(prototype) PyTorch 2 Export Post Training Quantization

Author: Jerry Zhang

This tutorial introduces the steps to do post training static quantization in graph mode based on torch._export.export. Compared to FX Graph Mode Quantization, this flow is expected to have significantly higher model coverage (88% on 14K models), better programmability, and a simplified UX.

Exportable by torch.export.export is a prerequisite to use the flow, you can find what are the constructs that’s supported in Export DB.

The high level architecture of quantization 2 with quantizer could look like this:

float_model(Python)                          Example Input
    \                                              /
     \                                            /
—-------------------------------------------------------
|                        export                        |
—-------------------------------------------------------
                            |
                    FX Graph in ATen     Backend Specific Quantizer
                            |                       /
—--------------------------------------------------------
|                     prepare_pt2e                      |
—--------------------------------------------------------
                            |
                     Calibrate/Train
                            |
—--------------------------------------------------------
|                    convert_pt2e                       |
—--------------------------------------------------------
                            |
                    Quantized Model
                            |
—--------------------------------------------------------
|                       Lowering                        |
—--------------------------------------------------------
                            |
        Executorch, Inductor or <Other Backends>

The PyTorch 2 export quantization API looks like this:

import torch
class M(torch.nn.Module):
   def __init__(self):
      super().__init__()
      self.linear = torch.nn.Linear(5, 10)

   def forward(self, x):
      return self.linear(x)


example_inputs = (torch.randn(1, 5),)
m = M().eval()

# Step 1. program capture
# This is available for pytorch 2.5+, for more details on lower pytorch versions
# please check `Export the model with torch.export` section
m = torch.export.export_for_training(m, example_inputs).module()
# we get a model with aten ops


# Step 2. quantization
from torch.ao.quantization.quantize_pt2e import (
  prepare_pt2e,
  convert_pt2e,
)

from torch.ao.quantization.quantizer.xnnpack_quantizer import (
  XNNPACKQuantizer,
  get_symmetric_quantization_config,
)
# backend developer will write their own Quantizer and expose methods to allow
# users to express how they
# want the model to be quantized
quantizer = XNNPACKQuantizer().set_global(get_symmetric_quantization_config())
m = prepare_pt2e(m, quantizer)

# calibration omitted

m = convert_pt2e(m)
# we have a model with aten ops doing integer computations when possible

Motivation of PyTorch 2 Export Quantization

In PyTorch versions prior to 2, we have FX Graph Mode Quantization that uses QConfigMapping and BackendConfig for customizations. QConfigMapping allows modeling users to specify how they want their model to be quantized, BackendConfig allows backend developers to specify the supported ways of quantization in their backend. While that API covers most use cases relatively well, it is not fully extensible. There are two main limitations for the current API:

• Limitation around expressing quantization intentions for complicated operator patterns (how an operator pattern should be observed/quantized) using existing objects: QConfig and QConfigMapping.

• Limited support on how user can express their intention of how they want their model to be quantized. For example, if users want to quantize the every other linear in the model, or the quantization behavior has some dependency on the actual shape of the Tensor (for example, only observe/quantize inputs and outputs when the linear has a 3D input), backend developer or modeling users need to change the core quantization API/flow.

A few improvements could make the existing flow better:

• We use QConfigMapping and BackendConfig as separate objects, QConfigMapping describes user’s intention of how they want their model to be quantized, BackendConfig describes what kind of quantization a backend supports. BackendConfig is backend-specific, but QConfigMapping is not, and the user can provide a QConfigMapping that is incompatible with a specific BackendConfig, this is not a great UX. Ideally, we can structure this better by making both configuration (QConfigMapping) and quantization capability (BackendConfig) backend-specific, so there will be less confusion about incompatibilities.

• In QConfig we are exposing observer/ fake_quant observer classes as an object for the user to configure quantization, this increases the things that the user may need to care about. For example, not only the dtype but also how the observation should happen, these could potentially be hidden from the user so that the user flow is simpler.

Here is a summary of the benefits of the new API:

• Programmability (addressing 1. and 2.): When a user’s quantization needs are not covered by available quantizers, users can build their own quantizer and compose it with other quantizers as mentioned above.

• Simplified UX (addressing 3.): Provides a single instance with which both backend and users interact. Thus you no longer have the user facing quantization config mapping to map users intent and a separate quantization config that backends interact with to configure what backend support. We will still have a method for users to query what is supported in a quantizer. With a single instance, composing different quantization capabilities also becomes more natural than previously.

  For example XNNPACK does not support embedding_byte and we have natively support for this in ExecuTorch. Thus, if we had ExecuTorchQuantizer that only quantized embedding_byte, then it can be composed with XNNPACKQuantizer. (Previously, this used to be concatenating the two BackendConfig together and since options in QConfigMapping are not backend specific, user also need to figure out how to specify the configurations by themselves that matches the quantization capabilities of the combined backend. With a single quantizer instance, we can compose two quantizers and query the composed quantizer for capabilities, which makes it less error prone and cleaner, for example, composed_quantizer.quantization_capabilities()).

