Intel® Extension for Transformers: Accelerating Transformer-based Models on Intel Platforms

Intel® Extension for Transformers is an innovative toolkit to accelerate Transformer-based models on Intel platforms. The toolkit helps developers to improve the productivity through ease-of-use model compression APIs by extending Hugging Face transformers APIs. The compression infrastructure leverages Intel® Neural Compressor which provides a rich set of model compression techniques: quantization, pruning, distillation and so on. The toolkit provides Transformers-accelerated Libraries and Neural Engine to demonstrate the performance of extremely compressed models, and therefore significantly improve the inference efficiency on Intel platforms. Some of the key features have been published in NeurIPS 2021 and 2022.

What does Intel® Extension for Transformers offer?

This toolkit helps developers to improve the productivity of inference deployment by extending Hugging Face transformers APIs for Transformer-based models in natural language processing (NLP) domain. With extremely compressed models, the toolkit can greatly improve the inference efficiency on Intel platforms.

Model Compression

Framework	Quantization	Pruning/Sparsity	Distillation	Neural Architecture Search
PyTorch	✔	✔	✔	✔
TensorFlow	✔	✔	✔	Stay tuned ⭐

Data Augmentation for NLP Datasets
Transformers-accelerated Neural Engine
Transformers-accelerated Libraries
Domain Algorithms |Length Adaptive Transformer | |-| |PyTorch ✔ |
Architecture of Intel® Extension for Transformers

Documentation

OVERVIEW
Model Compression	Neural Engine	Kernel Libraries	Examples
MODEL COMPRESSION
Quantization	Pruning	Distillation	Orchestration
Neural Architecture Search	Export	Metrics/Objectives	Pipeline
NEURAL ENGINE
Model Compilation	Custom Pattern	Deployment	Profiling
KERNEL LIBRARIES
Sparse GEMM Kernels	Custom INT8 Kernels	Profiling	Benchmark
ALGORITHMS
Length Adaptive		Data Augmentation
TUTORIALS AND RESULTS
Tutorials	Supported Models	Model Performance	Kernel Performance

Installation

Release Binary Install

pip install intel-extension-for-transformers

Install From Source

Install Intel® Extension for Transformers

git clone https://github.com/intel/intel-extension-for-transformers.git intel_extension_for_transformers
cd intel_extension_for_transformers
# Install Dependency
pip install -r requirements.txt
git submodule update --init --recursive
# Install intel_extension_for_transformers
python setup.py install

Note: Recommend install protobuf <= 3.20.0 if use onnxruntime <= 1.11

Getting Started

Quantization

from intel_extension_for_transformers.optimization import QuantizationConfig, metric, objectives
from intel_extension_for_transformers.optimization.trainer import NLPTrainer

# Replace transformers.Trainer with NLPTrainer
# trainer = transformers.Trainer(...)
trainer = NLPTrainer(...)
metric = metrics.Metric(name="eval_f1", is_relative=True, criterion=0.01)
q_config = QuantizationConfig(
    approach="PostTrainingStatic",
    metrics=[metric],
    objectives=[objectives.performance]
)
model = trainer.quantize(quant_config=q_config)

Please refer to quantization document for more details.

Pruning

from intel_extension_for_transformers.optimization import PrunerConfig, PruningConfig
from intel_extension_for_transformers.optimization.trainer import NLPTrainer

# Replace transformers.Trainer with NLPTrainer
# trainer = transformers.Trainer(...)
trainer = NLPTrainer(...)
metric = metrics.Metric(name="eval_accuracy")
pruner_config = PrunerConfig(prune_type='BasicMagnitude', target_sparsity_ratio=0.9)
p_conf = PruningConfig(pruner_config=[pruner_config], metrics=metric)
model = trainer.prune(pruning_config=p_conf)

Please refer to pruning document for more details.

Distillation

from intel_extension_for_transformers.optimization import DistillationConfig, Criterion
from intel_extension_for_transformers.optimization.trainer import NLPTrainer

# Replace transformers.Trainer with NLPTrainer
# trainer = transformers.Trainer(...)
teacher_model = ... # exist model
trainer = NLPTrainer(...)
metric = metrics.Metric(name="eval_accuracy")
d_conf = DistillationConfig(metrics=metric)
model = trainer.distill(distillation_config=d_conf, teacher_model=teacher_model)

Please refer to distillation document for more details.

Data Augmentation

Data augmentation provides the facilities to generate synthesized NLP dataset for further model optimization. The data augmentation supports text generation on popular fine-tuned models like GPT, GPT2, and other text synthesis approaches from nlpaug.

from intel_extension_for_transformers.preprocessing.data_augmentation import DataAugmentation
aug = DataAugmentation(augmenter_type="TextGenerationAug")
aug.input_dataset = "original_dataset.csv" # example: https://huggingface.co/datasets/glue/viewer/sst2/train
aug.column_names = "sentence"
aug.output_path = os.path.join(self.result_path, "test2.cvs")
aug.augmenter_arguments = {'model_name_or_path': 'gpt2-medium'}
aug.data_augment()
raw_datasets = load_dataset("csv", data_files=aug.output_path, delimiter="\t", split="train")

Please refer to data augmentation document for more details.

