Mobile-Bert-Uncased-Google: Optimized for Qualcomm Devices

MOBILEBERT is a lightweight BERT model designed for efficient self-supervised learning of language representations. It can be used for masked language modeling and as a backbone for various NLP tasks.

This is based on the implementation of Mobile-Bert-Uncased-Google 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
QNN_DLC	float	Universal	QAIRT 2.43	Download
QNN_DLC	w8a16	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 Mobile-Bert-Uncased-Google 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 Mobile-Bert-Uncased-Google on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.text_generation

Model Stats:

Model checkpoint: mobile_bert_uncased_google
Input resolution: 1x384
Number of parameters: 25.3M
Model size (float): 130 MB

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
Mobile-Bert-Uncased-Google	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	14.096 ms	0 - 230 MB	NPU
Mobile-Bert-Uncased-Google	ONNX	float	Snapdragon® X2 Elite	14.312 ms	80 - 80 MB	NPU
Mobile-Bert-Uncased-Google	ONNX	float	Snapdragon® X Elite	25.782 ms	80 - 80 MB	NPU
Mobile-Bert-Uncased-Google	ONNX	float	Snapdragon® 8 Gen 3 Mobile	20.089 ms	0 - 403 MB	NPU
Mobile-Bert-Uncased-Google	ONNX	float	Qualcomm® QCS8550 (Proxy)	26.339 ms	0 - 91 MB	NPU
Mobile-Bert-Uncased-Google	ONNX	float	Qualcomm® QCS9075	28.822 ms	0 - 3 MB	NPU
Mobile-Bert-Uncased-Google	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	15.467 ms	0 - 206 MB	NPU
Mobile-Bert-Uncased-Google	ONNX	w8a16	Snapdragon® 8 Elite Gen 5 Mobile	10.877 ms	0 - 412 MB	NPU
Mobile-Bert-Uncased-Google	ONNX	w8a16	Snapdragon® X2 Elite	11.162 ms	46 - 46 MB	NPU
Mobile-Bert-Uncased-Google	ONNX	w8a16	Snapdragon® X Elite	18.306 ms	47 - 47 MB	NPU
Mobile-Bert-Uncased-Google	ONNX	w8a16	Snapdragon® 8 Gen 3 Mobile	12.924 ms	0 - 536 MB	NPU
Mobile-Bert-Uncased-Google	ONNX	w8a16	Qualcomm® QCS6490	1259.219 ms	124 - 140 MB	CPU
Mobile-Bert-Uncased-Google	ONNX	w8a16	Qualcomm® QCS8550 (Proxy)	17.56 ms	0 - 61 MB	NPU
Mobile-Bert-Uncased-Google	ONNX	w8a16	Qualcomm® QCS9075	21.229 ms	0 - 3 MB	NPU
Mobile-Bert-Uncased-Google	ONNX	w8a16	Qualcomm® QCM6690	564.377 ms	231 - 263 MB	CPU
Mobile-Bert-Uncased-Google	ONNX	w8a16	Snapdragon® 8 Elite For Galaxy Mobile	11.608 ms	0 - 389 MB	NPU
Mobile-Bert-Uncased-Google	ONNX	w8a16	Snapdragon® 7 Gen 4 Mobile	519.5 ms	219 - 250 MB	CPU
Mobile-Bert-Uncased-Google	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	14.022 ms	0 - 208 MB	NPU
Mobile-Bert-Uncased-Google	QNN_DLC	float	Snapdragon® X2 Elite	14.707 ms	1 - 1 MB	NPU
Mobile-Bert-Uncased-Google	QNN_DLC	float	Snapdragon® X Elite	26.062 ms	0 - 0 MB	NPU
Mobile-Bert-Uncased-Google	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	20.052 ms	0 - 319 MB	NPU
Mobile-Bert-Uncased-Google	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	57.495 ms	0 - 205 MB	NPU