• Separation of concerns (addressing 4.): As we design the quantizer API, we also decouple specification of quantization, as expressed in terms of dtype, min/max (# of bits), symmetric, and so on, from the observer concept. Currently, the observer captures both quantization specification and how to observe (Histogram vs MinMax observer). Modeling users are freed from interacting with observer and fake quant objects with this change.

Define Helper Functions and Prepare Dataset

We’ll start by doing the necessary imports, defining some helper functions and prepare the data. These steps are identitcal to Static Quantization with Eager Mode in PyTorch.

To run the code in this tutorial using the entire ImageNet dataset, first download Imagenet by following the instructions at here ImageNet Data. Unzip the downloaded file into the data_path folder.

Download the torchvision resnet18 model and rename it to data/resnet18_pretrained_float.pth.

import os
import sys
import time
import numpy as np

import torch
import torch.nn as nn
from torch.utils.data import DataLoader

import torchvision
from torchvision import datasets
from torchvision.models.resnet import resnet18
import torchvision.transforms as transforms

# Set up warnings
import warnings
warnings.filterwarnings(
    action='ignore',
    category=DeprecationWarning,
    module=r'.*'
)
warnings.filterwarnings(
    action='default',
    module=r'torch.ao.quantization'
)

# Specify random seed for repeatable results
_ = torch.manual_seed(191009)


class AverageMeter(object):
    """Computes and stores the average and current value"""
    def __init__(self, name, fmt=':f'):
        self.name = name
        self.fmt = fmt
        self.reset()

    def reset(self):
        self.val = 0
        self.avg = 0
        self.sum = 0
        self.count = 0

    def update(self, val, n=1):
        self.val = val
        self.sum += val * n
        self.count += n
        self.avg = self.sum / self.count

    def __str__(self):
        fmtstr = '{name} {val' + self.fmt + '} ({avg' + self.fmt + '})'
        return fmtstr.format(**self.__dict__)


def accuracy(output, target, topk=(1,)):
    """
    Computes the accuracy over the k top predictions for the specified
    values of k.
    """
    with torch.no_grad():
        maxk = max(topk)
        batch_size = target.size(0)

        _, pred = output.topk(maxk, 1, True, True)
        pred = pred.t()
        correct = pred.eq(target.view(1, -1).expand_as(pred))

        res = []
        for k in topk:
            correct_k = correct[:k].reshape(-1).float().sum(0, keepdim=True)
            res.append(correct_k.mul_(100.0 / batch_size))
        return res


def evaluate(model, criterion, data_loader):
    model.eval()
    top1 = AverageMeter('Acc@1', ':6.2f')
    top5 = AverageMeter('Acc@5', ':6.2f')
    cnt = 0
    with torch.no_grad():
        for image, target in data_loader:
            output = model(image)
            loss = criterion(output, target)
            cnt += 1
            acc1, acc5 = accuracy(output, target, topk=(1, 5))
            top1.update(acc1[0], image.size(0))
            top5.update(acc5[0], image.size(0))
    print('')

    return top1, top5

def load_model(model_file):
    model = resnet18(pretrained=False)
    state_dict = torch.load(model_file, weights_only=True)
    model.load_state_dict(state_dict)
    model.to("cpu")
    return model

def print_size_of_model(model):
    torch.save(model.state_dict(), "temp.p")
    print("Size (MB):", os.path.getsize("temp.p")/1e6)
    os.remove("temp.p")

def prepare_data_loaders(data_path):
    normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                     std=[0.229, 0.224, 0.225])
    dataset = torchvision.datasets.ImageNet(
        data_path, split="train", transform=transforms.Compose([
            transforms.RandomResizedCrop(224),
            transforms.RandomHorizontalFlip(),
            transforms.ToTensor(),
            normalize,
        ]))
    dataset_test = torchvision.datasets.ImageNet(
        data_path, split="val", transform=transforms.Compose([
            transforms.Resize(256),
            transforms.CenterCrop(224),
            transforms.ToTensor(),
            normalize,
        ]))

    train_sampler = torch.utils.data.RandomSampler(dataset)
    test_sampler = torch.utils.data.SequentialSampler(dataset_test)

    data_loader = torch.utils.data.DataLoader(
        dataset, batch_size=train_batch_size,
        sampler=train_sampler)

    data_loader_test = torch.utils.data.DataLoader(
        dataset_test, batch_size=eval_batch_size,
        sampler=test_sampler)

    return data_loader, data_loader_test

data_path = '~/.data/imagenet'
saved_model_dir = 'data/'
float_model_file = 'resnet18_pretrained_float.pth'

train_batch_size = 30
eval_batch_size = 50

data_loader, data_loader_test = prepare_data_loaders(data_path)
example_inputs = (next(iter(data_loader))[0])
criterion = nn.CrossEntropyLoss()
float_model = load_model(saved_model_dir + float_model_file).to("cpu")
float_model.eval()

