Quark ONNX Quantization Tutorial For Block Floating Point (BFP)

This example demonstrates how to apply BFP16 quantization with Adaquant to the mobilenetv2_050.lamb_in1k model using the ONNX quantization tool in AMD Quark. Block Floating Point (BFP) quantization reduces computational complexity by dividing tensor values into blocks that share a common exponent while maintaining individual mantissas. This representation enables more efficient memory usage compared to full FP16/FP32 formats while retaining higher numerical precision than pure integer quantization. As a result, BFP provides an effective balance between performance efficiency and accuracy, making it well suited for edge and performance-sensitive inference deployments.

Standard INT8 quantization often introduces noticeable accuracy degradation, and in some cases BFP16 alone may still fall short of the desired accuracy. To further minimize quantization loss and maintain model fidelity, we integrate AdaQuant as an additional optimization step.

The example has the following parts:

• Install requirements

• Prepare model

• Prepare data

• Quantizatize with BFP16 only

• Quantizatize with BFP16 and AdaQuant

• Evaluate Models

1) Install The Necessary Python Packages:

In addition to Quark that must be installed as documented at here, extra packages are require for this tutorial.

%pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu
%pip install amd-quark
%pip install -r ./requirements.txt

2) Export ONNX Model From mobilenetv2_050.lamb_in1k Torch Model.

You don’t need to download this model manually. If you’re curious about its source, the corresponding model link is: https://huggingface.co/timm/mobilenetv2_050.lamb_in1k

Before exporting, let’s create a directory for models:

!mkdir -p models

import os
import shutil

import timm
import torch

model_name = "mobilenetv2_050.lamb_in1k"

model = timm.create_model(model_name, pretrained=True)
model = model.eval()
device = torch.device("cpu")

data_config = timm.data.resolve_model_data_config(
    model=model,
    use_test_size=True,
)

batch_size = 1
torch.manual_seed(42)
dummy_input = torch.randn((batch_size,) + tuple(data_config["input_size"])).to(device)

torch.onnx.export(
    model,
    dummy_input,
    "models/" + model_name + ".onnx",
    export_params=True,
    do_constant_folding=True,
    opset_version=17,
    input_names=["input"],
    output_names=["output"],
    dynamic_axes={"input": {0: "batch_size"}, "output": {0: "batch_size"}},
    verbose=False,
    dynamo=False,
)
print("Onnx model is saved at models/" + model_name + ".onnx")

3) Prepare data

ILSVRC 2012, commonly known as ‘ImageNet’. This dataset provides access to ImageNet (ILSVRC) 2012 which is the most commonly used subset of ImageNet. This dataset spans 1000 object classes and contains 50,000 validation images.

If you already have an ImageNet datasets, you can directly use your dataset path.

To prepare the test data, please check the download section of the main website: https://huggingface.co/datasets/imagenet-1k/tree/main/data. You need to register and download val_images.tar.gz to the current directory.

Then, create a val_data folder and decompress the .gz file to the folder.

!mkdir -p val_data && tar -xzf val_images.tar.gz -C val_data

If you have a local cache to store the dataset, you can use and environment variable like LOCAL_DATA_CACHE to specify its path. This is useful to organize and store all your datasets for different experiments in a central place. Otherwise, the current folder is used, and validation dataset and calibration dataset will be created under current directory.

import sys

source_folder = "val_data"
calib_data_path = "calib_data"

if os.environ.get("LOCAL_DATA_CACHE") is not None:
    data_path = os.environ["LOCAL_DATA_CACHE"]
    source_folder = os.path.join(data_path, "Imagenet/val")
    calib_data_path = os.path.join(data_path, "Imagenet/calib_100")
else:
    files = os.listdir(source_folder)

    for filename in files:
        if not filename.startswith("ILSVRC2012_val_") or not filename.endswith(".JPEG"):
            continue

        n_identifier = filename.split("_")[-1].split(".")[0]
        folder_name = n_identifier
        folder_path = os.path.join(source_folder, folder_name)
        if not os.path.exists(folder_path):
            os.makedirs(folder_path)
        file_path = os.path.join(source_folder, filename)
        destination = os.path.join(folder_path, filename)
        shutil.move(file_path, destination)

    print("File organization complete.")

