Introduction to Quantization¶

Quantization refers to techniques for performing computations and storing tensors at lower bitwidths than floating point precision. A quantized model executes some or all of the operations on tensors with integers rather than floating point values. This allows for a more compact model representation and the use of high performance vectorized operations on many hardware platforms. PyTorch supports INT8 quantization compared to typical FP32 models allowing for a 4x reduction in the model size and a 4x reduction in memory bandwidth requirements. Hardware support for INT8 computations is typically 2 to 4 times faster compared to FP32 compute. Quantization is primarily a technique to speed up inference and only the forward pass is supported for quantized operators.

PyTorch supports multiple approaches to quantizing a deep learning model. In most cases the model is trained in FP32 and then the model is converted to INT8. In addition, PyTorch also supports quantization aware training, which models quantization errors in both the forward and backward passes using fake-quantization modules. Note that the entire computation is carried out in floating point. At the end of quantization aware training, PyTorch provides conversion functions to convert the trained model into lower precision.

At lower level, PyTorch provides a way to represent quantized tensors and perform operations with them. They can be used to directly construct models that perform all or part of the computation in lower precision. Higher-level APIs are provided that incorporate typical workflows of converting FP32 model to lower precision with minimal accuracy loss.

Today, PyTorch supports the following backends for running quantized operators efficiently:

• x86 CPUs with AVX2 support or higher (without AVX2 some operations have inefficient implementations)

• ARM CPUs (typically found in mobile/embedded devices)

The corresponding implementation is chosen automatically based on the PyTorch build mode.

Quantization API Summary¶

There are three types of quantization supported in PyTorch:

1. dynamic quantization (weights quantized with activations read/stored in floating point and quantized for compute.)

2. static quantization (weights quantized, activations quantized, calibration required post training)

3. quantization aware training (weights quantized, activations quantized, quantization numerics modeled during training)

Please see our Introduction to Quantization on Pytorch blog post for a more comprehensive overview of the tradeoffs between these quantization types.

# original model
# all tensors and computations are in floating point
previous_layer_fp32 -- linear_fp32 -- activation_fp32 -- next_layer_fp32
                 /
linear_weight_fp32

# dynamically quantized model
# linear and conv weights are in int8
previous_layer_fp32 -- linear_int8_w_fp32_inp -- activation_fp32 -- next_layer_fp32
                     /
   linear_weight_int8

import torch

# define a floating point model
class M(torch.nn.Module):
    def __init__(self):
        super(M, self).__init__()
        self.fc = torch.nn.Linear(4, 4)

    def forward(self, x):
        x = self.fc(x)
        return x

# create a model instance
model_fp32 = M()
# create a quantized model instance
model_int8 = torch.quantization.quantize_dynamic(
    model_fp32,  # the original model
    {torch.nn.Linear},  # a set of layers to dynamically quantize
    dtype=torch.qint8)  # the target dtype for quantized weights

# run the model
input_fp32 = torch.randn(4, 4, 4, 4)
res = model_int8(input_fp32)

# original model
# all tensors and computations are in floating point
previous_layer_fp32 -- linear_fp32 -- activation_fp32 -- next_layer_fp32
                    /
    linear_weight_fp32

# statically quantized model
# weights and activations are in int8
previous_layer_int8 -- linear_with_activation_int8 -- next_layer_int8
                    /
  linear_weight_int8

import torch

# define a floating point model where some layers could be statically quantized
class M(torch.nn.Module):
    def __init__(self):
        super(M, self).__init__()
        # QuantStub converts tensors from floating point to quantized
        self.quant = torch.quantization.QuantStub()
        self.conv = torch.nn.Conv2d(1, 1, 1)
        self.relu = torch.nn.ReLU()
        # DeQuantStub converts tensors from quantized to floating point
        self.dequant = torch.quantization.DeQuantStub()

    def forward(self, x):
        # manually specify where tensors will be converted from floating
        # point to quantized in the quantized model
        x = self.quant(x)
        x = self.conv(x)
        x = self.relu(x)
        # manually specify where tensors will be converted from quantized
        # to floating point in the quantized model
        x = self.dequant(x)
        return x

