Deploy a Quantized Model on Cuda

Author: Wuwei Lin

This article is an introductory tutorial of automatic quantization with TVM. Automatic quantization is one of the quantization modes in TVM. More details on the quantization story in TVM can be found here. In this tutorial, we will import a GluonCV pre-trained model on ImageNet to Relay, quantize the Relay model and then perform the inference.

import tvm
from tvm import te
from tvm import relay
import mxnet as mx
from import download_testdata
from mxnet import gluon
import logging
import os

batch_size = 1
model_name = "resnet18_v1"
target = "cuda"
dev = tvm.device(target)

Prepare the Dataset

We will demonstrate how to prepare the calibration dataset for quantization. We first download the validation set of ImageNet and pre-process the dataset.

calibration_rec = download_testdata(

def get_val_data(num_workers=4):
    mean_rgb = [123.68, 116.779, 103.939]
    std_rgb = [58.393, 57.12, 57.375]

    def batch_fn(batch):
        return[0].asnumpy(), batch.label[0].asnumpy()

    img_size = 299 if model_name == "inceptionv3" else 224
    val_data =
        data_shape=(3, img_size, img_size),
    return val_data, batch_fn

The calibration dataset should be an iterable object. We define the calibration dataset as a generator object in Python. In this tutorial, we only use a few samples for calibration.

calibration_samples = 10

def calibrate_dataset():
    val_data, batch_fn = get_val_data()
    for i, batch in enumerate(val_data):
        if i * batch_size >= calibration_samples:
        data, _ = batch_fn(batch)
        yield {"data": data}

Import the model

We use the Relay MxNet frontend to import a model from the Gluon model zoo.

def get_model():
    gluon_model =, pretrained=True)
    img_size = 299 if model_name == "inceptionv3" else 224
    data_shape = (batch_size, 3, img_size, img_size)
    mod, params = relay.frontend.from_mxnet(gluon_model, {"data": data_shape})
    return mod, params

Quantize the Model

In quantization, we need to find the scale for each weight and intermediate feature map tensor of each layer.

For weights, the scales are directly calculated based on the value of the weights. Two modes are supported: power2 and max. Both modes find the maximum value within the weight tensor first. In power2 mode, the maximum is rounded down to power of two. If the scales of both weights and intermediate feature maps are power of two, we can leverage bit shifting for multiplications. This make it computationally more efficient. In max mode, the maximum is used as the scale. Without rounding, max mode might have better accuracy in some cases. When the scales are not powers of two, fixed point multiplications will be used.

For intermediate feature maps, we can find the scales with data-aware quantization. Data-aware quantization takes a calibration dataset as the input argument. Scales are calculated by minimizing the KL divergence between distribution of activation before and after quantization. Alternatively, we can also use pre-defined global scales. This saves the time for calibration. But the accuracy might be impacted.

def quantize(mod, params, data_aware):
    if data_aware:
        with relay.quantize.qconfig(calibrate_mode="kl_divergence", weight_scale="max"):
            mod = relay.quantize.quantize(mod, params, dataset=calibrate_dataset())
        with relay.quantize.qconfig(calibrate_mode="global_scale", global_scale=8.0):
            mod = relay.quantize.quantize(mod, params)
    return mod

Run Inference

We create a Relay VM to build and execute the model.

def run_inference(mod):
    model = relay.create_executor("vm", mod, dev, target).evaluate()
    val_data, batch_fn = get_val_data()
    for i, batch in enumerate(val_data):
        data, label = batch_fn(batch)
        prediction = model(data)
        if i > 10:  # only run inference on a few samples in this tutorial

def main():
    mod, params = get_model()
    mod = quantize(mod, params, data_aware=True)

if __name__ == "__main__":
/workspace/python/tvm/relay/ DeprecationWarning: Please use input parameter mod (tvm.IRModule) instead of deprecated parameter mod (tvm.relay.function.Function)

Total running time of the script: ( 1 minutes 39.156 seconds)

Gallery generated by Sphinx-Gallery