EfficientViT-l2-cls: Optimized for Qualcomm Devices

EfficientViT 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 EfficientViT-l2-cls 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	w8a16_mixed_fp16	Universal	QAIRT 2.42, ONNX Runtime 1.24.1	Download
QNN_DLC	float	Universal	QAIRT 2.43	Download
TFLITE	float	Universal	QAIRT 2.43, TFLite 2.17.0	Download

For more device-specific assets and performance metrics, visit EfficientViT-l2-cls 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 EfficientViT-l2-cls 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: 63.7M
• Model size (float): 243 MB

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
EfficientViT-l2-cls	ONNX	float	Snapdragon® X Elite	7.971 ms	131 - 131 MB	NPU
EfficientViT-l2-cls	ONNX	float	Snapdragon® 8 Gen 3 Mobile	5.264 ms	0 - 246 MB	NPU
EfficientViT-l2-cls	ONNX	float	Qualcomm® QCS8550 (Proxy)	7.33 ms	0 - 162 MB	NPU
EfficientViT-l2-cls	ONNX	float	Qualcomm® QCS9075	8.39 ms	0 - 4 MB	NPU
EfficientViT-l2-cls	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	3.956 ms	0 - 132 MB	NPU
EfficientViT-l2-cls	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	3.223 ms	1 - 158 MB	NPU
EfficientViT-l2-cls	ONNX	float	Snapdragon® X2 Elite	3.516 ms	132 - 132 MB	NPU
EfficientViT-l2-cls	QNN_DLC	float	Snapdragon® X Elite	7.918 ms	1 - 1 MB	NPU
EfficientViT-l2-cls	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	5.293 ms	0 - 233 MB	NPU
EfficientViT-l2-cls	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	24.397 ms	1 - 141 MB	NPU
EfficientViT-l2-cls	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	7.339 ms	1 - 2 MB	NPU
EfficientViT-l2-cls	QNN_DLC	float	Qualcomm® QCS9075	8.628 ms	3 - 5 MB	NPU
EfficientViT-l2-cls	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	14.909 ms	0 - 222 MB	NPU
EfficientViT-l2-cls	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	3.946 ms	0 - 150 MB	NPU
EfficientViT-l2-cls	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	3.225 ms	0 - 145 MB	NPU
EfficientViT-l2-cls	QNN_DLC	float	Snapdragon® X2 Elite	3.977 ms	1 - 1 MB	NPU
EfficientViT-l2-cls	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	5.245 ms	0 - 369 MB	NPU
EfficientViT-l2-cls	TFLITE	float	Qualcomm® QCS8275 (Proxy)	24.401 ms	0 - 276 MB	NPU
EfficientViT-l2-cls	TFLITE	float	Qualcomm® QCS8550 (Proxy)	6.995 ms	0 - 7 MB	NPU
EfficientViT-l2-cls	TFLITE	float	Qualcomm® QCS9075	8.57 ms	0 - 134 MB	NPU
EfficientViT-l2-cls	TFLITE	float	Qualcomm® QCS8450 (Proxy)	14.81 ms	0 - 363 MB	NPU
EfficientViT-l2-cls	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	3.988 ms	0 - 267 MB	NPU
EfficientViT-l2-cls	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	3.21 ms	0 - 280 MB	NPU

License

• The license for the original implementation of EfficientViT-l2-cls can be found here.

References

• EfficientViT: Multi-Scale Linear Attention for High-Resolution Dense Prediction
• 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.