Lunderberg commented on PR #14901: URL: https://github.com/apache/tvm/pull/14901#issuecomment-1599011076
Agreed, the device-side codegen should not handle host-to-device calls. If any such calls exist (e.g. resulting from SplitHostDevice), they are lowered to `PackedFunc` API in `LowerDeviceKernelLaunch`. I'm investigating the issue at the moment, and can reproduce it locally. (Thank you @LeiWang1999 for the test case!) It looks like it is a discrepancy between the `PrimFunc::ret_type` and the `CallNode::dtype`, currently checking where that discrepancy gets introduced. -- 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]
