Bert-Base-Uncased-Hf: Optimized for Qualcomm Devices

Bert 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 Bert-Base-Uncased-Hf 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	w8a16	Universal	QAIRT 2.45, ONNX Runtime 1.25.0	Download
QNN_DLC	float	Universal	QAIRT 2.45	Download
QNN_DLC	w8a16	Universal	QAIRT 2.45	Download
TFLITE	float	Universal	QAIRT 2.45	Download

For more device-specific assets and performance metrics, visit Bert-Base-Uncased-Hf 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 Bert-Base-Uncased-Hf on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.text_generation

Model Stats:

• Model checkpoint: google-bert/bert-base-uncased
• Input resolution: 1x384
• Number of parameters: 110M
• Model size (float): 418 MB

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
Bert-Base-Uncased-Hf	ONNX	w8a16	Snapdragon® X2 Elite	4.854 ms	213 - 213 MB	NPU
Bert-Base-Uncased-Hf	ONNX	w8a16	Snapdragon® X Elite	10.451 ms	153 - 153 MB	NPU
Bert-Base-Uncased-Hf	ONNX	w8a16	Snapdragon® 8 Gen 3 Mobile	7.047 ms	0 - 512 MB	NPU
Bert-Base-Uncased-Hf	ONNX	w8a16	Qualcomm® QCS8550 (Proxy)	9.966 ms	0 - 164 MB	NPU
Bert-Base-Uncased-Hf	ONNX	w8a16	Qualcomm® QCS9075	10.835 ms	0 - 45 MB	NPU
Bert-Base-Uncased-Hf	ONNX	w8a16	Snapdragon® 8 Elite Gen 5 Mobile	4.591 ms	0 - 458 MB	NPU
Bert-Base-Uncased-Hf	ONNX	w8a16	Snapdragon® 8 Elite Mobile	6.164 ms	0 - 456 MB	NPU
Bert-Base-Uncased-Hf	ONNX	w8a16	Qualcomm® QCM6690	121.571 ms	12 - 536 MB	NPU
Bert-Base-Uncased-Hf	ONNX	w8a16	Snapdragon® 7 Gen 4 Mobile	16.5 ms	12 - 474 MB	NPU
Bert-Base-Uncased-Hf	ONNX	w8a16	Qualcomm® QCS7790	16.5 ms	12 - 474 MB	NPU
Bert-Base-Uncased-Hf	ONNX	w8a16	Qualcomm® QCS8750	6.164 ms	0 - 456 MB	NPU
Bert-Base-Uncased-Hf	ONNX	w8a16	Qualcomm® QCS7181	10.451 ms	153 - 153 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	float	Snapdragon® X2 Elite	9.779 ms	1 - 1 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	float	Snapdragon® X Elite	21.026 ms	0 - 0 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	16.201 ms	0 - 655 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	float	Snapdragon® 8 Gen 1 Mobile	48.811 ms	0 - 489 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	float	Qualcomm® QCS8275	79.504 ms	0 - 473 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	21.83 ms	0 - 3 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	float	Qualcomm® QCS8450	48.811 ms	0 - 489 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	float	Snapdragon® 8 Elite Mobile	10.982 ms	0 - 470 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	float	Qualcomm® SA8295P	35.225 ms	0 - 287 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	8.748 ms	0 - 492 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	float	Qualcomm® SA7255P	79.504 ms	0 - 473 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	float	Qualcomm® QCS9075	27.951 ms	2 - 4 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	float	Qualcomm® QCS8750	10.982 ms	0 - 470 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	float	Qualcomm® QCS7181	21.026 ms	0 - 0 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	w8a16	Snapdragon® X2 Elite	5.384 ms	1 - 1 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	w8a16	Snapdragon® X Elite	11.513 ms	0 - 0 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	w8a16	Snapdragon® 8 Gen 3 Mobile	7.476 ms	0 - 519 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	w8a16	Qualcomm® QCS8275	26.025 ms	0 - 444 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	w8a16	Qualcomm® QCS8550 (Proxy)	10.873 ms	0 - 23 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	w8a16	Qualcomm® QCS9075	13.637 ms	2 - 4 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	w8a16	Snapdragon® 8 Elite Gen 5 Mobile	4.425 ms	0 - 464 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	w8a16	Snapdragon® 8 Elite Mobile	6.07 ms	0 - 460 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	w8a16	Qualcomm® SA7255P	26.025 ms	0 - 444 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	w8a16	Qualcomm® QCS8750	6.07 ms	0 - 460 MB	NPU
Bert-Base-Uncased-Hf	QNN_DLC	w8a16	Qualcomm® QCS7181	11.513 ms	0 - 0 MB	NPU
Bert-Base-Uncased-Hf	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	16.422 ms	0 - 663 MB	NPU
Bert-Base-Uncased-Hf	TFLITE	float	Snapdragon® 8 Gen 1 Mobile	48.744 ms	0 - 494 MB	NPU
Bert-Base-Uncased-Hf	TFLITE	float	Qualcomm® QCS8275	79.802 ms	0 - 486 MB	NPU
Bert-Base-Uncased-Hf	TFLITE	float	Qualcomm® QCS8550 (Proxy)	22.327 ms	0 - 3 MB	NPU
Bert-Base-Uncased-Hf	TFLITE	float	Qualcomm® SA8775P	298.339 ms	2 - 28 MB	GPU
Bert-Base-Uncased-Hf	TFLITE	float	Qualcomm® SA8650P	298.339 ms	2 - 28 MB	GPU
Bert-Base-Uncased-Hf	TFLITE	float	Qualcomm® SA8255P	298.339 ms	2 - 28 MB	GPU
Bert-Base-Uncased-Hf	TFLITE	float	Qualcomm® QCS8450	48.744 ms	0 - 494 MB	NPU
Bert-Base-Uncased-Hf	TFLITE	float	Snapdragon® 8 Elite Mobile	11.154 ms	0 - 479 MB	NPU
Bert-Base-Uncased-Hf	TFLITE	float	Qualcomm® SA8295P	35.141 ms	0 - 290 MB	NPU
Bert-Base-Uncased-Hf	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	8.999 ms	0 - 494 MB	NPU
Bert-Base-Uncased-Hf	TFLITE	float	Qualcomm® SA7255P	79.802 ms	0 - 486 MB	NPU
Bert-Base-Uncased-Hf	TFLITE	float	Qualcomm® QCS9075	28.22 ms	0 - 259 MB	NPU
Bert-Base-Uncased-Hf	TFLITE	float	Qualcomm® QCS8750	11.154 ms	0 - 479 MB	NPU

License

• The license for the original implementation of Bert-Base-Uncased-Hf can be found here.

References

• BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
• 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.