Inference pipelines with the ONNX Runtime accelerator

The pipeline() function makes it simple to use models from the Model Hub for accelerated inference on a variety of tasks such as text classification, question answering and image classification.

You can also use the pipeline() function from Transformers and provide your Optimum model class.

Currently the supported tasks are:

feature-extraction
text-classification
token-classification
question-answering
zero-shot-classification
text-generation
text2text-generation
summarization
translation
image-classification
automatic-speech-recognition
image-to-text

Optimum pipeline usage

While each task has an associated pipeline class, it is simpler to use the general pipeline() function which wraps all the task-specific pipelines in one object. The pipeline() function automatically loads a default model and tokenizer/feature-extractor capable of performing inference for your task.

Start by creating a pipeline by specifying an inference task:

>>> from optimum.pipelines import pipeline

>>> classifier = pipeline(task="text-classification", accelerator="ort")

Pass your input text/image to the pipeline() function:

>>> classifier("I like you. I love you.")
[{'label': 'POSITIVE', 'score': 0.9998838901519775}]

Note: The default models used in the pipeline() function are not optimized for inference or quantized, so there won’t be a performance improvement compared to their PyTorch counterparts.

Using vanilla Transformers model and converting to ONNX

The pipeline() function accepts any supported model from the Hugging Face Hub. There are tags on the Model Hub that allow you to filter for a model you’d like to use for your task.

To be able to load the model with the ONNX Runtime backend, the export to ONNX needs to be supported for the considered architecture.

You can check the list of supported architectures here.

Once you have picked an appropriate model, you can create the pipeline() by specifying the model repo:

>>> from optimum.pipelines import pipeline

# The model will be loaded to an ORTModelForQuestionAnswering.
>>> onnx_qa = pipeline("question-answering", model="deepset/roberta-base-squad2", accelerator="ort")
>>> question = "What's my name?"
>>> context = "My name is Philipp and I live in Nuremberg."

>>> pred = onnx_qa(question=question, context=context)

It is also possible to load it with the from_pretrained(model_name_or_path, export=True) method associated with the ORTModelForXXX class.

For example, here is how you can load the ORTModelForQuestionAnswering class for question answering:

>>> from transformers import AutoTokenizer
>>> from optimum.onnxruntime import ORTModelForQuestionAnswering
>>> from optimum.pipelines import pipeline

>>> tokenizer = AutoTokenizer.from_pretrained("deepset/roberta-base-squad2")

>>> # Loading the PyTorch checkpoint and converting to the ONNX format by providing
>>> # export=True
>>> model = ORTModelForQuestionAnswering.from_pretrained(
...     "deepset/roberta-base-squad2",
...     export=True
... )

>>> onnx_qa = pipeline("question-answering", model=model, tokenizer=tokenizer, accelerator="ort")
>>> question = "What's my name?"
>>> context = "My name is Philipp and I live in Nuremberg."

>>> pred = onnx_qa(question=question, context=context)

Using Optimum models

The pipeline() function is tightly integrated with the Hugging Face Hub and can load ONNX models directly.

>>> from optimum.pipelines import pipeline

>>> onnx_qa = pipeline("question-answering", model="optimum/roberta-base-squad2", accelerator="ort")
>>> question = "What's my name?"
>>> context = "My name is Philipp and I live in Nuremberg."

>>> pred = onnx_qa(question=question, context=context)

It is also possible to load it with the from_pretrained(model_name_or_path) method associated with the ORTModelForXXX class.

For example, here is how you can load the ORTModelForQuestionAnswering class for question answering:

>>> from transformers import AutoTokenizer
>>> from optimum.onnxruntime import ORTModelForQuestionAnswering
>>> from optimum.pipelines import pipeline

>>> tokenizer = AutoTokenizer.from_pretrained("optimum/roberta-base-squad2")

>>> # Loading directly an ONNX model from a model repo.
>>> model = ORTModelForQuestionAnswering.from_pretrained("optimum/roberta-base-squad2")

>>> onnx_qa = pipeline("question-answering", model=model, tokenizer=tokenizer, accelerator="ort")
>>> question = "What's my name?"
>>> context = "My name is Philipp and I live in Nuremberg."

