Transformers documentation

QDQBERT

Transformers

You are viewing v4.29.1 version. A newer version v4.46.3 is available.

Join the Hugging Face community

and get access to the augmented documentation experience

Collaborate on models, datasets and Spaces

Faster examples with accelerated inference

Switch between documentation themes

to get started

QDQBERT

Overview

The QDQBERT model can be referenced in Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation by Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius.

The abstract from the paper is the following:

Quantization techniques can reduce the size of Deep Neural Networks and improve inference latency and throughput by taking advantage of high throughput integer instructions. In this paper we review the mathematical aspects of quantization parameters and evaluate their choices on a wide range of neural network models for different application domains, including vision, speech, and language. We focus on quantization techniques that are amenable to acceleration by processors with high-throughput integer math pipelines. We also present a workflow for 8-bit quantization that is able to maintain accuracy within 1% of the floating-point baseline on all networks studied, including models that are more difficult to quantize, such as MobileNets and BERT-large.

Tips:

QDQBERT model adds fake quantization operations (pair of QuantizeLinear/DequantizeLinear ops) to (i) linear layer inputs and weights, (ii) matmul inputs, (iii) residual add inputs, in BERT model.
QDQBERT requires the dependency of Pytorch Quantization Toolkit. To install pip install pytorch-quantization --extra-index-url https://pypi.ngc.nvidia.com
QDQBERT model can be loaded from any checkpoint of HuggingFace BERT model (for example bert-base-uncased), and perform Quantization Aware Training/Post Training Quantization.
A complete example of using QDQBERT model to perform Quatization Aware Training and Post Training Quantization for SQUAD task can be found at transformers/examples/research_projects/quantization-qdqbert/.

This model was contributed by shangz.

Set default quantizers

QDQBERT model adds fake quantization operations (pair of QuantizeLinear/DequantizeLinear ops) to BERT by TensorQuantizer in Pytorch Quantization Toolkit. TensorQuantizer is the module for quantizing tensors, with QuantDescriptor defining how the tensor should be quantized. Refer to Pytorch Quantization Toolkit userguide for more details.

Before creating QDQBERT model, one has to set the default QuantDescriptor defining default tensor quantizers.

Example:

>>> import pytorch_quantization.nn as quant_nn
>>> from pytorch_quantization.tensor_quant import QuantDescriptor

>>> # The default tensor quantizer is set to use Max calibration method
>>> input_desc = QuantDescriptor(num_bits=8, calib_method="max")
>>> # The default tensor quantizer is set to be per-channel quantization for weights
>>> weight_desc = QuantDescriptor(num_bits=8, axis=((0,)))
>>> quant_nn.QuantLinear.set_default_quant_desc_input(input_desc)
>>> quant_nn.QuantLinear.set_default_quant_desc_weight(weight_desc)

Calibration

Calibration is the terminology of passing data samples to the quantizer and deciding the best scaling factors for tensors. After setting up the tensor quantizers, one can use the following example to calibrate the model:

>>> # Find the TensorQuantizer and enable calibration
>>> for name, module in model.named_modules():
...     if name.endswith("_input_quantizer"):
...         module.enable_calib()
...         module.disable_quant()  # Use full precision data to calibrate

>>> # Feeding data samples
>>> model(x)
>>> # ...

>>> # Finalize calibration
>>> for name, module in model.named_modules():
...     if name.endswith("_input_quantizer"):
...         module.load_calib_amax()
...         module.enable_quant()

>>> # If running on GPU, it needs to call .cuda() again because new tensors will be created by calibration process
>>> model.cuda()

>>> # Keep running the quantized model
>>> # ...

Export to ONNX

The goal of exporting to ONNX is to deploy inference by TensorRT. Fake quantization will be broken into a pair of QuantizeLinear/DequantizeLinear ONNX ops. After setting static member of TensorQuantizer to use Pytorch’s own fake quantization functions, fake quantized model can be exported to ONNX, follow the instructions in torch.onnx. Example:

>>> from pytorch_quantization.nn import TensorQuantizer

>>> TensorQuantizer.use_fb_fake_quant = True

>>> # Load the calibrated model
>>> ...
>>> # ONNX export
>>> torch.onnx.export(...)

Transformers

QDQBERT

Overview

Set default quantizers

Calibration

Export to ONNX

Documentation resources

QDQBertConfig

class transformers.QDQBertConfig

QDQBertModel

class transformers.QDQBertModel

forward

QDQBertLMHeadModel

class transformers.QDQBertLMHeadModel

forward

QDQBertForMaskedLM

class transformers.QDQBertForMaskedLM

forward

QDQBertForSequenceClassification

class transformers.QDQBertForSequenceClassification

forward

QDQBertForNextSentencePrediction

class transformers.QDQBertForNextSentencePrediction

forward

QDQBertForMultipleChoice

class transformers.QDQBertForMultipleChoice

forward

QDQBertForTokenClassification

class transformers.QDQBertForTokenClassification

forward

QDQBertForQuestionAnswering

class transformers.QDQBertForQuestionAnswering

forward