Hugging Face's logo

qualcomm
/
OpusMT-Zh-En

Text Generation
PyTorch
LiteRT
foundation
android
Model card Files
xet
Community

OpusMT-Zh-En: Optimized for Mobile Deployment

OpusMT Chinese to English neural machine translation model based on MarianMT transformer architecture

OpusMT Chinese to English translation model is a state-of-the-art neural machine translation system designed for translating Chinese text into English. This model is based on the Marian transformer architecture and has been optimized for edge inference by splitting into encoder and decoder components with modified attention mechanisms. It exhibits robust performance for real-world translation tasks, making it highly reliable for practical applications. The model supports input sequences up to 256 tokens and can generate English translations with high accuracy.

This model is an implementation of OpusMT-Zh-En found here.

This repository provides scripts to run OpusMT-Zh-En 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: Helsinki-NLP/opus-mt-zh-en
    • Input resolution: 256 tokens (Chinese text)
    • Max input sequence length: 256 tokens
    • Max output sequence length: 256 tokens
    • Number of parameters (OpusMTEncoder): ~74M
    • Model size (OpusMTEncoder) (float): ~280 MB
    • Number of parameters (OpusMTDecoder): ~74M
    • Model size (OpusMTDecoder) (float): ~280 MB
    • Number of encoder layers: 6
    • Number of decoder layers: 6
    • Attention heads: 8
    • Hidden dimension: 512
Model Precision Device Chipset Target Runtime Inference Time (ms) Peak Memory Range (MB) Primary Compute Unit Target Model
OpusMTEncoder float QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) TFLITE 12.804 ms 6 - 181 MB NPU OpusMT-Zh-En.tflite
OpusMTEncoder float QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) QNN_DLC 12.516 ms 0 - 139 MB NPU OpusMT-Zh-En.dlc
OpusMTEncoder float QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) TFLITE 5.108 ms 6 - 340 MB NPU OpusMT-Zh-En.tflite
OpusMTEncoder float QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) QNN_DLC 4.741 ms 0 - 171 MB NPU OpusMT-Zh-En.dlc
OpusMTEncoder float QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) TFLITE 3.63 ms 0 - 2 MB NPU OpusMT-Zh-En.tflite
OpusMTEncoder float QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) QNN_DLC 3.526 ms 0 - 2 MB NPU OpusMT-Zh-En.dlc
OpusMTEncoder float QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) ONNX 5.731 ms 0 - 114 MB NPU OpusMT-Zh-En.onnx.zip
OpusMTEncoder float QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) TFLITE 4.877 ms 6 - 181 MB NPU OpusMT-Zh-En.tflite
OpusMTEncoder float QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) QNN_DLC 19.322 ms 0 - 139 MB NPU OpusMT-Zh-En.dlc
OpusMTEncoder float SA7255P ADP Qualcomm® SA7255P TFLITE 12.804 ms 6 - 181 MB NPU OpusMT-Zh-En.tflite
OpusMTEncoder float SA7255P ADP Qualcomm® SA7255P QNN_DLC 12.516 ms 0 - 139 MB NPU OpusMT-Zh-En.dlc
OpusMTEncoder float SA8295P ADP Qualcomm® SA8295P TFLITE 5.525 ms 6 - 180 MB NPU OpusMT-Zh-En.tflite
OpusMTEncoder float SA8295P ADP Qualcomm® SA8295P QNN_DLC 5.262 ms 0 - 139 MB NPU OpusMT-Zh-En.dlc
OpusMTEncoder float SA8775P ADP Qualcomm® SA8775P TFLITE 4.877 ms 6 - 181 MB NPU OpusMT-Zh-En.tflite
OpusMTEncoder float SA8775P ADP Qualcomm® SA8775P QNN_DLC 19.322 ms 0 - 139 MB NPU OpusMT-Zh-En.dlc
OpusMTEncoder float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile TFLITE 2.667 ms 0 - 341 MB NPU OpusMT-Zh-En.tflite
OpusMTEncoder float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile QNN_DLC 2.562 ms 0 - 174 MB NPU OpusMT-Zh-En.dlc
OpusMTEncoder float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile ONNX 4.141 ms 15 - 329 MB NPU OpusMT-Zh-En.onnx.zip
OpusMTEncoder float Samsung Galaxy S25 Snapdragon® 8 Elite For Galaxy Mobile TFLITE 2.257 ms 0 - 355 MB NPU OpusMT-Zh-En.tflite
OpusMTEncoder float Samsung Galaxy S25 Snapdragon® 8 Elite For Galaxy Mobile QNN_DLC 2.12 ms 0 - 142 MB NPU OpusMT-Zh-En.dlc
OpusMTEncoder float Samsung Galaxy S25 Snapdragon® 8 Elite For Galaxy Mobile ONNX 3.556 ms 0 - 335 MB NPU OpusMT-Zh-En.onnx.zip
OpusMTEncoder float Snapdragon 8 Elite Gen 5 QRD Snapdragon® 8 Elite Gen 5 Mobile TFLITE 1.763 ms 0 - 314 MB NPU OpusMT-Zh-En.tflite
OpusMTEncoder float Snapdragon 8 Elite Gen 5 QRD Snapdragon® 8 Elite Gen 5 Mobile QNN_DLC 1.621 ms 0 - 146 MB NPU OpusMT-Zh-En.dlc
