๐ค Optimum is an extension of ๐ค Transformers, providing a set of optimization tools enabling maximum efficiency to train and run models on targeted hardware.
The AI ecosystem evolves quickly and more and more specialized hardware along with their own optimizations are emerging every day. As such, Optimum enables users to efficiently use any of these platforms with the same ease inherent to transformers.
Optimum aims at providing more diversity towards the kind of hardware users can target to train and finetune their models.
To achieve this, we are collaborating with the following hardware manufacturers in order to provide the best transformers integration:
- Graphcore IPUs - IPUs are a completely new kind of massively parallel processor to accelerate machine intelligence. More information here.
- Habana Gaudi Processor (HPU) - HPUs are designed to maximize training throughput and efficiency. More information here.
- Intel - Enabling the usage of Intel tools to accelerate inference on Intel architectures. More information about Neural Compressor and OpenVINO.
- More to come soon! โญ
๐ค Optimum can be installed using pip
as follows:
python -m pip install optimum
If you'd like to use the accelerator-specific features of ๐ค Optimum, you can install the required dependencies according to the table below:
Accelerator | Installation |
---|---|
ONNX Runtime | python -m pip install optimum[onnxruntime] |
Intel Neural Compressor | python -m pip install optimum[neural-compressor] |
OpenVINO | python -m pip install optimum[openvino,nncf] |
Graphcore IPU | python -m pip install optimum[graphcore] |
Habana Gaudi Processor (HPU) | python -m pip install optimum[habana] |
If you'd like to play with the examples or need the bleeding edge of the code and can't wait for a new release, you can install the base library from source as follows:
python -m pip install git+https://github.com/huggingface/optimum.git
For the accelerator-specific features, you can install them by appending #egg=optimum[accelerator_type]
to the pip
command, e.g.
python -m pip install git+https://github.com/huggingface/optimum.git#egg=optimum[onnxruntime]
Along with supporting dedicated AI hardware for training, Optimum also provides inference optimizations towards various frameworks and platforms.
Optimum enables the usage of popular compression techniques such as quantization and pruning by supporting ONNX Runtime along with Intel Neural Compressor and OpenVINO NNCF.
Features | ONNX Runtime | Neural Compressor | OpenVINO |
---|---|---|---|
Post-training Dynamic Quantization | โ๏ธ | โ๏ธ | N/A |
Post-training Static Quantization | โ๏ธ | โ๏ธ | โ๏ธ |
Quantization Aware Training (QAT) | Stay tuned! โญ | โ๏ธ | โ๏ธ |
Pruning | N/A | โ๏ธ | Stay tuned! โญ |
Check out the examples below to see how ๐ค Optimum can be used to train and run inference on various hardware accelerators.
To accelerate inference with ONNX Runtime, ๐ค Optimum uses configuration objects to define parameters for graph optimization and quantization. These objects are then used to instantiate dedicated optimizers and quantizers.
Before applying quantization or optimization, first we need to load our model. To load a model and run inference with ONNX Runtime, you can just replace the canonical Transformers AutoModelForXxx
class with the corresponding ORTModelForXxx
class. If you want to load from a PyTorch checkpoint, set from_transformers=True
to export your model to the ONNX format.
from optimum.onnxruntime import ORTModelForSequenceClassification
from transformers import AutoTokenizer
model_checkpoint = "distilbert-base-uncased-finetuned-sst-2-english"
save_directory = "tmp/onnx/"
# Load a model from transformers and export it to ONNX
tokenizer = AutoTokenizer.from_pretrained(model_checkpoint)
ort_model = ORTModelForSequenceClassification.from_pretrained(model_checkpoint, from_transformers=True)
# Save the ONNX model and tokenizer
ort_model.save_pretrained(save_directory)
tokenizer.save_pretrained(save_directory)
Let's see now how we can apply dynamic quantization with ONNX Runtime:
from optimum.onnxruntime.configuration import AutoQuantizationConfig
from optimum.onnxruntime import ORTQuantizer
# Define the quantization methodology
qconfig = AutoQuantizationConfig.arm64(is_static=False, per_channel=False)
quantizer = ORTQuantizer.from_pretrained(ort_model)
# Apply dynamic quantization on the model
quantizer.quantize(save_dir=save_directory, quantization_config=qconfig)
In this example, we've quantized a model from the Hugging Face Hub, in the same manner we can quantize a model hosted locally by providing the path to the directory containing the model weights. The result from applying the quantize()
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 optimum.onnxruntime import ORTModelForSequenceClassification
from transformers import pipeline, AutoTokenizer
model = ORTModelForSequenceClassification.from_pretrained(save_directory, file_name="model_quantized.onnx")
tokenizer = AutoTokenizer.from_pretrained(save_directory)
classifier = pipeline("text-classification", model=model, tokenizer=tokenizer)
results = classifier("I love burritos!")
You can find more examples in the documentation and in the examples.
To load a model and run inference with OpenVINO Runtime, you can just replace your AutoModelForXxx
class with the corresponding OVModelForXxx
class.
If you want to load a PyTorch checkpoint, set from_transformers=True
to convert your model to the OpenVINO IR (Intermediate Representation).
- from transformers import AutoModelForSequenceClassification
+ from optimum.intel.openvino import OVModelForSequenceClassification
from transformers import AutoTokenizer, pipeline
# Download a tokenizer and model from the Hub and convert to OpenVINO format
tokenizer = AutoTokenizer.from_pretrained(model_id)
model_id = "distilbert-base-uncased-finetuned-sst-2-english"
- model = AutoModelForSequenceClassification.from_pretrained(model_id)
+ model = OVModelForSequenceClassification.from_pretrained(model_id, from_transformers=True)
# Run inference!
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.
To train transformers on Habana's Gaudi processors, ๐ค Optimum provides a GaudiTrainer
that is very similar to the ๐ค Transformers Trainer. Here is a simple example:
- 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.
To train transformers on Graphcore's IPUs, ๐ค Optimum provides a IPUTrainer
that is very similar to the ๐ค Transformers Trainer. Here is a simple example:
- from transformers import Trainer, TrainingArguments
+ from optimum.graphcore import IPUConfig, IPUTrainer, IPUTrainingArguments
# Download a pretrained model from the Hub
model = AutoModelForXxx.from_pretrained("bert-base-uncased")
# Define the training arguments
- training_args = TrainingArguments(
+ training_args = IPUTrainingArguments(
output_dir="path/to/save/folder/",
+ ipu_config_name="Graphcore/bert-base-ipu", # Any IPUConfig on the Hub or stored locally
...
)
# Define the configuration to compile and put the model on the IPU
+ ipu_config = IPUConfig.from_pretrained(training_args.ipu_config_name)
# Initialize the trainer
- trainer = Trainer(
+ trainer = IPUTrainer(
model=model,
+ ipu_config=ipu_config
args=training_args,
train_dataset=train_dataset
...
)
# Use Graphcore IPU for training!
trainer.train()
You can find more examples in the documentation and in the examples.
To train transformers with ONNX Runtime's acceleration features, ๐ค Optimum provides a ORTTrainer
that is very similar to the ๐ค Transformers Trainer. Here is a simple example:
- 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,
+ feature="sequence-classification", # The model type to export to ONNX
...
)
# Use ONNX Runtime for training!
trainer.train()
You can find more examples in the documentation and in the examples.