At its core, 🤗 Optimum uses configuration objects to define parameters for optimization on different accelerators. These objects are then used to instantiate dedicated optimizers, quantizers, and pruners. For example, here’s how you can apply dynamic quantization with ONNX Runtime:

>>> from optimum.onnxruntime import ORTConfig, ORTQuantizer

>>> # The model we wish to quantize
>>> model_ckpt = "distilbert-base-uncased-finetuned-sst-2-english"
>>> # The type of quantization to apply
>>> ort_config = ORTConfig(quantization_approach="dynamic")
>>> quantizer = ORTQuantizer(ort_config)
>>> # Quantize the model!
>>> quantizer.fit(model_ckpt, output_dir=".", feature="sequence-classification")

In this example, we’ve quantized a model from the Hugging Face Hub, but it could also be a path to a local model directory. The feature argument in the fit() method corresponds to the type of task that we wish to quantize the model for. The result from applying the fit() method is a model-quantized.onnx file that can be used to run inference. Here’s an example of how to load an ONNX Runtime model and generate predictions with it:

>>> from datasets import Dataset
>>> from transformers import AutoTokenizer
>>> from optimum.onnxruntime import ORTModel

>>> # Load quantized model
>>> ort_model = ORTModel("model-quantized.onnx", quantizer.onnx_config)
>>> # Create a dataset or load one from the Hub
>>> ds = Dataset.from_dict({"sentence": ["I love burritos!"]})
>>> # Tokenize the inputs & convert to PyTorch tensors
>>> tokenizer = AutoTokenizer.from_pretrained(model_ckpt)

>>> def preprocess_fn(ex):
...     return tokenizer(ex["sentence"])

>>> tokenized_ds = ds.map(preprocess_fn, remove_columns=ds.column_names)
>>> tokenized_ds.set_format("torch")
>>> # Create dataloader and run evaluation
>>> dataloader = DataLoader(tokenized_ds)
>>> ort_outputs = ort_model.evaluation_loop(dataloader)
>>> # Extract logits!
>>> ort_outputs.predictions

Similarly, you can apply static quantization by simply changing the quantization_approach in the ORTConfig object:

>>> ort_config = ORTConfig(quantization_approach="static")

Static quantization relies on feeding batches of data through the model to observe the activation patterns ahead of inference time. The ideal quantization scheme is then calculated and saved. To support this, 🤗 Optimum allows you to provide a calibration dataset. The calibration dataset can be a simple Dataset object from the 🤗 Datasets library, or any dataset that’s hosted on the Hugging Face Hub. For this example, we’ll pick the sst2 dataset that the model was originally trained on:

>>> from transformers import DataCollatorWithPadding

>>> # We use a data collator to pad the examples in a batch
>>> data_collator = DataCollatorWithPadding(tokenizer)
>>> # For calibration we define the dataset and preprocessing function
>>> quantizer = ORTQuantizer(
...     ort_config,
...     dataset_name="glue",
...     dataset_config_name="sst2",
...     preprocess_function=preprocess_fn,
...     data_collator=data_collator,
... )
>>> # Quantize the same way we did for dynamic quantization!
>>> quantizer.fit(model_ckpt, output_dir=".", feature="sequence-classification")

As a final example, let’s take a look at applying graph optimizations techniques such as operator fusion and constant folding. As before, we load a configuration object, but this time by setting the optimization level instead of the quantization approach:

>>> # opt_level=99 enables all graph optimisations
>>> ort_config = ORTConfig(opt_level=99)

>>> from optimum.onnxruntime import ORTOptimizer

>>> optimizer = ORTOptimizer(ort_config)
>>> optimizer.fit(model_ckpt, output_dir=".", feature="sequence-classification")

And that’s it - the model is now optimized and ready for inference! As you can see, the process is similar in each case:

Quickstart