Engine cache and model security

[MickaMickaMicka]

I am using ONNX Runtime to generate a TensorRt engine from ONNX model file.
I am using trt_options.trt_engine_cache_enable = 1; and trt_options.trt_engine_cache_path = "./path"; to generate the engine file once and to load it on later runs, which works well (speeds up strongly).

However I'm not sure about what exactly that engines file is.
Does it include models weights?
Can it be used to load the full model, without having access to the .onnx model file?

My question arises because we are encrypting our model files and load them from RAM during runtime, so that other people with access to the system don't have access to our models. If the cache is basically a full access to the model, we will need a different solution.

[asfiyab-nvidia]

@yuanyao-nv can you help take a look at the ONNX runtime query?

[yuanyao-nv]

Does it include models weights?
By default the engine will include the weights.

Can it be used to load the full model, without having access to the .onnx model file?
Once the engine is built, the onnx model is no longer needed for inference. Note that the engine only works for the configs you specify during engine build (optimization profile, precisions, GPU version, etc.)

See more info in our developer guide: https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/index.html#prog-model