred designs. In a production environment repeatedly running the same model, the file copying should only be done once, with each session loading the model that has already been uploaded. In this use case, the saved binary file is fundamental; created by the user, uploaded by the launcher, and loaded by the session. In a testing environment where each model is different, the file copying must be done prior to each session. In this use case, the saved binary file is a temporary intermediate, whose entire purpose is replicate the local built module onto the remote session. I think passing the executor factories makes sense to support the second use case, because those are the objects that have sufficient information to fully describe the state being replicated onto the remote session. -- 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]
