Support pre-registered plug-in cuda execution provider library

[baijumeswani]

Up until now, for the cuda execution provider, onnxruntime-genai tried using the built-in cuda ep using the legacy OrtCUDAProviderOptionsV2 and AppendExecutionProvider_CUDA_V2 API. These functions are designed to make use of the built-in cuda execution provider and are not compatible with the world of plug-in eps.

The changes in this pull-request now extend support to executing the model with a pre-registered (plugged-in) CUDAExecutionProvider.

In order to use the plug-in capabilities, application layer needs to do the following:

• C++

  #include "onnxruntime_cxx_api.h"
  #include "ort_genai.h"
  
  auto env = Ort::Env();
  env.RegisterExecutionProviderLibrary("CUDAExecutionProvider", "path\to\onnxruntime_providers_cuda.dll");
  
  auto model = OgaModel("path\to\model\directory");
  ...

• Python

  import onnxruntime_genai as og
  
  og.register_execution_provider_library("CUDAExecutionProvider", "path\to\onnxruntime_providers_cuda.dll")
  
  model = og.Model("path\to\model\directory")
  ...

• C#

  using Microsoft.ML.OnnxRuntime;
  using Microsoft.ML.OnnxRuntimeGenAI;
  
  var ortEnv = OrtEnv.Instance();
  ortEnv.RegisterExecutionProviderLibrary("CUDAExecutionProvider", "path\to\onnxruntime_providers_cuda.dll");
  
  using Model model = new Model("path\to\model\directory");
  ...

[kunal-vaishnavi]

Can we provide a direct API from ORT GenAI for registering a provider library for non-Python language bindings?

[baijumeswani]

Can we provide a direct API from ORT GenAI for registering a provider library for non-Python language bindings?

We do have a C API:

onnxruntime-genai/src/ort_genai_c.h

Lines 1071 to 1084 in 240fa1b

	/**
	* \brief Registers an execution provider library with ONNXRuntime API.
	* \param registration_name name for registration.
	* \param path provider path.
	*
	*/
	OGA_EXPORT void OGA_API_CALL OgaRegisterExecutionProviderLibrary(const char* registration_name, const char* library_path);
	/**
	* \brief Unregisters an execution provider library with ONNXRuntime API.
	* \param registration_name name for registration.
	*
	*/
	OGA_EXPORT void OGA_API_CALL OgaUnregisterExecutionProviderLibrary(const char* registration_name);

But people are probably more used to using the onnxruntime env since those APIs offer more control. onnxruntime-genai API fr library registration is needed for Python in particular because the onnxruntime python library is not the one we load from onnxruntime-genai's Python package.