OpusMTEncoder float Snapdragon 8 Elite Gen 5 QRD Snapdragon® 8 Elite Gen 5 Mobile ONNX 2.843 ms 0 - 298 MB NPU OpusMT-Zh-En.onnx.zip
OpusMTEncoder float Snapdragon X Elite CRD Snapdragon® X Elite QNN_DLC 3.855 ms 0 - 0 MB NPU OpusMT-Zh-En.dlc
OpusMTEncoder float Snapdragon X Elite CRD Snapdragon® X Elite ONNX 5.801 ms 108 - 108 MB NPU OpusMT-Zh-En.onnx.zip
OpusMTDecoder float QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) TFLITE 6.408 ms 0 - 425 MB NPU OpusMT-Zh-En.tflite
OpusMTDecoder float QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) QNN_DLC 5.386 ms 6 - 242 MB NPU OpusMT-Zh-En.dlc
OpusMTDecoder float QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) TFLITE 4.715 ms 0 - 426 MB NPU OpusMT-Zh-En.tflite
OpusMTDecoder float QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) QNN_DLC 3.999 ms 2 - 246 MB NPU OpusMT-Zh-En.dlc
OpusMTDecoder float QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) TFLITE 3.485 ms 0 - 4 MB NPU OpusMT-Zh-En.tflite
OpusMTDecoder float QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) QNN_DLC 2.927 ms 1 - 402 MB NPU OpusMT-Zh-En.dlc
OpusMTDecoder float QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) ONNX 3.846 ms 12 - 15 MB NPU OpusMT-Zh-En.onnx.zip
OpusMTDecoder float QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) TFLITE 4.142 ms 0 - 292 MB NPU OpusMT-Zh-En.tflite
OpusMTDecoder float QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) QNN_DLC 3.503 ms 6 - 243 MB NPU OpusMT-Zh-En.dlc
OpusMTDecoder float SA7255P ADP Qualcomm® SA7255P TFLITE 6.408 ms 0 - 425 MB NPU OpusMT-Zh-En.tflite
OpusMTDecoder float SA7255P ADP Qualcomm® SA7255P QNN_DLC 5.386 ms 6 - 242 MB NPU OpusMT-Zh-En.dlc
OpusMTDecoder float SA8295P ADP Qualcomm® SA8295P TFLITE 4.782 ms 0 - 285 MB NPU OpusMT-Zh-En.tflite
OpusMTDecoder float SA8295P ADP Qualcomm® SA8295P QNN_DLC 4.095 ms 6 - 234 MB NPU OpusMT-Zh-En.dlc
OpusMTDecoder float SA8775P ADP Qualcomm® SA8775P TFLITE 4.142 ms 0 - 292 MB NPU OpusMT-Zh-En.tflite
OpusMTDecoder float SA8775P ADP Qualcomm® SA8775P QNN_DLC 3.503 ms 6 - 243 MB NPU OpusMT-Zh-En.dlc
OpusMTDecoder float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile TFLITE 2.622 ms 0 - 575 MB NPU OpusMT-Zh-En.tflite
OpusMTDecoder float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile QNN_DLC 2.216 ms 0 - 259 MB NPU OpusMT-Zh-En.dlc
OpusMTDecoder float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile ONNX 2.942 ms 0 - 416 MB NPU OpusMT-Zh-En.onnx.zip
OpusMTDecoder float Samsung Galaxy S25 Snapdragon® 8 Elite For Galaxy Mobile TFLITE 2.334 ms 0 - 484 MB NPU OpusMT-Zh-En.tflite
OpusMTDecoder float Samsung Galaxy S25 Snapdragon® 8 Elite For Galaxy Mobile QNN_DLC 1.964 ms 0 - 234 MB NPU OpusMT-Zh-En.dlc
OpusMTDecoder float Samsung Galaxy S25 Snapdragon® 8 Elite For Galaxy Mobile ONNX 2.573 ms 0 - 443 MB NPU OpusMT-Zh-En.onnx.zip
OpusMTDecoder float Snapdragon 8 Elite Gen 5 QRD Snapdragon® 8 Elite Gen 5 Mobile TFLITE 2.156 ms 0 - 433 MB NPU OpusMT-Zh-En.tflite
OpusMTDecoder float Snapdragon 8 Elite Gen 5 QRD Snapdragon® 8 Elite Gen 5 Mobile QNN_DLC 1.884 ms 0 - 242 MB NPU OpusMT-Zh-En.dlc
OpusMTDecoder float Snapdragon 8 Elite Gen 5 QRD Snapdragon® 8 Elite Gen 5 Mobile ONNX 2.386 ms 1 - 406 MB NPU OpusMT-Zh-En.onnx.zip
OpusMTDecoder float Snapdragon X Elite CRD Snapdragon® X Elite QNN_DLC 2.755 ms 6 - 6 MB NPU OpusMT-Zh-En.dlc
OpusMTDecoder float Snapdragon X Elite CRD Snapdragon® X Elite ONNX 3.17 ms 161 - 161 MB NPU OpusMT-Zh-En.onnx.zip

Installation

Install the package via pip:

# NOTE: 3.10 <= PYTHON_VERSION < 3.14 is supported.
pip install "qai-hub-models[opus-mt-zh-en]"

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.opus_mt_zh_en.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.opus_mt_zh_en.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.opus_mt_zh_en.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.opus_mt_zh_en 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.

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 OpusMT-Zh-En's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

  • The license for the original implementation of OpusMT-Zh-En can be found here.

References

Community

Downloads last month
9
Inference Providers NEW
Text Generation
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support