>>> pred = onnx_qa(question=question, context=context)

Optimizing and quantizing in pipelines

The pipeline() function can not only run inference on vanilla ONNX Runtime checkpoints - you can also use checkpoints optimized with the ORTQuantizer and the ORTOptimizer.

Below you can find two examples of how you could use the ORTOptimizer and the ORTQuantizer to optimize/quantize your model and use it for inference afterwards.

Quantizing with the ORTQuantizer

>>> from transformers import AutoTokenizer
>>> from optimum.onnxruntime import (
...     AutoQuantizationConfig,
...     ORTModelForSequenceClassification,
...     ORTQuantizer
... )
>>> from optimum.pipelines import pipeline

>>> # Load the tokenizer and export the model to the ONNX format
>>> model_id = "distilbert-base-uncased-finetuned-sst-2-english"
>>> save_dir = "distilbert_quantized"

>>> model = ORTModelForSequenceClassification.from_pretrained(model_id, export=True)

>>> # Load the quantization configuration detailing the quantization we wish to apply
>>> qconfig = AutoQuantizationConfig.avx512_vnni(is_static=False, per_channel=True)
>>> quantizer = ORTQuantizer.from_pretrained(model)

>>> # Apply dynamic quantization and save the resulting model
>>> quantizer.quantize(save_dir=save_dir, quantization_config=qconfig)
>>> # Load the quantized model from a local repository
>>> model = ORTModelForSequenceClassification.from_pretrained(save_dir)

>>> # Create the transformers pipeline
>>> onnx_clx = pipeline("text-classification", model=model, accelerator="ort")
>>> text = "I like the new ORT pipeline"
>>> pred = onnx_clx(text)
>>> print(pred)
>>> # [{'label': 'POSITIVE', 'score': 0.9974810481071472}]

>>> # Save and push the model to the hub (in practice save_dir could be used here instead)
>>> model.save_pretrained("new_path_for_directory")
>>> model.push_to_hub("new_path_for_directory", repository_id="my-onnx-repo", use_auth_token=True)

Optimizing with ORTOptimizer

>>> from transformers import AutoTokenizer
>>> from optimum.onnxruntime import (
...     AutoOptimizationConfig,
...     ORTModelForSequenceClassification,
...     ORTOptimizer
... )
>>> from optimum.onnxruntime.configuration import OptimizationConfig
>>> from optimum.pipelines import pipeline

>>> # Load the tokenizer and export the model to the ONNX format
>>> model_id = "distilbert-base-uncased-finetuned-sst-2-english"
>>> save_dir = "distilbert_optimized"

>>> tokenizer = AutoTokenizer.from_pretrained(model_id)
>>> model = ORTModelForSequenceClassification.from_pretrained(model_id, export=True)

>>> # Load the optimization configuration detailing the optimization we wish to apply
>>> optimization_config = AutoOptimizationConfig.O3()
>>> optimizer = ORTOptimizer.from_pretrained(model)

>>> optimizer.optimize(save_dir=save_dir, optimization_config=optimization_config)
# Load the optimized model from a local repository
>>> model = ORTModelForSequenceClassification.from_pretrained(save_dir)

# Create the transformers pipeline
>>> onnx_clx = pipeline("text-classification", model=model, accelerator="ort")
>>> text = "I like the new ORT pipeline"
>>> pred = onnx_clx(text)
>>> print(pred)
>>> # [{'label': 'POSITIVE', 'score': 0.9973127245903015}]

# Save and push the model to the hub
>>> tokenizer.save_pretrained("new_path_for_directory")
>>> model.save_pretrained("new_path_for_directory")
>>> model.push_to_hub("new_path_for_directory", repository_id="my-onnx-repo", use_auth_token=True)