[GitHub] Caenorst opened a new pull request #14040: Reformat of TensorRT to use subgraph API

[GitBox <gi...@apache.org>]

Caenorst opened a new pull request #14040: Reformat of TensorRT to use subgraph API
URL: https://github.com/apache/incubator-mxnet/pull/14040
 
 
   ## Description ##
   This PR modify TensorRT usage by relying on the Subgraph API, the backend is named 'TensorRT'
   ## Checklist ##
   ### Essentials ###
   Please feel free to remove inapplicable items for your PR.
   - [x] Changes are complete (i.e. I finished coding on this PR)
   - [x] All changes have test coverage:
   - [ ] Code is well-documented: 
   - Need an update in documentation for general use cases
   - For user-facing API changes, API doc string has been updated. 
   - For new C++ functions in header files, their functionalities and arguments are documented. 
   - [ ] To the best of my knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change
   
   ### Changes ###
   - TensorRT now just rely on the Subgraph API, and a contrib function for loading weights
   - Add FP16 support, computation in FP16 forced by default
   - Graph partitioning have changed so if the same input of a subgraph is used multiple times in the subgraph then it stay as a single node in the subgraph
   
   ## Comments ##
   - Now that it's using the Subgraph API, you can use TensorRT with both module API and symbolic API
   - The way weights are loaded in the TensorRT node is a little bit dirty, we will do a follow up PR (including one on NNVM) to natively have a node attributes own subgraph parameters including a setter on the c api
   

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services