Could a onnx model use several gpu devices to accelerate?

[posa88]

`
optionsOfSession1.addCuda(0);
optionsOfSession1.addCuda(1);

optionsOfSession2.addCuda(3);
optionsOfSession2.addCuda(4);
`
Could the code works fine?

[snnn]

No. They need be in different processes.

Our CUDA EP has a threadlocal cache: https://github.com/microsoft/onnxruntime/blob/master/onnxruntime/core/providers/cuda/cuda_execution_provider.h#L113 . Each PerThreadContext maps to one single GPU.