Bert-Base-Uncased-Hf: Optimized for Mobile Deployment

Language model for masked language modeling and general-purpose NLP tasks

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 model is an implementation of Bert-Base-Uncased-Hf found here.

This repository provides scripts to run Bert-Base-Uncased-Hf on Qualcomm® devices. More details on model performance across various devices, can be found here.

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

Model	Precision	Device	Chipset	Target Runtime	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit	Target Model
Bert-Base-Uncased-Hf	float	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	TFLITE	80.892 ms	0 - 586 MB	NPU	Bert-Base-Uncased-Hf.tflite
Bert-Base-Uncased-Hf	float	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	QNN_DLC	80.741 ms	0 - 575 MB	NPU	Bert-Base-Uncased-Hf.dlc
Bert-Base-Uncased-Hf	float	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	TFLITE	48.017 ms	0 - 604 MB	NPU	Bert-Base-Uncased-Hf.tflite
Bert-Base-Uncased-Hf	float	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	QNN_DLC	48.253 ms	0 - 610 MB	NPU	Bert-Base-Uncased-Hf.dlc
Bert-Base-Uncased-Hf	float	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	TFLITE	22.04 ms	0 - 4 MB	NPU	Bert-Base-Uncased-Hf.tflite
Bert-Base-Uncased-Hf	float	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	QNN_DLC	22.975 ms	0 - 8 MB	NPU	Bert-Base-Uncased-Hf.dlc
Bert-Base-Uncased-Hf	float	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	ONNX	32.699 ms	0 - 323 MB	NPU	Bert-Base-Uncased-Hf.onnx.zip
Bert-Base-Uncased-Hf	float	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	TFLITE	28.418 ms	0 - 586 MB	NPU	Bert-Base-Uncased-Hf.tflite
Bert-Base-Uncased-Hf	float	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	QNN_DLC	28.193 ms	0 - 573 MB	NPU	Bert-Base-Uncased-Hf.dlc
Bert-Base-Uncased-Hf	float	SA7255P ADP	Qualcomm® SA7255P	TFLITE	80.892 ms	0 - 586 MB	NPU	Bert-Base-Uncased-Hf.tflite
Bert-Base-Uncased-Hf	float	SA7255P ADP	Qualcomm® SA7255P	QNN_DLC	80.741 ms	0 - 575 MB	NPU	Bert-Base-Uncased-Hf.dlc
Bert-Base-Uncased-Hf	float	SA8295P ADP	Qualcomm® SA8295P	TFLITE	35.893 ms	0 - 545 MB	NPU	Bert-Base-Uncased-Hf.tflite
Bert-Base-Uncased-Hf	float	SA8295P ADP	Qualcomm® SA8295P	QNN_DLC	35.864 ms	0 - 543 MB	NPU	Bert-Base-Uncased-Hf.dlc
Bert-Base-Uncased-Hf	float	SA8775P ADP	Qualcomm® SA8775P	TFLITE	28.418 ms	0 - 586 MB	NPU	Bert-Base-Uncased-Hf.tflite
Bert-Base-Uncased-Hf	float	SA8775P ADP	Qualcomm® SA8775P	QNN_DLC	28.193 ms	0 - 573 MB	NPU	Bert-Base-Uncased-Hf.dlc
Bert-Base-Uncased-Hf	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	TFLITE	17.064 ms	0 - 648 MB	NPU	Bert-Base-Uncased-Hf.tflite
Bert-Base-Uncased-Hf	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	QNN_DLC	16.783 ms	0 - 641 MB	NPU	Bert-Base-Uncased-Hf.dlc
Bert-Base-Uncased-Hf	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	ONNX	24.105 ms	0 - 646 MB	NPU	Bert-Base-Uncased-Hf.onnx.zip
Bert-Base-Uncased-Hf	float	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	TFLITE	12.091 ms	0 - 590 MB	NPU	Bert-Base-Uncased-Hf.tflite
Bert-Base-Uncased-Hf	float	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	QNN_DLC	11.984 ms	0 - 580 MB	NPU	Bert-Base-Uncased-Hf.dlc
Bert-Base-Uncased-Hf	float	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	ONNX	16.861 ms	0 - 590 MB	NPU	Bert-Base-Uncased-Hf.onnx.zip
Bert-Base-Uncased-Hf	float	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen 5 Mobile	TFLITE	9.995 ms	0 - 577 MB	NPU	Bert-Base-Uncased-Hf.tflite
Bert-Base-Uncased-Hf	float	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen 5 Mobile	QNN_DLC	9.733 ms	0 - 571 MB	NPU	Bert-Base-Uncased-Hf.dlc
Bert-Base-Uncased-Hf	float	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen 5 Mobile	ONNX	13.828 ms	0 - 604 MB	NPU	Bert-Base-Uncased-Hf.onnx.zip
Bert-Base-Uncased-Hf	float	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN_DLC	22.194 ms	0 - 0 MB	NPU	Bert-Base-Uncased-Hf.dlc
Bert-Base-Uncased-Hf	float	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	31.902 ms	265 - 265 MB	NPU	Bert-Base-Uncased-Hf.onnx.zip

