dlexplorer opened a new pull request #7387: URL: https://github.com/apache/tvm/pull/7387
In case of non cpython flavour of python, the task passed to measure process should be serialized using pickle approach. The task includes workload which is a list of Tensors. This list should be serialized and deserialized as an atomic object. If list is not serialized as the atomic object, each tensor will be serialized/deserialized independently. And during deserialization of output tensor, it will recreate input tensors one more time instead of reusing existing from the list because it will not know that input tensors are created during deserialization of list elements. As a result, the function `SchedulePostProcToPrimFunc` will fail on assert in GetOrAllocBuffer dues to missmatching of input tensors. ---------------------------------------------------------------- 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]
