comaniac commented on pull request #8775: URL: https://github.com/apache/tvm/pull/8775#issuecomment-902349542
Thanks for the response :) I completely understand the concern of unifying the lowering mechanism and feel this is the right track as well. Meanwhile, I have two cases: 1. We do lower Relay function one-by-one with JIT. It seems to me that this might be a reasonable exception, because we could naturally have a JIT compiler that compiles each function when visiting them at the first time. Meanwhile, the current Relay interpreter wraps each function to a single module and performs lowering, which seems not necessary but a workaround. cc @icemelon @yzhliu 2. The case of using Python could be a relatively miner. Specifically, if we lower a Relay function using TE compiler directly in C++, then we have no way to configure AutoTVM and AutoScheduler (e.g., tracing and logging mode), since both tuning frameworks have some logic purely in Python. Finally, while I agree that TE compiler could be as private as possible, it doesn't hurt to keep the Python interface for debugging and experiments. We just need to make sure they are only used for developments and no one really uses them in the submitted PRs. -- 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]