Neural Engine

Neural Engine is one of reference deployments that Intel Extension for Transformers provides. Neural Engine aims to demonstrate the optimal performance of extremely compressed NLP models by exploring the optimization opportunities from both HW and SW.

from intel_extension_for_transformers.backends.neural_engine.compile import compile
# /path/to/your/model is a TensorFlow pb model or ONNX model
model = compile('/path/to/your/model')
inputs = ... # [input_ids, segment_ids, input_mask]
model.inference(inputs)

Please refer to Neural Engine for more details.

Quantized Length Adaptive Transformer

Quantized Length Adaptive Transformer leverages sequence-length reduction and low-bit representation techniques to further enhance model inference performance, enabling adaptive sequence-length sizes to accommodate different computational budget requirements with an optimal accuracy efficiency tradeoff.

from intel_extension_for_transformers.optimization import QuantizationConfig, DynamicLengthConfig, metric, objectives
from intel_extension_for_transformers.optimization.trainer import NLPTrainer

# Replace transformers.Trainer with NLPTrainer
# trainer = transformers.Trainer(...)
trainer = NLPTrainer(...)
metric = metrics.Metric(name="eval_f1", is_relative=True, criterion=0.01)
q_config = QuantizationConfig(
    approach="PostTrainingStatic",
    metrics=[metric],
    objectives=[objectives.performance]
)
# Apply the length config
dynamic_length_config = DynamicLengthConfig(length_config=length_config)
trainer.set_dynamic_config(dynamic_config=dynamic_length_config)
# Quantization
model = trainer.quantize(quant_config=q_config)

Please refer to paper QuaLA-MiniLM and code for details

Transformers-accelerated Neural Engine

Transformers-accelerated Neural Engine is one of reference deployments that Intel® Extension for Transformers provides. Neural Engine aims to demonstrate the optimal performance of extremely compressed NLP models by exploring the optimization opportunities from both HW and SW.

from intel_extension_for_transformers.backends.neural_engine.compile import compile
# /path/to/your/model is a TensorFlow pb model or ONNX model
model = compile('/path/to/your/model')
inputs = ... # [input_ids, segment_ids, input_mask]
model.inference(inputs)

Please refer to example in Transformers-accelerated Neural Engine and paper Fast Distilbert on CPUs for more details.

Transformers-accelerated Libraries

Transformers-accelerated Libraries is a high-performance operator computing library implemented by assembly. Transformers-accelerated Libraries contains a JIT domain, a kernel domain, and a scheduling proxy framework.

#include "interface.hpp"
  ...
  operator_desc op_desc(ker_kind, ker_prop, eng_kind, ts_descs, op_attrs);
  sparse_matmul_desc spmm_desc(op_desc);
  sparse_matmul spmm_kern(spmm_desc);
  std::vector<const void*> rt_data = {data0, data1, data2, data3, data4};
  spmm_kern.execute(rt_data);

Please refer to Transformers-accelerated Libraries for more details.

System Requirements

Validated Hardware Environment

Intel® Extension for Transformers supports systems based on Intel 64 architecture or compatible processors that are specifically optimized for the following CPUs:

Intel Xeon Scalable processor (formerly Cascade Lake, Icelake)
Future Intel Xeon Scalable processor (code name Sapphire Rapids)

Validated Software Environment

OS version: CentOS 8.4, Ubuntu 20.04
Python version: 3.7, 3.8, 3.9

Framework	Intel TensorFlow	PyTorch	IPEX
Version	2.10.0 2.9.1	1.13.0+cpu 1.12.0+cpu 1.11.0+cpu	1.13.0 1.12.0

OS version: Windows 10
Python version: 3.7, 3.8, 3.9

Framework	Intel TensorFlow	PyTorch
Version	2.9.1	1.13.0+cpu

Selected Publications/Events

Blog published on Medium: MLefficiency — Optimizing transformer models for efficiency (Dec 2022)
NeurIPS'2022: Fast Distilbert on CPUs (Nov 2022)
NeurIPS'2022: QuaLA-MiniLM: a Quantized Length Adaptive MiniLM (Nov 2022)
Blog published by Cohere: Top NLP Papers—November 2022 (Nov 2022)
Blog published by Alibaba: Deep learning inference optimization for Address Purification (Aug 2022)
NeurIPS'2021: Prune Once for All: Sparse Pre-Trained Language Models (Nov 2021)

kiminh / intel-extension-for-transformers