comaniac opened a new pull request #7145: URL: https://github.com/apache/tvm/pull/7145
The previous implementation of task serialization mechanism has a potential problem, which is a bit tricky so I didn't notice that before. Previously, I defined a set of attributes in the Python object (e.g., `hardware_params`) and use them to be the state when serializing a task object in Python. However, in the case that users don't provide `hardware_params`, the SearchTask constructor in C++ invokes `GetDefaultHardwareParams` to get the default hardware parameters. These default hardware parameters aren't exposed to the Python object because of the same name attribute. It wouldn't be a bit deal because when a task is deserialized, we follow the same process to get the default hardware parameters. However, since `GetDefaultHardwareParams` may access hardware context (e.g., CUDA context), this prevents tasks from being deserialized in parallel (with multiprocessing). This PR removes all attributes in `SearchTask` Python object to make sure we access the up-to-date C++ object attributes when serializing a task. The only exception is `self.dag` because we need it to reconstruct a ComputeDAG. ---------------------------------------------------------------- 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: us...@infra.apache.org
