C, C++, and Python API for Adapters (Multi-LoRA and others)

[baijumeswani]

This pull-request introduces a way to apply adapters to a model with onnxruntime-genai. It leverages onnxruntime API changes from microsoft/onnxruntime#22046 and allows users to load, apply, and unload adapters as needed.

This pull-request adds the C, C++ and the Python API. Other language bindings will follow soon.

Example usage:

C++

auto model = OgaModel::Create("<adapter/model/path>");
auto adapters = OgaAdapters::Create(*model);
adapters->LoadAdapter("</absolute/adapter/file/path>", "<unique_user_supplied_adapter_identifier>");

auto tokenizer = OgaTokenizer::Create(*model);

const char* input_strings[] = {
    "This is a test.",
    "Rats are awesome pets!",
    "The quick brown fox jumps over the lazy dog.",
};

auto input_sequences = OgaSequences::Create();
for (auto& string : input_strings)
  tokenizer->Encode(string, *input_sequences);

{
  auto params = OgaGeneratorParams::Create(*model);
  params->SetSearchOption("max_length", 20);
  params->SetInputSequences(*input_sequences);

  auto generator = OgaGenerator::Create(*model, *params);
  generator->SetActiveAdapter(*adapters, "<unique_user_supplied_adapter_identifier>");

  while (!generator->IsDone()) {
    generator->ComputeLogits();
    generator->GenerateNextToken();
  }
}

// Optionally, unload the adapter. Will error out if the adapter is still active.
// So, the generator must go out of scope before the adapter can be unloaded.
adapters->UnloadAdapter("<unique_user_supplied_adapter_identifier>");

Python

model = og.Model("<model/path>")
adapters = og.Adapters(model)
adapters.load("</absolute/path/to/adapters>", "<unique_user_supplied_adapter_identifier>")

tokenizer = og.Tokenizer(model)
prompts = [
    "This is a test.",
    "Rats are awesome pets!",
    "The quick brown fox jumps over the lazy dog.",
]

params = og.GeneratorParams(model)
params.set_search_options(max_length=20)
params.input_ids = tokenizer.encode_batch(prompts)

generator = og.Generator(model, params)
generator.set_active_adapter(adapters, "<unique_user_suppled_adapter_identifier>")

while not generator.is_done():
    generator.compute_logits()
    generator.generate_next_token()

[pranavsharma]

In the interest of time, let's resolve the ref counting issues later as they don't affect the public APIs.