Hugging Face's logo

qualcomm
/
LeViT

Image Classification
PyTorch
LiteRT
bu_auto
android
Model card Files
xet
Community

LeViT: Optimized for Mobile Deployment

Imagenet classifier and general purpose backbone

LeViT is a vision transformer model that can classify images from the Imagenet dataset.

This model is an implementation of LeViT found here.

This repository provides scripts to run LeViT on Qualcomm® devices. More details on model performance across various devices, can be found here.

Model Details

  • Model Type: Model_use_case.image_classification
  • Model Stats:
    • Model checkpoint: LeViT-128S
    • Input resolution: 224x224
    • Number of parameters: 7.82M
    • Model size (float): 29.9 MB
    • Model size (w8a16): 8.83 MB
Model Precision Device Chipset Target Runtime Inference Time (ms) Peak Memory Range (MB) Primary Compute Unit Target Model
LeViT float QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) TFLITE 3.915 ms 0 - 150 MB NPU LeViT.tflite
LeViT float QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) QNN_DLC 3.743 ms 1 - 147 MB NPU LeViT.dlc
LeViT float QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) TFLITE 2.187 ms 0 - 177 MB NPU LeViT.tflite
LeViT float QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) QNN_DLC 2.174 ms 1 - 178 MB NPU LeViT.dlc
LeViT float QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) TFLITE 1.446 ms 0 - 3 MB NPU LeViT.tflite
LeViT float QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) QNN_DLC 1.479 ms 1 - 3 MB NPU LeViT.dlc
LeViT float QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) ONNX 1.498 ms 0 - 22 MB NPU LeViT.onnx.zip
LeViT float QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) TFLITE 1.945 ms 0 - 150 MB NPU LeViT.tflite
LeViT float QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) QNN_DLC 1.931 ms 0 - 148 MB NPU LeViT.dlc
LeViT float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile TFLITE 0.987 ms 0 - 184 MB NPU LeViT.tflite
LeViT float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile QNN_DLC 1.032 ms 1 - 178 MB NPU LeViT.dlc
LeViT float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile ONNX 0.995 ms 0 - 153 MB NPU LeViT.onnx.zip
LeViT float Samsung Galaxy S25 Snapdragon® 8 Elite For Galaxy Mobile TFLITE 0.773 ms 0 - 150 MB NPU LeViT.tflite
LeViT float Samsung Galaxy S25 Snapdragon® 8 Elite For Galaxy Mobile QNN_DLC 0.817 ms 0 - 152 MB NPU LeViT.dlc
LeViT float Samsung Galaxy S25 Snapdragon® 8 Elite For Galaxy Mobile ONNX 0.847 ms 0 - 122 MB NPU LeViT.onnx.zip
LeViT float Snapdragon 8 Elite Gen 5 QRD Snapdragon® 8 Elite Gen5 Mobile TFLITE 0.678 ms 0 - 155 MB NPU LeViT.tflite
LeViT float Snapdragon 8 Elite Gen 5 QRD Snapdragon® 8 Elite Gen5 Mobile QNN_DLC 0.741 ms 1 - 151 MB NPU LeViT.dlc
LeViT float Snapdragon 8 Elite Gen 5 QRD Snapdragon® 8 Elite Gen5 Mobile ONNX 0.813 ms 1 - 124 MB NPU LeViT.onnx.zip
LeViT float Snapdragon X Elite CRD Snapdragon® X Elite QNN_DLC 1.711 ms 1 - 1 MB NPU LeViT.dlc
LeViT float Snapdragon X Elite CRD Snapdragon® X Elite ONNX 1.459 ms 16 - 16 MB NPU LeViT.onnx.zip
LeViT w8a16 Dragonwing Q-6690 MTP Qualcomm® Qcm6690 QNN_DLC 5.802 ms 0 - 141 MB NPU LeViT.dlc
LeViT w8a16 QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) QNN_DLC 2.972 ms 0 - 131 MB NPU LeViT.dlc
LeViT w8a16 QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) QNN_DLC 1.419 ms 0 - 3 MB NPU LeViT.dlc
LeViT w8a16 QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) QNN_DLC 7.214 ms 0 - 131 MB NPU LeViT.dlc
LeViT w8a16 Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile QNN_DLC 0.996 ms 0 - 160 MB NPU LeViT.dlc
LeViT w8a16 Samsung Galaxy S25 Snapdragon® 8 Elite For Galaxy Mobile QNN_DLC 0.736 ms 0 - 131 MB NPU LeViT.dlc
LeViT w8a16 Snapdragon 7 Gen 4 QRD Snapdragon® 7 Gen 4 Mobile QNN_DLC 1.48 ms 0 - 136 MB NPU LeViT.dlc
LeViT w8a16 Snapdragon 8 Elite Gen 5 QRD Snapdragon® 8 Elite Gen5 Mobile QNN_DLC 0.632 ms 0 - 133 MB NPU LeViT.dlc
LeViT w8a16 Snapdragon X Elite CRD Snapdragon® X Elite QNN_DLC 1.63 ms 0 - 0 MB NPU LeViT.dlc
LeViT w8a16_mixed_int16 Dragonwing Q-6690 MTP Qualcomm® Qcm6690 QNN_DLC 6.094 ms 0 - 142 MB NPU LeViT.dlc
LeViT w8a16_mixed_int16 QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) QNN_DLC 3.039 ms 0 - 131 MB NPU LeViT.dlc
LeViT w8a16_mixed_int16 QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) QNN_DLC 1.46 ms 0 - 3 MB NPU LeViT.dlc
LeViT w8a16_mixed_int16 QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) QNN_DLC 1.741 ms 0 - 131 MB NPU LeViT.dlc
LeViT w8a16_mixed_int16 Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile QNN_DLC 1.015 ms 0 - 161 MB NPU LeViT.dlc
LeViT w8a16_mixed_int16 Samsung Galaxy S25 Snapdragon® 8 Elite For Galaxy Mobile QNN_DLC 0.751 ms 0 - 134 MB NPU LeViT.dlc
LeViT w8a16_mixed_int16 Snapdragon 7 Gen 4 QRD Snapdragon® 7 Gen 4 Mobile QNN_DLC 1.51 ms 0 - 135 MB NPU LeViT.dlc
LeViT w8a16_mixed_int16 Snapdragon 8 Elite Gen 5 QRD Snapdragon® 8 Elite Gen5 Mobile QNN_DLC 0.648 ms 0 - 132 MB NPU LeViT.dlc
LeViT w8a16_mixed_int16 Snapdragon X Elite CRD Snapdragon® X Elite QNN_DLC 1.669 ms 0 - 0 MB NPU LeViT.dlc