# create another instance of the model since
# we need to keep the original model around
model_to_quantize = load_model(saved_model_dir + float_model_file).to("cpu")

Set the model to eval mode

For post training quantization, we’ll need to set the model to the eval mode.

model_to_quantize.eval()

Export the model with torch.export

Here is how you can use torch.export to export the model:

example_inputs = (torch.rand(2, 3, 224, 224),)
# for pytorch 2.5+
exported_model = torch.export.export_for_training(model_to_quantize, example_inputs).module()

# for pytorch 2.4 and before
# from torch._export import capture_pre_autograd_graph
# exported_model = capture_pre_autograd_graph(model_to_quantize, example_inputs)

# or capture with dynamic dimensions
# for pytorch 2.5+
dynamic_shapes = tuple(
  {0: torch.export.Dim("dim")} if i == 0 else None
  for i in range(len(example_inputs))
)
exported_model = torch.export.export_for_training(model_to_quantize, example_inputs, dynamic_shapes=dynamic_shapes).module()

# for pytorch 2.4 and before
# dynamic_shape API may vary as well
# from torch._export import dynamic_dim
# exported_model = capture_pre_autograd_graph(model_to_quantize, example_inputs, constraints=[dynamic_dim(example_inputs[0], 0)])

Import the Backend Specific Quantizer and Configure how to Quantize the Model

The following code snippets describes how to quantize the model:

from torch.ao.quantization.quantizer.xnnpack_quantizer import (
  XNNPACKQuantizer,
  get_symmetric_quantization_config,
)
quantizer = XNNPACKQuantizer()
quantizer.set_global(get_symmetric_quantization_config())

Quantizer is backend specific, and each Quantizer will provide their own way to allow users to configure their model. Just as an example, here is the different configuration APIs supported by XNNPackQuantizer:

quantizer.set_global(qconfig_opt)  # qconfig_opt is an optional quantization config
    .set_object_type(torch.nn.Conv2d, qconfig_opt) # can be a module type
    .set_object_type(torch.nn.functional.linear, qconfig_opt) # or torch functional op
    .set_module_name("foo.bar", qconfig_opt)

Note

Check out our tutorial that describes how to write a new Quantizer.

Prepare the Model for Post Training Quantization

prepare_pt2e folds BatchNorm operators into preceding Conv2d operators, and inserts observers in appropriate places in the model.

prepared_model = prepare_pt2e(exported_model, quantizer)
print(prepared_model.graph)

Calibration

The calibration function is run after the observers are inserted in the model. The purpose for calibration is to run through some sample examples that is representative of the workload (for example a sample of the training data set) so that the observers in themodel are able to observe the statistics of the Tensors and we can later use this information to calculate quantization parameters.

def calibrate(model, data_loader):
    model.eval()
    with torch.no_grad():
        for image, target in data_loader:
            model(image)
calibrate(prepared_model, data_loader_test)  # run calibration on sample data

Convert the Calibrated Model to a Quantized Model

convert_pt2e takes a calibrated model and produces a quantized model.

quantized_model = convert_pt2e(prepared_model)
print(quantized_model)

At this step, we currently have two representations that you can choose from, but exact representation we offer in the long term might change based on feedback from PyTorch users.

• Q/DQ Representation (default)

  Previous documentation for representations all quantized operators are represented as dequantize -> fp32_op -> qauntize.

def quantized_linear(x_int8, x_scale, x_zero_point, weight_int8, weight_scale, weight_zero_point, bias_fp32, output_scale, output_zero_point):
    x_fp32 = torch.ops.quantized_decomposed.dequantize_per_tensor(
             x_i8, x_scale, x_zero_point, x_quant_min, x_quant_max, torch.int8)
    weight_fp32 = torch.ops.quantized_decomposed.dequantize_per_tensor(
             weight_i8, weight_scale, weight_zero_point, weight_quant_min, weight_quant_max, torch.int8)
    weight_permuted = torch.ops.aten.permute_copy.default(weight_fp32, [1, 0]);
    out_fp32 = torch.ops.aten.addmm.default(bias_fp32, x_fp32, weight_permuted)
    out_i8 = torch.ops.quantized_decomposed.quantize_per_tensor(
    out_fp32, out_scale, out_zero_point, out_quant_min, out_quant_max, torch.int8)
    return out_i8

• Reference Quantized Model Representation

  We will have a special representation for selected ops, for example, quantized linear. Other ops are represented as dq -> float32_op -> q and q/dq are decomposed into more primitive operators. You can get this representation by using convert_pt2e(..., use_reference_representation=True).