Mobile-Bert-Uncased-Google	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	26.454 ms	0 - 29 MB	NPU
Mobile-Bert-Uncased-Google	QNN_DLC	float	Qualcomm® SA8775P	28.682 ms	0 - 205 MB	NPU
Mobile-Bert-Uncased-Google	QNN_DLC	float	Qualcomm® QCS9075	29.012 ms	0 - 2 MB	NPU
Mobile-Bert-Uncased-Google	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	45.987 ms	0 - 315 MB	NPU
Mobile-Bert-Uncased-Google	QNN_DLC	float	Qualcomm® SA7255P	57.495 ms	0 - 205 MB	NPU
Mobile-Bert-Uncased-Google	QNN_DLC	float	Qualcomm® SA8295P	33.831 ms	0 - 204 MB	NPU
Mobile-Bert-Uncased-Google	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	15.485 ms	0 - 209 MB	NPU
Mobile-Bert-Uncased-Google	QNN_DLC	w8a16	Snapdragon® 8 Elite Gen 5 Mobile	10.79 ms	0 - 356 MB	NPU
Mobile-Bert-Uncased-Google	QNN_DLC	w8a16	Snapdragon® X2 Elite	11.4 ms	1 - 1 MB	NPU
Mobile-Bert-Uncased-Google	QNN_DLC	w8a16	Snapdragon® X Elite	18.548 ms	0 - 0 MB	NPU
Mobile-Bert-Uncased-Google	QNN_DLC	w8a16	Snapdragon® 8 Gen 3 Mobile	12.699 ms	0 - 386 MB	NPU
Mobile-Bert-Uncased-Google	QNN_DLC	w8a16	Qualcomm® QCS8275 (Proxy)	33.265 ms	0 - 364 MB	NPU
Mobile-Bert-Uncased-Google	QNN_DLC	w8a16	Qualcomm® QCS8550 (Proxy)	17.71 ms	0 - 2 MB	NPU
Mobile-Bert-Uncased-Google	QNN_DLC	w8a16	Qualcomm® SA8775P	18.204 ms	0 - 304 MB	NPU
Mobile-Bert-Uncased-Google	QNN_DLC	w8a16	Qualcomm® QCS9075	21.083 ms	0 - 2 MB	NPU
Mobile-Bert-Uncased-Google	QNN_DLC	w8a16	Qualcomm® QCS8450 (Proxy)	22.253 ms	0 - 384 MB	NPU
Mobile-Bert-Uncased-Google	QNN_DLC	w8a16	Qualcomm® SA7255P	33.265 ms	0 - 364 MB	NPU
Mobile-Bert-Uncased-Google	QNN_DLC	w8a16	Qualcomm® SA8295P	21.196 ms	0 - 368 MB	NPU
Mobile-Bert-Uncased-Google	QNN_DLC	w8a16	Snapdragon® 8 Elite For Galaxy Mobile	11.605 ms	0 - 360 MB	NPU
Mobile-Bert-Uncased-Google	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	13.706 ms	0 - 267 MB	NPU
Mobile-Bert-Uncased-Google	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	19.218 ms	0 - 355 MB	NPU
Mobile-Bert-Uncased-Google	TFLITE	float	Qualcomm® QCS8275 (Proxy)	55.12 ms	1 - 248 MB	NPU
Mobile-Bert-Uncased-Google	TFLITE	float	Qualcomm® QCS8550 (Proxy)	25.64 ms	0 - 3 MB	NPU
Mobile-Bert-Uncased-Google	TFLITE	float	Qualcomm® SA8775P	27.822 ms	0 - 248 MB	NPU
Mobile-Bert-Uncased-Google	TFLITE	float	Qualcomm® QCS9075	28.158 ms	0 - 82 MB	NPU
Mobile-Bert-Uncased-Google	TFLITE	float	Qualcomm® QCS8450 (Proxy)	44.357 ms	0 - 343 MB	NPU
Mobile-Bert-Uncased-Google	TFLITE	float	Qualcomm® SA7255P	55.12 ms	1 - 248 MB	NPU
Mobile-Bert-Uncased-Google	TFLITE	float	Qualcomm® SA8295P	32.512 ms	0 - 236 MB	NPU
Mobile-Bert-Uncased-Google	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	14.901 ms	0 - 272 MB	NPU

License

The license for the original implementation of Mobile-Bert-Uncased-Google 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/Mobile-Bert-Uncased-Google

MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices

Paper • 2004.02984 • Published Apr 6, 2020 • 1