VIT: Optimized for Qualcomm Devices
VIT 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 VIT 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 2.45, ONNX Runtime 1.25.0 | Download |
| ONNX | w8a16 | Universal | QAIRT 2.45, ONNX Runtime 1.25.0 | Download |
| ONNX | w8a8 | Universal | QAIRT 2.45, ONNX Runtime 1.25.0 | Download |
| ONNX | w8a8_mixed_int16 | 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 |
| QNN_DLC | w8a8 | Universal | QAIRT 2.45 | Download |
| TFLITE | float | Universal | QAIRT 2.45 | Download |
| TFLITE | w8a8 | Universal | QAIRT 2.45 | Download |
For more device-specific assets and performance metrics, visit VIT 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 VIT 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: 86.6M
- Model size (float): 330 MB
- Model size (w8a16): 86.2 MB
- Model size (w8a8): 83.2 MB
Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit |
|---|---|---|---|---|---|---|
| VIT | ONNX | float | Snapdragon® X2 Elite | 4.221 ms | 2 - 2 MB | NPU |
| VIT | ONNX | float | Snapdragon® X Elite | 7.523 ms | 171 - 171 MB | NPU |
| VIT | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 5.019 ms | 0 - 304 MB | NPU |
| VIT | ONNX | float | Snapdragon® 8 Gen 1 Mobile | 12.105 ms | 1 - 283 MB | NPU |
| VIT | ONNX | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 7.15 ms | 0 - 194 MB | NPU |
| VIT | ONNX | float | Qualcomm® QCS8450 | 12.105 ms | 1 - 283 MB | NPU |
| VIT | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 2.819 ms | 0 - 163 MB | NPU |
| VIT | ONNX | float | Snapdragon® 8 Elite Mobile | 3.447 ms | 0 - 175 MB | NPU |
| VIT | ONNX | float | Qualcomm® Dragonwing™ IQ-9075 | 9.532 ms | 0 - 4 MB | NPU |
| VIT | ONNX | float | Qualcomm® Dragonwing™ Q-8750 | 3.447 ms | 0 - 175 MB | NPU |
| VIT | ONNX | float | Qualcomm® Dragonwing™ IQ-X7181 | 7.523 ms | 171 - 171 MB | NPU |
| VIT | ONNX | w8a16 | Snapdragon® X2 Elite | 2.007 ms | 1 - 1 MB | NPU |
| VIT | ONNX | w8a16 | Snapdragon® X Elite | 4.743 ms | 86 - 86 MB | NPU |
| VIT | ONNX | w8a16 | Snapdragon® 8 Gen 3 Mobile | 3.096 ms | 0 - 314 MB | NPU |
| VIT | ONNX | w8a16 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 4.583 ms | 0 - 99 MB | NPU |
| VIT | ONNX | w8a16 | Snapdragon® 8 Elite Mobile | 2.42 ms | 0 - 261 MB | NPU |
| VIT | ONNX | w8a16 | Qualcomm® Dragonwing™ IQ-9075 | 4.457 ms | 0 - 3 MB | NPU |
| VIT | ONNX | w8a16 | Snapdragon® 7 Gen 4 Mobile | 5.853 ms | 0 - 397 MB | NPU |
| VIT | ONNX | w8a16 | Snapdragon® 8 Elite Gen 5 Mobile | 1.801 ms | 0 - 271 MB | NPU |
| VIT | ONNX | w8a16 | Qualcomm® Dragonwing™ Q-6690 | 50.452 ms | 0 - 404 MB | NPU |
| VIT | ONNX | w8a16 | Qualcomm® Dragonwing™ Q-7790 | 5.853 ms | 0 - 397 MB | NPU |
| VIT | ONNX | w8a16 | Qualcomm® Dragonwing™ Q-8750 | 2.42 ms | 0 - 261 MB | NPU |
| VIT | ONNX | w8a16 | Qualcomm® Dragonwing™ IQ-X7181 | 4.743 ms | 86 - 86 MB | NPU |
| VIT | ONNX | w8a8 | Snapdragon® X2 Elite | 4.119 ms | 1 - 1 MB | NPU |
| VIT | ONNX | w8a8 | Snapdragon® X Elite | 9.76 ms | 86 - 86 MB | NPU |
| VIT | ONNX | w8a8 | Snapdragon® 8 Gen 3 Mobile | 6.254 ms | 0 - 334 MB | NPU |
| VIT | ONNX | w8a8 | Snapdragon® 8 Gen 1 Mobile | 12.728 ms | 0 - 341 MB | NPU |
| VIT | ONNX | w8a8 | Qualcomm® Dragonwing™ QCS6490 | 61.245 ms | 0 - 3 MB | NPU |
| VIT | ONNX | w8a8 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 9.354 ms | 0 - 99 MB | NPU |
| VIT | ONNX | w8a8 | Qualcomm® QCS8450 | 12.728 ms | 0 - 341 MB | NPU |
