inference mode issue

[beijinggao]

After exporting the model for prediction, I found that setting de.enable_inference_mode() is required.

and comment is : de.enable_inference_mode()

TrainableWrapper.read_value` is not thread-safe that causes threads competition and Out-Of-Bound exception in concurrent serving scenario.

To resolve this, we define the ModelMode APIs to instruct the TrainableWrapper to build a different thread-safe sub-graph for 'TrainableWrapper.read_value' on inference mode.

But after export the model and use tensorflow serving or trion to inference，how to set enable_inference_mode()，or is it necessary ?

[MoFHeka]

model = build_model(xxx)
de.enable_inference_mode()
model.save(export_dir)

enable_inference_mode is used to change the graph building logic inner TFRA. It would be reduce two times memory copy in TrainableWrapper which are IDs and Embedding Values.