# create a model instance
model_fp32 = M()

# model must be set to eval mode for static quantization logic to work
model_fp32.eval()

# attach a global qconfig, which contains information about what kind
# of observers to attach. Use 'fbgemm' for server inference and
# 'qnnpack' for mobile inference. Other quantization configurations such
# as selecting symmetric or assymetric quantization and MinMax or L2Norm
# calibration techniques can be specified here.
model_fp32.qconfig = torch.quantization.get_default_qconfig('fbgemm')

# Fuse the activations to preceding layers, where applicable.
# This needs to be done manually depending on the model architecture.
# Common fusions include `conv + relu` and `conv + batchnorm + relu`
model_fp32_fused = torch.quantization.fuse_modules(model_fp32, [['conv', 'relu']])

# Prepare the model for static quantization. This inserts observers in
# the model that will observe activation tensors during calibration.
model_fp32_prepared = torch.quantization.prepare(model_fp32_fused)

# calibrate the prepared model to determine quantization parameters for activations
# in a real world setting, the calibration would be done with a representative dataset
input_fp32 = torch.randn(4, 1, 4, 4)
model_fp32_prepared(input_fp32)

# Convert the observed model to a quantized model. This does several things:
# quantizes the weights, computes and stores the scale and bias value to be
# used with each activation tensor, and replaces key operators with quantized
# implementations.
model_int8 = torch.quantization.convert(model_fp32_prepared)

# run the model, relevant calculations will happen in int8
res = model_int8(input_fp32)

# original model
# all tensors and computations are in floating point
previous_layer_fp32 -- linear_fp32 -- activation_fp32 -- next_layer_fp32
                      /
    linear_weight_fp32

# model with fake_quants for modeling quantization numerics during training
previous_layer_fp32 -- fq -- linear_fp32 -- activation_fp32 -- fq -- next_layer_fp32
                           /
   linear_weight_fp32 -- fq

# quantized model
# weights and activations are in int8
previous_layer_int8 -- linear_with_activation_int8 -- next_layer_int8
                     /
   linear_weight_int8

import torch

# define a floating point model where some layers could benefit from QAT
class M(torch.nn.Module):
    def __init__(self):
        super(M, self).__init__()
        # QuantStub converts tensors from floating point to quantized
        self.quant = torch.quantization.QuantStub()
        self.conv = torch.nn.Conv2d(1, 1, 1)
        self.bn = torch.nn.BatchNorm2d(1)
        self.relu = torch.nn.ReLU()
        # DeQuantStub converts tensors from quantized to floating point
        self.dequant = torch.quantization.DeQuantStub()

    def forward(self, x):
        x = self.quant(x)
        x = self.conv(x)
        x = self.bn(x)
        x = self.relu(x)
        x = self.dequant(x)
        return x

# create a model instance
model_fp32 = M()

# model must be set to train mode for QAT logic to work
model_fp32.train()

# attach a global qconfig, which contains information about what kind
# of observers to attach. Use 'fbgemm' for server inference and
# 'qnnpack' for mobile inference. Other quantization configurations such
# as selecting symmetric or assymetric quantization and MinMax or L2Norm
# calibration techniques can be specified here.
model_fp32.qconfig = torch.quantization.get_default_qat_qconfig('fbgemm')

# fuse the activations to preceding layers, where applicable
# this needs to be done manually depending on the model architecture
model_fp32_fused = torch.quantization.fuse_modules(model_fp32,
    [['conv', 'bn', 'relu']])

# Prepare the model for QAT. This inserts observers and fake_quants in
# the model that will observe weight and activation tensors during calibration.
model_fp32_prepared = torch.quantization.prepare_qat(model_fp32_fused)

# run the training loop (not shown)
training_loop(model_fp32_prepared)

# Convert the observed model to a quantized model. This does several things:
# quantizes the weights, computes and stores the scale and bias value to be
# used with each activation tensor, fuses modules where appropriate,
# and replaces key operators with quantized implementations.
model_fp32_prepared.eval()
model_int8 = torch.quantization.convert(model_fp32_prepared)

# run the model, relevant calculations will happen in int8
res = model_int8(input_fp32)

Quantized Tensors¶

PyTorch supports both per tensor and per channel asymmetric linear quantization. Per tensor means that all the values within the tensor are scaled the same way. Per channel means that for each dimension, typically the channel dimension of a tensor, the values in the tensor are scaled and offset by a different value (effectively the scale and offset become vectors). This allows for lesser error in converting tensors to quantized values.

