Olive

Why we included this project

Olive is Microsoft's toolkit for pushing models onto the ONNX Runtime, and its main draw is that it automates the search for the right optimizations. You point it at a model, name the hardware it will run on (CPU, GPU, or NPU) and the constraints that matter, such as accuracy or latency, and it works through combinations of its 40-plus built-in passes to hand back an efficient ONNX model. That covers conversion, finetuning, quantization, graph optimization, and packaging the result for deployment, so the pipeline from a PyTorch or Hugging Face checkpoint to something you can actually serve is handled in one place. The CLI covers common jobs like `olive quantize` and `olive auto-opt`, while YAML workflows give teams finer control and caching speeds up repeated runs. Recent releases added NPU support and Multi-LoRA serving, which matters if you are deploying to several accelerators and would rather skip the usual trial and error.