Hugging Face Optimum

🤗 Optimum is an extension of 🤗 Transformers and Diffusers, providing a set of optimization tools enabling maximum efficiency to train and run models on targeted hardware, while keeping things easy to use.

Installation

🤗 Optimum can be installed using pip as follows:

python -m pip install optimum

If you'd like to use the accelerator-specific features of 🤗 Optimum, you can install the required dependencies according to the table below:

The --upgrade --upgrade-strategy eager option is needed to ensure the different packages are upgraded to the latest possible version.

To install from source:

python -m pip install git+https://github.com/huggingface/optimum.git

For the accelerator-specific features, append optimum[accelerator_type] to the above command:

python -m pip install optimum[onnxruntime]@git+https://github.com/huggingface/optimum.git

Accelerated Inference

🤗 Optimum provides multiple tools to export and run optimized models on various ecosystems:

• ONNX / ONNX Runtime
• TensorFlow Lite
• OpenVINO
• Habana first-gen Gaudi / Gaudi2, more details here
• AWS Inferentia 2 / Inferentia 1, more details here
• NVIDIA TensorRT-LLM , more details here

The export and optimizations can be done both programmatically and with a command line.

Features summary

OpenVINO

Before you begin, make sure you have all the necessary libraries installed :

pip install --upgrade --upgrade-strategy eager optimum[openvino]

It is possible to export 🤗 Transformers and Diffusers models to the OpenVINO format easily:

optimum-cli export openvino --model distilbert-base-uncased-finetuned-sst-2-english distilbert_sst2_ov

If you add --weight-format int8, the weights will be quantized to int8, check out our documentation for more detail. To apply quantization on both weights and activations, you can find more information here.

To load a model and run inference with OpenVINO Runtime, you can just replace your AutoModelForXxx class with the corresponding OVModelForXxx class. To load a PyTorch checkpoint and convert it to the OpenVINO format on-the-fly, you can set export=True when loading your model.

- from transformers import AutoModelForSequenceClassification
+ from optimum.intel import OVModelForSequenceClassification
  from transformers import AutoTokenizer, pipeline

  model_id = "distilbert-base-uncased-finetuned-sst-2-english"
  tokenizer = AutoTokenizer.from_pretrained(model_id)
- model = AutoModelForSequenceClassification.from_pretrained(model_id)
+ model = OVModelForSequenceClassification.from_pretrained(model_id, export=True)

  classifier = pipeline("text-classification", model=model, tokenizer=tokenizer)
  results = classifier("He's a dreadful magician.")

You can find more examples in the documentation and in the examples.

Neural Compressor

Before you begin, make sure you have all the necessary libraries installed :

pip install --upgrade --upgrade-strategy eager optimum[neural-compressor]

Dynamic quantization can be applied on your model:

optimum-cli inc quantize --model distilbert-base-cased-distilled-squad --output ./quantized_distilbert

To load a model quantized with Intel Neural Compressor, hosted locally or on the 🤗 hub, you can do as follows :

from optimum.intel import INCModelForSequenceClassification

model_id = "Intel/distilbert-base-uncased-finetuned-sst-2-english-int8-dynamic"
model = INCModelForSequenceClassification.from_pretrained(model_id)

You can find more examples in the documentation and in the examples.

ONNX + ONNX Runtime

Before you begin, make sure you have all the necessary libraries installed :

pip install optimum[exporters,onnxruntime]

It is possible to export 🤗 Transformers and Diffusers models to the ONNX format and perform graph optimization as well as quantization easily:

optimum-cli export onnx -m deepset/roberta-base-squad2 --optimize O2 roberta_base_qa_onnx

The model can then be quantized using onnxruntime:

optimum-cli onnxruntime quantize \
  --avx512 \
  --onnx_model roberta_base_qa_onnx \
  -o quantized_roberta_base_qa_onnx

These commands will export deepset/roberta-base-squad2 and perform O2 graph optimization on the exported model, and finally quantize it with the avx512 configuration.

For more information on the ONNX export, please check the documentation.

