[Documentation] Clearer TRT dependencies

[gedoensmax]

Describe the documentation issue

Is it true that the TRT EP needs the same dependencies as CUDA EP ? The documentation kind of implies that this is the case.
But if TRT is built without CUDA EP the dependencies should be limited to TRT and cudart correct ? Given that would not link against any other library. At least the ORT source should not require any other linking that that.

Page / URL

https://onnxruntime.ai/docs/execution-providers/TensorRT-ExecutionProvider.html#requirements

[RyanUnderhill]

The documentation says this: "Note that it is recommended you also register CUDAExecutionProvider to allow Onnx Runtime to assign nodes to CUDA execution provider that TensorRT does not support."

So it's not required, but recommended to avoid having the CPU provider as the fallback for anything that TensorRT doesn't support.

In terms of building, it appears to depend on cuda/cudnn/cublas:

https://github.com/microsoft/onnxruntime/blob/9c323106735535b6dab6b476648faac0ad185e21/cmake/onnxruntime_providers_tensorrt.cmake#L85C8-L85C8

[gedoensmax]

I see thanks. I think we should double check with every TRT release if cudnn and cutlass is really a strict dependency. TRT kind of has dropped the dependency on those in the latest releases. I will check internally how we stand towards that.

[github-actions]

This issue has been automatically marked as stale due to inactivity and will be closed in 7 days if no further activity occurs. If further support is needed, please provide an update and/or more details.

[gedoensmax]

This is related to #18542.

[github-actions]

This issue has been automatically marked as stale due to inactivity and will be closed in 30 days if no further activity occurs. If further support is needed, please provide an update and/or more details.