VIT: Optimized for Qualcomm Devices

VIT is a machine learning model that can classify images from the Imagenet dataset. It can also be used as a backbone in building more complex models for specific use cases.

This is based on the implementation of VIT 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 2.42, ONNX Runtime 1.24.1	Download
ONNX	w8a16	Universal	QAIRT 2.42, ONNX Runtime 1.24.1	Download
ONNX	w8a8	Universal	QAIRT 2.42, ONNX Runtime 1.24.1	Download
ONNX	w8a8_mixed_int16	Universal	QAIRT 2.42, ONNX Runtime 1.24.1	Download
QNN_DLC	float	Universal	QAIRT 2.43	Download
QNN_DLC	w8a16	Universal	QAIRT 2.43	Download
QNN_DLC	w8a8	Universal	QAIRT 2.43	Download
TFLITE	float	Universal	QAIRT 2.43, TFLite 2.17.0	Download
TFLITE	w8a8	Universal	QAIRT 2.43, TFLite 2.17.0	Download

For more device-specific assets and performance metrics, visit VIT 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 VIT on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.image_classification

Model Stats:

Model checkpoint: Imagenet
Input resolution: 224x224
Number of parameters: 86.6M
Model size (float): 330 MB
Model size (w8a16): 86.2 MB
Model size (w8a8): 83.2 MB

