Kh4L opened a new pull request #18490: URL: https://github.com/apache/incubator-mxnet/pull/18490
## Description ## This PR upgrades the MXNet TensorRT integration to use the new TensorRT 7.0 API and refactor it to fully use the new Partition API (described in #17623 and previous related PRs). This allows the user to use the new `optimize_for` method to run the full optimization (graph partitioning + TensorRT creation) in a single pass instead of the previous `get_backend_symbol` + `init_tensorrt_params` two stage optimization. ### Changes: - Update the TensorRT API calls to TensorRT 7.0 - The new Partition API adds a `PrePartition` member function to `SubgraphProperty` that allows us to modify the graph before the partitioning and Subgraph optimization pass. It also give us a way to access the params (args and aux) during the partitioning. Add PrePartition to the TensorRT partitioning backend and change to logic to pass the params to the TensorRT backend directly during the optimization (hence making obsolete `init_tensorrt_params`). - Update the tests to use `optimize_for` instead of `get_backend_symbol`. :warning: This PR requires #18378, #18405 and #18350 to be merged. ---------------------------------------------------------------- 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]
