craig-martinson/onnxruntime-genai

Generative AI extensions for onnxruntime

ONNX Runtime Generative AI

Run generative AI models with ONNX Runtime.

This library provides the generative AI loop for ONNX models, including inference with ONNX Runtime, logits processing, search and sampling, and KV cache management.

Users can call a high level generate() method, or run each iteration of the model in a loop.

• Support greedy/beam search and TopP, TopK sampling to generate token sequences
• Built in logits processing like repetition penalties
• Easy custom scoring

See full documentation at [https://onnxruntime.ai/docs/genai].

Features

• Supported model architectures:
  • Phi-3
  • Phi-2
  • Gemma
  • LLaMA
  • Mistral
• Supported targets:
  • GPU (DirectML)
  • GPU (CUDA)
  • CPU
• Supported sampling features
  • Beam search
  • Greedy search
  • Top P/Top K
• APIs
  • Python
  • C#
  • C/C++

Coming very soon

• Support for the encoder decoder model architectures, such as whisper, T5 and BART.

Coming soon

• Support for mobile devices (Android and iOS) with Java and Objective-C bindings

Roadmap

• Stable diffusion pipeline
• Automatic model download and cache
• More model architectures

Installation

DirectML

pip install [--pre] numpy onnxruntime-genai-directml

CPU

pip install [--pre] numpy onnxruntime-genai

CUDA

pip install numpy onnxruntime-genai-cuda --pre --index-url=https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/onnxruntime-genai/pypi/simple/

Sample code for phi-2 in Python

Install the onnxruntime-genai Python package.

1. Build the model

python -m onnxruntime_genai.models.builder -m microsoft/phi-2 -e cpu -p int4 -o ./models/phi2
# You can append --extra_options enable_cuda_graph=1 to build an onnx model that supports using cuda graph in ORT.

2. Run inference

import os
import onnxruntime_genai as og

model_path = os.path.abspath("./models/phi2")

model = og.Model(model_path)

tokenizer = og.Tokenizer(model)

prompt = '''def print_prime(n):
    """
    Print all primes between 1 and n
    """'''

tokens = tokenizer.encode(prompt)

params = og.GeneratorParams(model)
params.set_search_options({"max_length":200})
# Add the following line to enable cuda graph by passing the maximum batch size.
# params.try_use_cuda_graph_with_max_batch_size(16)
params.input_ids = tokens

output_tokens = model.generate(params)

text = tokenizer.decode(output_tokens)

print("Output:")
print(text)

Model download and export

ONNX models are run from a local folder, via a string supplied to the Model() method.

You can bring your own ONNX model or use the model builder utility, included in this package.

Install model builder dependencies.

pip install numpy install transformers torch onnx onnxruntime

Export int4 CPU version

huggingface-cli login --token <your HuggingFace token>
python -m onnxruntime_genai.models.builder -m microsoft/phi-2 -p int4 -e cpu -o <model folder>

Contributing

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.

When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact [email protected] with any additional questions or comments.

Trademarks

This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.