Installation

Install the package via pip:

pip install qai-hub-models

Configure Qualcomm® AI Hub Workbench to run this model on a cloud-hosted device

Sign-in to Qualcomm® AI Hub Workbench with your Qualcomm® ID. Once signed in navigate to Account -> Settings -> API Token.

With this API token, you can configure your client to run models on the cloud hosted devices.

qai-hub configure --api_token API_TOKEN

Navigate to docs for more information.

Demo off target

The package contains a simple end-to-end demo that downloads pre-trained weights and runs this model on a sample input.

python -m qai_hub_models.models.bert_base_uncased_hf.demo

The above demo runs a reference implementation of pre-processing, model inference, and post processing.

NOTE: If you want running in a Jupyter Notebook or Google Colab like environment, please add the following to your cell (instead of the above).

%run -m qai_hub_models.models.bert_base_uncased_hf.demo

Run model on a cloud-hosted device

In addition to the demo, you can also run the model on a cloud-hosted Qualcomm® device. This script does the following:

• Performance check on-device on a cloud-hosted device
• Downloads compiled assets that can be deployed on-device for Android.
• Accuracy check between PyTorch and on-device outputs.

python -m qai_hub_models.models.bert_base_uncased_hf.export

How does this work?

This export script leverages Qualcomm® AI Hub to optimize, validate, and deploy this model on-device. Lets go through each step below in detail:

Step 1: Compile model for on-device deployment

To compile a PyTorch model for on-device deployment, we first trace the model in memory using the jit.trace and then call the submit_compile_job API.

import torch

import qai_hub as hub
from qai_hub_models.models.bert_base_uncased_hf import Model

# Load the model
torch_model = Model.from_pretrained()

# Device
device = hub.Device("Samsung Galaxy S25")

# Trace model
input_shape = torch_model.get_input_spec()
sample_inputs = torch_model.sample_inputs()

pt_model = torch.jit.trace(torch_model, [torch.tensor(data[0]) for _, data in sample_inputs.items()])

# Compile model on a specific device
compile_job = hub.submit_compile_job(
    model=pt_model,
    device=device,
    input_specs=torch_model.get_input_spec(),
)

# Get target model to run on-device
target_model = compile_job.get_target_model()

Step 2: Performance profiling on cloud-hosted device

After compiling models from step 1. Models can be profiled model on-device using the target_model. Note that this scripts runs the model on a device automatically provisioned in the cloud. Once the job is submitted, you can navigate to a provided job URL to view a variety of on-device performance metrics.

profile_job = hub.submit_profile_job(
    model=target_model,
    device=device,
)
        

Step 3: Verify on-device accuracy

To verify the accuracy of the model on-device, you can run on-device inference on sample input data on the same cloud hosted device.

input_data = torch_model.sample_inputs()
inference_job = hub.submit_inference_job(
    model=target_model,
    device=device,
    inputs=input_data,
)
    on_device_output = inference_job.download_output_data()

With the output of the model, you can compute like PSNR, relative errors or spot check the output with expected output.

Note: This on-device profiling and inference requires access to Qualcomm® AI Hub Workbench. Sign up for access.

Run demo on a cloud-hosted device

You can also run the demo on-device.

python -m qai_hub_models.models.bert_base_uncased_hf.demo --eval-mode on-device

NOTE: If you want running in a Jupyter Notebook or Google Colab like environment, please add the following to your cell (instead of the above).

%run -m qai_hub_models.models.bert_base_uncased_hf.demo -- --eval-mode on-device

Deploying compiled model to Android

The models can be deployed using multiple runtimes:

• TensorFlow Lite (.tflite export): This tutorial provides a guide to deploy the .tflite model in an Android application.

• QNN (.so export ): This sample app provides instructions on how to use the .so shared library in an Android application.

View on Qualcomm® AI Hub

Get more details on Bert-Base-Uncased-Hf's performance across various devices here. Explore all available models on Qualcomm® AI Hub

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.