Run the exported model using ONNX Runtime

Once the model is exported to the ONNX format, we provide Python classes enabling you to run the exported ONNX model in a seemless manner using ONNX Runtime in the backend:

- from transformers import AutoModelForQuestionAnswering
+ from optimum.onnxruntime import ORTModelForQuestionAnswering
  from transformers import AutoTokenizer, pipeline

  model_id = "deepset/roberta-base-squad2"
  tokenizer = AutoTokenizer.from_pretrained(model_id)
- model = AutoModelForQuestionAnswering.from_pretrained(model_id)
+ model = ORTModelForQuestionAnswering.from_pretrained("roberta_base_qa_onnx")
  qa_pipe = pipeline("question-answering", model=model, tokenizer=tokenizer)
  question = "What's Optimum?"
  context = "Optimum is an awesome library everyone should use!"
  results = qa_pipe(question=question, context=context)

More details on how to run ONNX models with ORTModelForXXX classes here.

TensorFlow Lite

Before you begin, make sure you have all the necessary libraries installed :

pip install optimum[exporters-tf]

Just as for ONNX, it is possible to export models to TensorFlow Lite and quantize them:

optimum-cli export tflite \
  -m deepset/roberta-base-squad2 \
  --sequence_length 384  \
  --quantize int8-dynamic roberta_tflite_model

Accelerated training

🤗 Optimum provides wrappers around the original 🤗 Transformers Trainer to enable training on powerful hardware easily. We support many providers:

• Habana's Gaudi processors
• AWS Trainium instances, check here
• ONNX Runtime (optimized for GPUs)

Habana

Before you begin, make sure you have all the necessary libraries installed :

pip install --upgrade --upgrade-strategy eager optimum[habana]

- from transformers import Trainer, TrainingArguments
+ from optimum.habana import GaudiTrainer, GaudiTrainingArguments

  # Download a pretrained model from the Hub
  model = AutoModelForXxx.from_pretrained("bert-base-uncased")

  # Define the training arguments
- training_args = TrainingArguments(
+ training_args = GaudiTrainingArguments(
      output_dir="path/to/save/folder/",
+     use_habana=True,
+     use_lazy_mode=True,
+     gaudi_config_name="Habana/bert-base-uncased",
      ...
  )

  # Initialize the trainer
- trainer = Trainer(
+ trainer = GaudiTrainer(
      model=model,
      args=training_args,
      train_dataset=train_dataset,
      ...
  )

  # Use Habana Gaudi processor for training!
  trainer.train()

You can find more examples in the documentation and in the examples.

ONNX Runtime

- from transformers import Trainer, TrainingArguments
+ from optimum.onnxruntime import ORTTrainer, ORTTrainingArguments

  # Download a pretrained model from the Hub
  model = AutoModelForSequenceClassification.from_pretrained("bert-base-uncased")

  # Define the training arguments
- training_args = TrainingArguments(
+ training_args = ORTTrainingArguments(
      output_dir="path/to/save/folder/",
      optim="adamw_ort_fused",
      ...
  )

  # Create a ONNX Runtime Trainer
- trainer = Trainer(
+ trainer = ORTTrainer(
      model=model,
      args=training_args,
      train_dataset=train_dataset,
      ...
  )

  # Use ONNX Runtime for training!
  trainer.train()

You can find more examples in the documentation and in the examples.

Quanto

Quanto is a pytorch quantization backend.

You can quantize a model either using the python API or the optimum-cli.

from transformers import AutoModelForCausalLM
from optimum.quanto import QuantizedModelForCausalLM, qint4

model = AutoModelForCausalLM.from_pretrained('meta-llama/Meta-Llama-3.1-8B')
qmodel = QuantizedModelForCausalLM.quantize(model, weights=qint4, exclude='lm_head')

The quantized model can be saved using save_pretrained:

qmodel.save_pretrained('./Llama-3.1-8B-quantized')

It can later be reloaded using from_pretrained:

from optimum.quanto import QuantizedModelForCausalLM

qmodel = QuantizedModelForCausalLM.from_pretrained('Llama-3.1-8B-quantized')

You can see more details and examples in the Quanto repository.