Triton Inference Server Feature Requests

Have a sample Java client in TRTIS in addition to the Python and C++ clients already provided. This would be to support customers who's frontend is primarily in Java.

Be able to directly pull models from NGC model repo instead of having to set up TRTIS model repo in a persistent volume.

Some would rather have request die in the queue than wait for an available instance to be ready for inference. TRTIS should let you configure the queue to have different lengths/functionality

TRTIS should automatically generate the model configuration file based on the model's metadata. Today this works for Tensorflow and TRT. Enable this functionality for all model formats accepted by TRTIS

Metrics for % of cores on GPU used over time

Add basic CPU metrics for models run on CPU in TRTIS. ie CPU Usage, CPU Memory

Add MXNet support to TRTIS as a backend so native MXNet models can be served without any conversion through ONNX.