Installation

Install the package via pip:

# NOTE: 3.10 <= PYTHON_VERSION < 3.14 is supported.
pip install "qai-hub-models[levit]"

Configure Qualcomm® AI Hub Workbench to run this model on a cloud-hosted device

Sign-in to Qualcomm® AI Hub Workbench with your Qualcomm® ID. Once signed in navigate to Account -> Settings -> API Token.

With this API token, you can configure your client to run models on the cloud hosted devices.

qai-hub configure --api_token API_TOKEN

Navigate to docs for more information.

Demo off target

The package contains a simple end-to-end demo that downloads pre-trained weights and runs this model on a sample input.

python -m qai_hub_models.models.levit.demo

The above demo runs a reference implementation of pre-processing, model inference, and post processing.

NOTE: If you want running in a Jupyter Notebook or Google Colab like environment, please add the following to your cell (instead of the above).

%run -m qai_hub_models.models.levit.demo

Run model on a cloud-hosted device

In addition to the demo, you can also run the model on a cloud-hosted Qualcomm® device. This script does the following:

  • Performance check on-device on a cloud-hosted device
  • Downloads compiled assets that can be deployed on-device for Android.
  • Accuracy check between PyTorch and on-device outputs.
python -m qai_hub_models.models.levit.export

How does this work?

This export script leverages Qualcomm® AI Hub to optimize, validate, and deploy this model on-device. Lets go through each step below in detail:

Step 1: Compile model for on-device deployment

To compile a PyTorch model for on-device deployment, we first trace the model in memory using the jit.trace and then call the submit_compile_job API.

import torch

import qai_hub as hub
from qai_hub_models.models.levit import Model

# Load the model
torch_model = Model.from_pretrained()

# Device
device = hub.Device("Samsung Galaxy S25")

# Trace model
input_shape = torch_model.get_input_spec()
sample_inputs = torch_model.sample_inputs()

pt_model = torch.jit.trace(torch_model, [torch.tensor(data[0]) for _, data in sample_inputs.items()])

# Compile model on a specific device
compile_job = hub.submit_compile_job(
    model=pt_model,
    device=device,
    input_specs=torch_model.get_input_spec(),
)

# Get target model to run on-device
target_model = compile_job.get_target_model()

Step 2: Performance profiling on cloud-hosted device

After compiling models from step 1. Models can be profiled model on-device using the target_model. Note that this scripts runs the model on a device automatically provisioned in the cloud. Once the job is submitted, you can navigate to a provided job URL to view a variety of on-device performance metrics.

profile_job = hub.submit_profile_job(
    model=target_model,
    device=device,
)

Step 3: Verify on-device accuracy

To verify the accuracy of the model on-device, you can run on-device inference on sample input data on the same cloud hosted device.

input_data = torch_model.sample_inputs()
inference_job = hub.submit_inference_job(
    model=target_model,
    device=device,
    inputs=input_data,
)
    on_device_output = inference_job.download_output_data()

With the output of the model, you can compute like PSNR, relative errors or spot check the output with expected output.

Note: This on-device profiling and inference requires access to Qualcomm® AI Hub Workbench. Sign up for access.

Run demo on a cloud-hosted device

You can also run the demo on-device.

python -m qai_hub_models.models.levit.demo --eval-mode on-device

NOTE: If you want running in a Jupyter Notebook or Google Colab like environment, please add the following to your cell (instead of the above).

%run -m qai_hub_models.models.levit.demo -- --eval-mode on-device

Deploying compiled model to Android

The models can be deployed using multiple runtimes:

  • TensorFlow Lite (.tflite export): This tutorial provides a guide to deploy the .tflite model in an Android application.

  • QNN (.so export ): This sample app provides instructions on how to use the .so shared library in an Android application.

View on Qualcomm® AI Hub

Get more details on LeViT's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

  • The license for the original implementation of LeViT can be found here.

References

Community

Downloads last month
41
Inference Providers NEW
Image Classification
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Paper for qualcomm/LeViT