merrymercy commented on issue #6096: URL: https://github.com/apache/incubator-tvm/issues/6096#issuecomment-663238657
To conclude, the current cache mechanism in relay works globally across different calls to `relay.build`. If we want to use two different schedules for the same relay subgraph in two different `relay.build`, we have to call `compile_engine.get().clear()` between two `relay.build`. In @tkonolige 's test, the autotvm task extraction calls `compile_engine.get().clear()` internllay, making @tkonolige draw the conclusion that task extraction is required. But actually `compile_engine.get().clear()` is what we need. To fix the issue, we all agree that we should make the relay cache work only within a single module. @tkonolige Could you send the fix? ---------------------------------------------------------------- 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]
