[Feature Request]

[lnotgm]

Describe the feature request

Can onnxruntime support directly loading *.engine or *.trt to initialize the session when using TensorRT EP?

Describe scenario use case

Currently using TensorRT EP, in order to ensure the initialization speed, the *.engine is written to the hard disk when using the cache. In my production environment, I want *.engine to be encrypted. And they can be directly loaded in the following way.
//Decrypt model to byte
byte=Decryptor.read("./encrypted_model.engine");

//Initialize to session
session = Ort::Session(env, byte.data(), byte.size(), session_options);

[jywu-msft]

@chilo-ms will share details

[chilo-ms]

@lnotgm
The "model" input to Ort::Session() should be ONNX format regardless of file path or byte stream.
TensorRT EP provides the feature of using an embedded engine model, similar to your request, it's just that the input model is a wrapper of engine (still an ONNX file). This can also fulfill your request of almost directly loading the *.engine to reduce session initialization time.
Please see the detail of using "Embedded engine model / EPContext model":
https://onnxruntime.ai/docs/execution-providers/TensorRT-ExecutionProvider.html#tensorrt-ep-caches