LeViT: Optimized for Qualcomm Devices

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

This is based on the implementation of LeViT found here. This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the Qualcomm® AI Hub Models library to export with custom configurations. More details on model performance across various devices, can be found here.

Qualcomm AI Hub Models uses Qualcomm AI Hub Workbench to compile, profile, and evaluate this model. Sign up to run these models on a hosted Qualcomm® device.

Getting Started

There are two ways to deploy this model on your device:

Option 1: Download Pre-Exported Models

Below are pre-exported model assets ready for deployment.

Runtime	Precision	Chipset	SDK Versions	Download
ONNX	float	Universal	QAIRT 2.37, ONNX Runtime 1.23.0	Download
ONNX	w8a16	Universal	QAIRT 2.37, ONNX Runtime 1.23.0	Download
ONNX	w8a16_mixed_int16	Universal	QAIRT 2.37, ONNX Runtime 1.23.0	Download
QNN_DLC	float	Universal	QAIRT 2.42	Download
QNN_DLC	w8a16	Universal	QAIRT 2.42	Download
QNN_DLC	w8a16_mixed_int16	Universal	QAIRT 2.42	Download
TFLITE	float	Universal	QAIRT 2.42, TFLite 2.17.0	Download

For more device-specific assets and performance metrics, visit LeViT on Qualcomm® AI Hub.

Option 2: Export with Custom Configurations

Use the Qualcomm® AI Hub Models Python library to compile and export the model with your own:

• Custom weights (e.g., fine-tuned checkpoints)
• Custom input shapes
• Target device and runtime configurations

This option is ideal if you need to customize the model beyond the default configuration provided here.

See our repository for LeViT on GitHub for usage instructions.

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

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
LeViT	ONNX	float	Snapdragon® X Elite	1.467 ms	16 - 16 MB	NPU
LeViT	ONNX	float	Snapdragon® 8 Gen 3 Mobile	1.094 ms	0 - 148 MB	NPU
LeViT	ONNX	float	Qualcomm® QCS8550 (Proxy)	1.51 ms	0 - 22 MB	NPU
LeViT	ONNX	float	Qualcomm® QCS9075	1.841 ms	1 - 3 MB	NPU
LeViT	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	0.899 ms	0 - 124 MB	NPU
LeViT	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	0.818 ms	0 - 123 MB	NPU
LeViT	QNN_DLC	float	Snapdragon® X Elite	1.796 ms	1 - 1 MB	NPU
LeViT	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	1.073 ms	0 - 82 MB	NPU
LeViT	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	3.802 ms	1 - 59 MB	NPU
LeViT	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	1.556 ms	1 - 2 MB	NPU
LeViT	QNN_DLC	float	Qualcomm® QCS9075	1.872 ms	3 - 5 MB	NPU
LeViT	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	2.373 ms	0 - 81 MB	NPU
LeViT	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	0.828 ms	0 - 58 MB	NPU
LeViT	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	0.733 ms	1 - 62 MB	NPU
LeViT	QNN_DLC	w8a16	Snapdragon® X Elite	1.633 ms	0 - 0 MB	NPU
LeViT	QNN_DLC	w8a16	Snapdragon® 8 Gen 3 Mobile	1.005 ms	0 - 61 MB	NPU
LeViT	QNN_DLC	w8a16	Qualcomm® QCS8275 (Proxy)	2.968 ms	0 - 41 MB	NPU
LeViT	QNN_DLC	w8a16	Qualcomm® QCS8550 (Proxy)	1.426 ms	0 - 14 MB	NPU
LeViT	QNN_DLC	w8a16	Qualcomm® QCS9075	1.725 ms	0 - 2 MB	NPU
LeViT	QNN_DLC	w8a16	Qualcomm® QCM6690	5.551 ms	0 - 165 MB	NPU
LeViT	QNN_DLC	w8a16	Snapdragon® 8 Elite For Galaxy Mobile	0.738 ms	0 - 42 MB	NPU
LeViT	QNN_DLC	w8a16	Snapdragon® 7 Gen 4 Mobile	1.461 ms	0 - 39 MB	NPU
LeViT	QNN_DLC	w8a16	Snapdragon® 8 Elite Gen 5 Mobile	0.634 ms	0 - 42 MB	NPU
LeViT	QNN_DLC	w8a16_mixed_int16	Snapdragon® X Elite	1.692 ms	0 - 0 MB	NPU
LeViT	QNN_DLC	w8a16_mixed_int16	Snapdragon® 8 Gen 3 Mobile	1.014 ms	0 - 62 MB	NPU
LeViT	QNN_DLC	w8a16_mixed_int16	Qualcomm® QCS8275 (Proxy)	3.041 ms	0 - 40 MB	NPU
LeViT	QNN_DLC	w8a16_mixed_int16	Qualcomm® QCS8550 (Proxy)	1.465 ms	0 - 2 MB	NPU
LeViT	QNN_DLC	w8a16_mixed_int16	Qualcomm® QCS9075	1.726 ms	0 - 2 MB	NPU
LeViT	QNN_DLC	w8a16_mixed_int16	Qualcomm® QCM6690	5.877 ms	0 - 166 MB	NPU
LeViT	QNN_DLC	w8a16_mixed_int16	Snapdragon® 8 Elite For Galaxy Mobile	0.753 ms	0 - 41 MB	NPU
LeViT	QNN_DLC	w8a16_mixed_int16	Snapdragon® 7 Gen 4 Mobile	1.496 ms	0 - 39 MB	NPU
LeViT	QNN_DLC	w8a16_mixed_int16	Snapdragon® 8 Elite Gen 5 Mobile	0.647 ms	0 - 41 MB	NPU
LeViT	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	1.035 ms	0 - 97 MB	NPU
LeViT	TFLITE	float	Qualcomm® QCS8275 (Proxy)	4.035 ms	0 - 70 MB	NPU
LeViT	TFLITE	float	Qualcomm® QCS8550 (Proxy)	1.514 ms	0 - 3 MB	NPU
LeViT	TFLITE	float	Qualcomm® QCS9075	1.883 ms	0 - 19 MB	NPU
LeViT	TFLITE	float	Qualcomm® QCS8450 (Proxy)	2.348 ms	0 - 93 MB	NPU
LeViT	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	0.787 ms	0 - 77 MB	NPU
LeViT	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	0.679 ms	0 - 78 MB	NPU

License

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

References

• LeViT: a Vision Transformer in ConvNet's Clothing for Faster Inference
• Source Model Implementation

Community

• Join our AI Hub Slack community to collaborate, post questions and learn more about on-device AI.
• For questions or feedback please reach out to us.