Are concurrent calls to the 'Run' method in the C interface serialized when using the CUDA provider?

[xgirones]

When employing the CPU provider, it is possible to run multiple inferences in parallel in the same session by calling the Run method concurrently from different threads. However, when attempting the same with the CUDA provider, the performance is very bad, and it appears that the ORT is serializing simultaneous calls to Run. Is this the case? If so, what can I do to run multiple concurrent queries efficiently in a single session when using the CUDA backend? My workflow involves running multiple inferences in parallel using tiny batch sizes and small models.