    if not os.path.exists(calib_data_path):
        os.makedirs(calib_data_path)

    destination_folder = calib_data_path

    subfolders = os.listdir(source_folder)

    for subfolder in subfolders:
        source_subfolder = os.path.join(source_folder, subfolder)
        destination_subfolder = os.path.join(destination_folder, subfolder)
        os.makedirs(destination_subfolder, exist_ok=True)
        files = os.listdir(source_subfolder)

        if files:
            file_to_copy = files[0]
            source_file = os.path.join(source_subfolder, file_to_copy)
            destination_file = os.path.join(destination_subfolder, file_to_copy)

            shutil.copy(source_file, destination_file)

    print("Creating calibration dataset complete.")

if not os.path.exists(source_folder):
    print("The provided data path does not exist.")
    sys.exit(1)

The storage format of the val_data of the ImageNet dataset organized as follows:

• val_data

  • n01440764

    • ILSVRC2012_val_00000293.JPEG

    • ILSVRC2012_val_00002138.JPEG

    • …

  • n01443537

    • ILSVRC2012_val_00000236.JPEG

    • ILSVRC2012_val_00000262.JPEG

    • …

  • …

The storage format of the calib_data of the ImageNet dataset organized as follows:

• calib_data

  • n01440764

    • ILSVRC2012_val_00000293.JPEG

  • n01443537

    • ILSVRC2012_val_00000236.JPEG

  • …

4) Quantization Procedure

First, create a data reader that gathers calibration statistics from the target dataset. Next, inside quantize_model, construct the quantized model and pass in your configuration.

import numpy as np
import onnxruntime
import torchvision
from timm.data import resolve_data_config
from timm.models import create_model
from torchvision import transforms

from quark.onnx.operators.custom_ops import get_library_path


def load_loader(model_name, data_dir, batch_size, workers):
    timm_model = create_model(
        model_name,
        pretrained=False,
    )
    data_config = resolve_data_config(model=timm_model, use_test_size=True)
    crop_pct = data_config["crop_pct"]
    input_size = data_config["input_size"]
    width = input_size[-1]
    data_transform = transforms.Compose(
        [
            transforms.Resize(int(width / crop_pct)),
            transforms.CenterCrop(width),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
        ]
    )
    dataset = torchvision.datasets.ImageFolder(data_dir, data_transform)
    data_loader = torch.utils.data.DataLoader(
        dataset, batch_size=batch_size, shuffle=False, num_workers=workers, pin_memory=True
    )
    return data_loader


class CalibrationDataReader:
    def __init__(self, dataloader):
        super().__init__()
        self.iterator = iter(dataloader)

    def get_next(self) -> dict:
        try:
            return {"input": next(self.iterator)[0].numpy()}
        except Exception:
            return None

Now let’s define the quantization process.

import copy

from quark.onnx import AdaQuantConfig, BFP16Spec, ModelQuantizer, QConfig, QLayerConfig


def quantize_model(args: dict) -> None:
    data_loader = load_loader(args["model_name"], args["calibration_dataset_path"], args["batch_size"], args["workers"])
    dr = CalibrationDataReader(data_loader)

    # Get quantization configuration
    algo_config = []
    if args.get("use_adaquant"):
        if args["device"] == "cpu":
            algo_config = [AdaQuantConfig(data_size=100, learning_rate=1e-6)]
        else:
            algo_config = [
                AdaQuantConfig(data_size=100, learning_rate=1e-6, optim_device=args["device"], infer_device="cuda:0")
            ]

    activation_spec = BFP16Spec()
    weight_spec = BFP16Spec()
    config = QConfig(
        global_config=QLayerConfig(activation=activation_spec, weight=weight_spec),
        algo_config=algo_config,
        AlignSlice=False,
        FoldRelu=True,
        AlignConcat=True,
        LogSeverityLevel=3,  # reduce print outs because too many prints will cause the notebook crash
    )
    print(f"The configuration for quantization is {config}")