The mapping is performed by converting the floating point tensors using

Note that, we ensure that zero in floating point is represented with no error after quantization, thereby ensuring that operations like padding do not cause additional quantization error.

In order to do quantization in PyTorch, we need to be able to represent quantized data in Tensors. A Quantized Tensor allows for storing quantized data (represented as int8/uint8/int32) along with quantization parameters like scale and zero_point. Quantized Tensors allow for many useful operations making quantized arithmetic easy, in addition to allowing for serialization of data in a quantized format.

Quantization Operation coverage¶

Quantized Tensors support a limited subset of data manipulation methods of the regular full-precision tensor. For NN operators included in PyTorch, we restrict support to:

1. 8 bit weights (data_type = qint8)

2. 8 bit activations (data_type = quint8)

Note that operator implementations currently only support per channel quantization for weights of the conv and linear operators. Furthermore the minimum and the maximum of the input data is mapped linearly to the minimum and the maximum of the quantized data type such that zero is represented with no quantization error.

Additional data types and quantization schemes can be implemented through the custom operator mechanism.

Many operations for quantized tensors are available under the same API as full float version in torch or torch.nn. Quantized version of NN modules that perform re-quantization are available in torch.nn.quantized. Those operations explicitly take output quantization parameters (scale and zero_point) in the operation signature.

In addition, we also support fused versions corresponding to common fusion patterns that impact quantization at: torch.nn.intrinsic.quantized.

For quantization aware training, we support modules prepared for quantization aware training at torch.nn.qat and torch.nn.intrinsic.qat

The list of supported operations is sufficient to cover typical CNN and RNN models

Quantization Customizations¶

While default implementations of observers to select the scale factor and bias based on observed tensor data are provided, developers can provide their own quantization functions. Quantization can be applied selectively to different parts of the model or configured differently for different parts of the model.

We also provide support for per channel quantization for conv2d(), conv3d() and linear()

Quantization workflows work by adding (e.g. adding observers as .observer submodule) or replacing (e.g. converting nn.Conv2d to nn.quantized.Conv2d) submodules in the model’s module hierarchy. It means that the model stays a regular nn.Module-based instance throughout the process and thus can work with the rest of PyTorch APIs.

Model Preparation for Quantization¶

It is necessary to currently make some modifications to the model definition prior to quantization. This is because currently quantization works on a module by module basis. Specifically, for all quantization techniques, the user needs to:

1. Convert any operations that require output requantization (and thus have additional parameters) from functionals to module form (for example, using torch.nn.ReLU instead of torch.nn.functional.relu).

2. Specify which parts of the model need to be quantized either by assigning .qconfig attributes on submodules or by specifying qconfig_dict. For example, setting model.conv1.qconfig = None means that the model.conv layer will not be quantized, and setting model.linear1.qconfig = custom_qconfig means that the quantization settings for model.linear1 will be using custom_qconfig instead of the global qconfig.

For static quantization techniques which quantize activations, the user needs to do the following in addition:

1. Specify where activations are quantized and de-quantized. This is done using QuantStub and DeQuantStub modules.

2. Use torch.nn.quantized.FloatFunctional to wrap tensor operations that require special handling for quantization into modules. Examples are operations like add and cat which require special handling to determine output quantization parameters.

3. Fuse modules: combine operations/modules into a single module to obtain higher accuracy and performance. This is done using the torch.quantization.fuse_modules() API, which takes in lists of modules to be fused. We currently support the following fusions: [Conv, Relu], [Conv, BatchNorm], [Conv, BatchNorm, Relu], [Linear, Relu]

Best Practices¶

1. Set the reduce_range argument on observers to True if you are using the fbgemm backend. This argument prevents overflow on some int8 instructions by reducing the range of quantized data type by 1 bit.

Modules that provide quantization functions and classes¶