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
VIT	ONNX	float	Snapdragon® X Elite	11.17 ms	170 - 170 MB	NPU
VIT	ONNX	float	Snapdragon® 8 Gen 3 Mobile	7.17 ms	0 - 379 MB	NPU
VIT	ONNX	float	Qualcomm® QCS8550 (Proxy)	10.473 ms	0 - 195 MB	NPU
VIT	ONNX	float	Qualcomm® QCS9075	14.319 ms	0 - 4 MB	NPU
VIT	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	4.984 ms	0 - 347 MB	NPU
VIT	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	3.665 ms	1 - 355 MB	NPU
VIT	ONNX	float	Snapdragon® X2 Elite	3.869 ms	170 - 170 MB	NPU
VIT	ONNX	w8a16	Snapdragon® X Elite	11.274 ms	86 - 86 MB	NPU
VIT	ONNX	w8a16	Snapdragon® 8 Gen 3 Mobile	7.368 ms	0 - 377 MB	NPU
VIT	ONNX	w8a16	Qualcomm® QCS6490	1117.365 ms	39 - 56 MB	CPU
VIT	ONNX	w8a16	Qualcomm® QCS8550 (Proxy)	10.702 ms	0 - 6 MB	NPU
VIT	ONNX	w8a16	Qualcomm® QCS9075	13.211 ms	0 - 3 MB	NPU
VIT	ONNX	w8a16	Qualcomm® QCM6690	620.244 ms	68 - 84 MB	CPU
VIT	ONNX	w8a16	Snapdragon® 8 Elite For Galaxy Mobile	5.297 ms	0 - 301 MB	NPU
VIT	ONNX	w8a16	Snapdragon® 7 Gen 4 Mobile	593.006 ms	61 - 77 MB	CPU
VIT	ONNX	w8a16	Snapdragon® 8 Elite Gen 5 Mobile	3.799 ms	0 - 304 MB	NPU
VIT	ONNX	w8a16	Snapdragon® X2 Elite	3.896 ms	86 - 86 MB	NPU
VIT	ONNX	w8a8	Snapdragon® X Elite	13.615 ms	85 - 85 MB	NPU
VIT	ONNX	w8a8	Snapdragon® 8 Gen 3 Mobile	8.822 ms	0 - 469 MB	NPU
VIT	ONNX	w8a8	Qualcomm® QCS6490	332.509 ms	20 - 69 MB	CPU
VIT	ONNX	w8a8	Qualcomm® QCS8550 (Proxy)	12.949 ms	0 - 108 MB	NPU
VIT	ONNX	w8a8	Qualcomm® QCS9075	13.705 ms	0 - 3 MB	NPU
VIT	ONNX	w8a8	Qualcomm® QCM6690	135.965 ms	22 - 43 MB	CPU
VIT	ONNX	w8a8	Snapdragon® 8 Elite For Galaxy Mobile	7.278 ms	0 - 324 MB	NPU
VIT	ONNX	w8a8	Snapdragon® 7 Gen 4 Mobile	129.433 ms	13 - 34 MB	CPU
VIT	ONNX	w8a8	Snapdragon® 8 Elite Gen 5 Mobile	4.584 ms	0 - 350 MB	NPU
VIT	ONNX	w8a8	Snapdragon® X2 Elite	5.104 ms	85 - 85 MB	NPU
VIT	ONNX	w8a8_mixed_int16	Snapdragon® X Elite	178.34 ms	79 - 79 MB	NPU
VIT	ONNX	w8a8_mixed_int16	Snapdragon® 8 Gen 3 Mobile	83.083 ms	67 - 435 MB	NPU
VIT	ONNX	w8a8_mixed_int16	Qualcomm® QCS6490	711.851 ms	96 - 126 MB	CPU
VIT	ONNX	w8a8_mixed_int16	Qualcomm® QCS8550 (Proxy)	99.555 ms	0 - 84 MB	NPU
VIT	ONNX	w8a8_mixed_int16	Qualcomm® QCS9075	130.215 ms	68 - 71 MB	NPU
VIT	ONNX	w8a8_mixed_int16	Qualcomm® QCM6690	386.713 ms	99 - 119 MB	CPU
VIT	ONNX	w8a8_mixed_int16	Snapdragon® 8 Elite For Galaxy Mobile	72.628 ms	68 - 339 MB	NPU
VIT	ONNX	w8a8_mixed_int16	Snapdragon® 7 Gen 4 Mobile	372.159 ms	43 - 63 MB	CPU
VIT	ONNX	w8a8_mixed_int16	Snapdragon® 8 Elite Gen 5 Mobile	63.104 ms	0 - 268 MB	NPU
VIT	ONNX	w8a8_mixed_int16	Snapdragon® X2 Elite	53.294 ms	79 - 79 MB	NPU
VIT	QNN_DLC	float	Snapdragon® X Elite	11.844 ms	1 - 1 MB	NPU
VIT	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	7.724 ms	0 - 371 MB	NPU
VIT	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	40.618 ms	1 - 339 MB	NPU
VIT	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	11.199 ms	1 - 3 MB	NPU
VIT	QNN_DLC	float	Qualcomm® SA8775P	13.884 ms	1 - 339 MB	NPU
VIT	QNN_DLC	float	Qualcomm® QCS9075	15.555 ms	1 - 3 MB	NPU
VIT	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	19.128 ms	0 - 348 MB	NPU
VIT	QNN_DLC	float	Qualcomm® SA7255P	40.618 ms	1 - 339 MB	NPU
VIT	QNN_DLC	float	Qualcomm® SA8295P	17.176 ms	1 - 332 MB	NPU
VIT	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	5.287 ms	1 - 339 MB	NPU
VIT	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	4.051 ms	1 - 346 MB	NPU
VIT	QNN_DLC	float	Snapdragon® X2 Elite	4.466 ms	1 - 1 MB	NPU
VIT	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	5.878 ms	0 - 324 MB	NPU
VIT	TFLITE	float	Qualcomm® QCS8275 (Proxy)	35.879 ms	0 - 288 MB	NPU
VIT	TFLITE	float	Qualcomm® QCS8550 (Proxy)	8.023 ms	0 - 3 MB	NPU
VIT	TFLITE	float	Qualcomm® SA8775P	11.125 ms	0 - 288 MB	NPU
VIT	TFLITE	float	Qualcomm® QCS9075	11.565 ms	0 - 174 MB	NPU
VIT	TFLITE	float	Qualcomm® QCS8450 (Proxy)	13.831 ms	0 - 291 MB	NPU
VIT	TFLITE	float	Qualcomm® SA7255P	35.879 ms	0 - 288 MB	NPU
VIT	TFLITE	float	Qualcomm® SA8295P	13.322 ms	0 - 261 MB	NPU
VIT	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	3.959 ms	0 - 293 MB	NPU
VIT	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	3.091 ms	0 - 278 MB	NPU
VIT	TFLITE	w8a8	Snapdragon® 8 Gen 3 Mobile	4.742 ms	0 - 184 MB	NPU
VIT	TFLITE	w8a8	Qualcomm® QCS6490	60.229 ms	1 - 99 MB	NPU
VIT	TFLITE	w8a8	Qualcomm® QCS8275 (Proxy)	14.331 ms	0 - 85 MB	NPU
VIT	TFLITE	w8a8	Qualcomm® QCS8550 (Proxy)	6.781 ms	0 - 3 MB	NPU
VIT	TFLITE	w8a8	Qualcomm® SA8775P	7.06 ms	0 - 86 MB	NPU
VIT	TFLITE	w8a8	Qualcomm® QCS9075	7.563 ms	0 - 89 MB	NPU
VIT	TFLITE	w8a8	Qualcomm® QCM6690	96.938 ms	2 - 185 MB	NPU
VIT	TFLITE	w8a8	Qualcomm® QCS8450 (Proxy)	8.769 ms	0 - 180 MB	NPU
VIT	TFLITE	w8a8	Qualcomm® SA7255P	14.331 ms	0 - 85 MB	NPU
VIT	TFLITE	w8a8	Qualcomm® SA8295P	9.674 ms	0 - 89 MB	NPU
VIT	TFLITE	w8a8	Snapdragon® 8 Elite For Galaxy Mobile	3.364 ms	0 - 85 MB	NPU
VIT	TFLITE	w8a8	Snapdragon® 7 Gen 4 Mobile	20.391 ms	1 - 74 MB	NPU
VIT	TFLITE	w8a8	Snapdragon® 8 Elite Gen 5 Mobile	2.293 ms	0 - 89 MB	NPU

License

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

References

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.

Downloads last month: -; Downloads are not tracked for this model. How to track

Paper for qualcomm/VIT

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

Paper • 2010.11929 • Published Oct 22, 2020 • 15