ONNX Runtime GenAI C# API

[baijumeswani]

This pull-request introduces the C# API for ONNX Runtime GenAI. The C# API follows the C API closely and leverages the C API as the interface with the ONNX Runtime GenAI core.

The API 📚

The C# API primarily revolves around three classes:

• Model - Entry point for running GenAI with the library. Used for loading the model and the config.
• GeneratorParams - Used to set the configurable parameters for search.
• Generator - Main computation engine. Used for computing the logits and retrieving the generated tokens.

Sample usage 👨‍💻

using Model model = new Model(modelPath, DeviceType.CPU)
using GeneratorParams generatorParams = new GeneratorParams(model)

generatorParams.SetMaxLength(maxLength);
generatorParams.SetInputIDs(inputIDs, sequenceLength, batchSize);

using Generator generator = new Generator(model, generatorParams)

while (!generator.IsDone())
{
    generator.ComputeLogits();
    generator.GenerateNextTokenTop();
}

for (ulong i = 0; i < batchSize; i++)
{
    var sequence = generator.GetSequence((int)i);
    ... // use the output tokens
}

What does this Pull-Request Include? 🤔

• The C# API for Windows. The build generates a managed dll called Microsoft.ML.OnnxRuntimeGenAI.dll for x64 using net8.0.
• Test that leverages the managed dll built.
• This pull-request does not include packaging.
• Module definition for exporting the native functions in the C library.

What's next? 🔄

• Run build and tests in CI pipeline.
• C# packaging.
• Linux C# support
• Build the C# API as part of the cmake build.
• Tokenizer functions to be added to C# API to complete the API

[yuslepukhin]

The summary example features nested using blocks. It would be better to use using statements that are scope based, so blocks are not required, although they may be useful when you want to dispose of something before the current scope ends. #Resolved

[baijumeswani]

The summary example features nested using blocks. It would be better to use using statements that are scope based, so blocks are not required, although they may be useful when you want to dispose of something before the current scope ends.

Updated the sample example to avoid using the nested using statements.

[yuslepukhin]

[baijumeswani]

Thank you for the review @yuslepukhin @RyanUnderhill @snnn :)