    # Create an ONNX quantizer
    quantizer = ModelQuantizer(config)

    # Quantize the ONNX model
    quantizer.quantize_model(args["input_model_path"], args["output_model_path"], dr)

The cell defines a quantization config with AdaQuant disabled, and then generates a quantized model to the models directory using the BFP16Spec configuration.

quant_config = {
    "model_name": "mobilenetv2_050.lamb_in1k",
    "input_model_path": "models/mobilenetv2_050.lamb_in1k.onnx",
    "output_model_path": "models/mobilenetv2_050.lamb_in1k_quantized.onnx",
    "calibration_dataset_path": calib_data_path,
    "batch_size": 1,
    "workers": 1,
    "device": "cpu",
}

quantize_model(quant_config)

The cell applies the same default quantization scheme and adds the AdaQuant option, then generates the quantized model into the models directory.

quant_config_with_adaquant = copy.deepcopy(quant_config)
quant_config_with_adaquant["output_model_path"] = "models/mobilenetv2_050.lamb_in1k_adaquant_quantized.onnx"
quant_config_with_adaquant["use_adaquant"] = True

quantize_model(quant_config_with_adaquant)

5) Evaluation and Expected Results

Evaluation is performed on the ImageNet validation set. We compare three models — (1) full-precision, (2) quantized without AdaQuant, and (3) quantized with AdaQuant — to assess AdaQuant’s effectiveness. The full-precision model serves as the baseline for measuring any accuracy change caused by quantization.

ImageNet has 1,000 classes, so we report both Prec@1 and Prec@5 to capture strict and relaxed accuracy. Both metrics are reported as percentages (higher is better). Prec@1 shows exact single-label correctness; Prec@5 is useful on large, fine-grained label spaces because it captures near-misses where the correct class is among the model’s top candidates.

import time


class AverageMeter:
    """Computes and stores the average and current value"""

    def __init__(self):
        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 accuracy_np(output, target):
    max_indices = np.argsort(output, axis=1)[:, ::-1]
    top5 = 100 * np.equal(max_indices[:, :5], target[:, np.newaxis]).sum(axis=1).mean()
    top1 = 100 * np.equal(max_indices[:, 0], target).mean()
    return top1, top5


def metrics(onnx_model_path, sess_options, providers, data_loader, print_freq):
    session = onnxruntime.InferenceSession(onnx_model_path, sess_options, providers=providers)
    input_name = session.get_inputs()[0].name

    batch_time = AverageMeter()
    top1 = AverageMeter()
    top5 = AverageMeter()
    end = time.time()
    for i, (input, target) in enumerate(data_loader):
        # run the net and return prediction
        output = session.run([], {input_name: input.data.numpy()})
        output = output[0]

        # measure accuracy and record loss
        prec1, prec5 = accuracy_np(output, target.numpy())
        top1.update(prec1.item(), input.size(0))
        top5.update(prec5.item(), input.size(0))

        # measure elapsed time
        batch_time.update(time.time() - end)
        end = time.time()

        if i % print_freq == 0:
            print(
                f"Test: [{i}/{len(data_loader)}]\t"
                f"Time {batch_time.val:.3f} ({batch_time.avg:.3f}, {input.size(0) / batch_time.avg:.3f}/s, "
                f"{100 * batch_time.avg / input.size(0):.3f} ms/sample) \t"
                f"Prec@1 {top1.val:.3f} ({top1.avg:.3f})\t"
                f"Prec@5 {top5.val:.3f} ({top5.avg:.3f})"
            )

    return top1, top5


def evaluate(args: dict):
    args["gpu_id"] = 0

    # Set graph optimization level
    sess_options = onnxruntime.SessionOptions()
    sess_options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
    if args.get("profile"):
        sess_options.enable_profiling = True
    if args.get("onnx_output_opt"):
        sess_options.optimized_model_filepath = args["onnx_output_opt"]
    if args.get("gpu"):
        if "ROCMExecutionProvider" in onnxruntime.get_available_providers():
            device = "ROCM"
            providers = ["ROCMExecutionProvider"]
        elif "CUDAExecutionProvider" in onnxruntime.get_available_providers():
            device = "CUDA"
            providers = ["CUDAExecutionProvider"]
        else:
            device = "CPU"
            providers = ["CPUExecutionProvider"]
            print("Warning: GPU is not available, use CPU instead.")
    else:
        device = "CPU"
        providers = ["CPUExecutionProvider"]
    sess_options.register_custom_ops_library(get_library_path(device))