# Reference Quantized Pattern for quantized linear
def quantized_linear(x_int8, x_scale, x_zero_point, weight_int8, weight_scale, weight_zero_point, bias_fp32, output_scale, output_zero_point):
    x_int16 = x_int8.to(torch.int16)
    weight_int16 = weight_int8.to(torch.int16)
    acc_int32 = torch.ops.out_dtype(torch.mm, torch.int32, (x_int16 - x_zero_point), (weight_int16 - weight_zero_point))
    bias_scale = x_scale * weight_scale
    bias_int32 = out_dtype(torch.ops.aten.div.Tensor, torch.int32, bias_fp32, bias_scale)
    acc_int32 = acc_int32 + bias_int32
    acc_int32 = torch.ops.out_dtype(torch.ops.aten.mul.Scalar, torch.int32, acc_int32, x_scale * weight_scale / output_scale) + output_zero_point
    out_int8 = torch.ops.aten.clamp(acc_int32, qmin, qmax).to(torch.int8)
    return out_int8

See here for the most up-to-date reference representations.

Checking Model Size and Accuracy Evaluation

Now we can compare the size and model accuracy with baseline model.

# Baseline model size and accuracy
print("Size of baseline model")
print_size_of_model(float_model)

top1, top5 = evaluate(float_model, criterion, data_loader_test)
print("Baseline Float Model Evaluation accuracy: %2.2f, %2.2f"%(top1.avg, top5.avg))

# Quantized model size and accuracy
print("Size of model after quantization")
# export again to remove unused weights
quantized_model = torch.export.export_for_training(quantized_model, example_inputs).module()
print_size_of_model(quantized_model)

top1, top5 = evaluate(quantized_model, criterion, data_loader_test)
print("[before serilaization] Evaluation accuracy on test dataset: %2.2f, %2.2f"%(top1.avg, top5.avg))

Note

We can’t do performance evaluation now since the model is not lowered to target device, it’s just a representation of quantized computation in ATen operators.

Note

The weights are still in fp32 right now, we may do constant propagation for quantize op to get integer weights in the future.

If you want to get better accuracy or performance, try configuring quantizer in different ways, and each quantizer will have its own way of configuration, so please consult the documentation for the quantizer you are using to learn more about how you can have more control over how to quantize a model.

Save and Load Quantized Model

We’ll show how to save and load the quantized model.

# 0. Store reference output, for example, inputs, and check evaluation accuracy:
example_inputs = (next(iter(data_loader))[0],)
ref = quantized_model(*example_inputs)
top1, top5 = evaluate(quantized_model, criterion, data_loader_test)
print("[before serialization] Evaluation accuracy on test dataset: %2.2f, %2.2f"%(top1.avg, top5.avg))

# 1. Export the model and Save ExportedProgram
pt2e_quantized_model_file_path = saved_model_dir + "resnet18_pt2e_quantized.pth"
# capture the model to get an ExportedProgram
quantized_ep = torch.export.export(quantized_model, example_inputs)
# use torch.export.save to save an ExportedProgram
torch.export.save(quantized_ep, pt2e_quantized_model_file_path)


# 2. Load the saved ExportedProgram
loaded_quantized_ep = torch.export.load(pt2e_quantized_model_file_path)
loaded_quantized_model = loaded_quantized_ep.module()

# 3. Check results for example inputs and check evaluation accuracy again:
res = loaded_quantized_model(*example_inputs)
print("diff:", ref - res)

top1, top5 = evaluate(loaded_quantized_model, criterion, data_loader_test)
print("[after serialization/deserialization] Evaluation accuracy on test dataset: %2.2f, %2.2f"%(top1.avg, top5.avg))

Output:

[before serialization] Evaluation accuracy on test dataset: 79.82, 94.55
diff: tensor([[0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.],
        ...,
        [0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.]])

[after serialization/deserialization] Evaluation accuracy on test dataset: 79.82, 94.55

Debugging the Quantized Model

You can use Numeric Suite that can help with debugging in eager mode and FX graph mode. The new version of Numeric Suite working with PyTorch 2 Export models is still in development.

Lowering and Performance Evaluation

The model produced at this point is not the final model that runs on the device, it is a reference quantized model that captures the intended quantized computation from the user, expressed as ATen operators and some additional quantize/dequantize operators, to get a model that runs on real devices, we’ll need to lower the model. For example, for the models that run on edge devices, we can lower with delegation and ExecuTorch runtime operators.

Conclusion

In this tutorial, we went through the overall quantization flow in PyTorch 2 Export Quantization using XNNPACKQuantizer and got a quantized model that could be further lowered to a backend that supports inference with XNNPACK backend. To use this for your own backend, please first follow the tutorial and implement a Quantizer for your backend, and then quantize the model with that Quantizer.