    if args.get("onnx_input"):
        val_loader = load_loader(args["model_name"], args["data"], args["batch_size"], args["workers"])
        f_top1, f_top5 = metrics(args["onnx_input"], sess_options, providers, val_loader, args["print_freq"])
        print(f" * Prec@1 {f_top1.avg:.3f} ({100 - f_top1.avg:.3f}) Prec@5 {f_top5.avg:.3f} ({100.0 - f_top5.avg:.3f})")
    elif args.get("onnx_float") and args.get("onnx_quant"):
        val_loader = load_loader(args[""], args["data"], args["batch_size"], args["workers"])
        f_top1, f_top5 = metrics(args["onnx_float"], sess_options, providers, val_loader, args["print_freq"])
        f_top1 = format(f_top1.avg, ".2f")
        f_top5 = format(f_top5.avg, ".2f")

        q_top1, q_top5 = metrics(args["onnx_quant"], sess_options, providers, val_loader, args["print_freq"])
        q_top1 = format(q_top1.avg, ".2f")
        q_top5 = format(q_top5.avg, ".2f")

        f_size = format(os.path.getsize(args["onnx_float"]) / (1024 * 1024), ".2f")
        q_size = format(os.path.getsize(args["onnx_quant"]) / (1024 * 1024), ".2f")
        """
        --------------------------------------------------------
        |             | float model    | quantized model |
        --------------------------------------------------------
        | ****        | ****           | ****             |
        --------------------------------------------------------
        | Model Size  | ****           | ****             |
        --------------------------------------------------------
        """
        from rich.console import Console
        from rich.table import Table

        console = Console()

        table = Table()
        table.add_column("")
        table.add_column("Float Model")
        table.add_column("Quantized Model", style="bold green1")

        table.add_row("Model", args["onnx_float"], args["onnx_quant"])
        table.add_row("Model Size", str(f_size) + " MB", str(q_size) + " MB")
        table.add_row("Prec@1", str(f_top1) + " %", str(q_top1) + " %")
        table.add_row("Prec@5", str(f_top5) + " %", str(q_top5) + " %")

        console.print(table)

    else:
        print("Please specify both model-float and model-quant or model-input for evaluation.")

First, define an evaluation config, and record accuracy of the Full Precision model on ImageNet val dataset

eval_config = {
    "data": source_folder,
    "model_name": "mobilenetv2_050.lamb_in1k",
    "batch_size": 1,
    "workers": 1,
    "gpu": False,
    "print_freq": 1000,
}

full_precision_eval_config = copy.deepcopy(eval_config)
full_precision_eval_config["onnx_input"] = "models/mobilenetv2_050.lamb_in1k.onnx"

evaluate(full_precision_eval_config)

Then, specify the path to the quantized model without AdaQuant and record its accuracy on ImageNet val dataset

quant_eval_config = copy.deepcopy(eval_config)
quant_eval_config["onnx_input"] = "models/mobilenetv2_050.lamb_in1k_quantized.onnx"

evaluate(quant_eval_config)

Last, specify the path to the quantized model with AdaQuant and record its accuracy on ImageNet val dataset

adaquant_eval_config = copy.deepcopy(eval_config)
adaquant_eval_config["onnx_input"] = "models/mobilenetv2_050.lamb_in1k_adaquant_quantized.onnx"

evaluate(adaquant_eval_config)

The following table contains the expected results, but please note that different machines can lead to minor variations in the accuracy of quantized model with AdaQuant.

	Float Model	Quantized Model without AdaQuant	Quantized Model with AdaQuant
Model Size	8.7 MB	8.4 MB	8.4 MB
Prec@1	65.424 %	60.838 %	65.220 %
Prec@5	85.788 %	82.658 %	85.584 %