| VIT | ONNX | w8a8 | Snapdragon® 8 Elite Mobile | 5.41 ms | 0 - 239 MB | NPU |
| VIT | ONNX | w8a8 | Snapdragon® 7 Gen 4 Mobile | 16.827 ms | 0 - 394 MB | NPU |
| VIT | ONNX | w8a8 | Qualcomm® Dragonwing™ IQ-9075 | 8.129 ms | 0 - 3 MB | NPU |
| VIT | ONNX | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 3.52 ms | 0 - 246 MB | NPU |
| VIT | ONNX | w8a8 | Qualcomm® Dragonwing™ Q-6690 | 209.772 ms | 0 - 529 MB | NPU |
| VIT | ONNX | w8a8 | Qualcomm® Dragonwing™ Q-7790 | 16.827 ms | 0 - 394 MB | NPU |
| VIT | ONNX | w8a8 | Qualcomm® Dragonwing™ Q-8750 | 5.41 ms | 0 - 239 MB | NPU |
| VIT | ONNX | w8a8 | Qualcomm® Dragonwing™ IQ-X7181 | 9.76 ms | 86 - 86 MB | NPU |
| VIT | ONNX | w8a8_mixed_int16 | Snapdragon® 8 Gen 3 Mobile | 70.995 ms | 48 - 409 MB | NPU |
| VIT | ONNX | w8a8_mixed_int16 | Snapdragon® 8 Elite Mobile | 55.466 ms | 30 - 295 MB | NPU |
| VIT | ONNX | w8a8_mixed_int16 | Snapdragon® 8 Elite Gen 5 Mobile | 48.767 ms | 58 - 334 MB | NPU |
| VIT | ONNX | w8a8_mixed_int16 | Qualcomm® Dragonwing™ Q-8750 | 55.466 ms | 30 - 295 MB | NPU |
| VIT | QNN_DLC | float | Snapdragon® X2 Elite | 3.545 ms | 1 - 1 MB | NPU |
| VIT | QNN_DLC | float | Snapdragon® X Elite | 8.304 ms | 1 - 1 MB | NPU |
| VIT | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 5.4 ms | 0 - 250 MB | NPU |
| VIT | QNN_DLC | float | Snapdragon® 8 Gen 1 Mobile | 13.117 ms | 0 - 229 MB | NPU |
| VIT | QNN_DLC | float | Qualcomm® QCS8275 | 34.797 ms | 1 - 167 MB | NPU |
| VIT | QNN_DLC | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 7.667 ms | 1 - 2 MB | NPU |
| VIT | QNN_DLC | float | Qualcomm® SA8775P | 10.497 ms | 1 - 167 MB | NPU |
| VIT | QNN_DLC | float | Qualcomm® SA8650P | 10.497 ms | 1 - 167 MB | NPU |
| VIT | QNN_DLC | float | Qualcomm® SA8255P | 10.497 ms | 1 - 167 MB | NPU |
| VIT | QNN_DLC | float | Qualcomm® QCS8450 | 13.117 ms | 0 - 229 MB | NPU |
| VIT | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 2.94 ms | 1 - 157 MB | NPU |
| VIT | QNN_DLC | float | Snapdragon® 8 Elite Mobile | 3.76 ms | 1 - 156 MB | NPU |
| VIT | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-9075 | 10.169 ms | 3 - 5 MB | NPU |
| VIT | QNN_DLC | float | Qualcomm® SA7255P | 34.797 ms | 1 - 167 MB | NPU |
| VIT | QNN_DLC | float | Qualcomm® SA8295P | 12.937 ms | 1 - 144 MB | NPU |
| VIT | QNN_DLC | float | Qualcomm® Dragonwing™ Q-8750 | 3.76 ms | 1 - 156 MB | NPU |
| VIT | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-X7181 | 8.304 ms | 1 - 1 MB | NPU |
| VIT | QNN_DLC | w8a16 | Snapdragon® X2 Elite | 3.596 ms | 0 - 0 MB | NPU |
| VIT | QNN_DLC | w8a16 | Snapdragon® X Elite | 8.211 ms | 0 - 0 MB | NPU |
| VIT | QNN_DLC | w8a16 | Snapdragon® 8 Gen 3 Mobile | 5.095 ms | 0 - 319 MB | NPU |
| VIT | QNN_DLC | w8a16 | Qualcomm® QCS8275 | 17.35 ms | 0 - 258 MB | NPU |
| VIT | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 7.621 ms | 0 - 2 MB | NPU |
| VIT | QNN_DLC | w8a16 | Qualcomm® SA8775P | 7.808 ms | 0 - 259 MB | NPU |
| VIT | QNN_DLC | w8a16 | Qualcomm® SA8650P | 7.808 ms | 0 - 259 MB | NPU |
| VIT | QNN_DLC | w8a16 | Qualcomm® SA8255P | 7.808 ms | 0 - 259 MB | NPU |
| VIT | QNN_DLC | w8a16 | Snapdragon® 8 Elite Mobile | 3.902 ms | 0 - 259 MB | NPU |
| VIT | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ IQ-9075 | 7.58 ms | 0 - 2 MB | NPU |
| VIT | QNN_DLC | w8a16 | Qualcomm® SA7255P | 17.35 ms | 0 - 258 MB | NPU |
| VIT | QNN_DLC | w8a16 | Snapdragon® 7 Gen 4 Mobile | 12.914 ms | 0 - 403 MB | NPU |
| VIT | QNN_DLC | w8a16 | Snapdragon® 8 Elite Gen 5 Mobile | 2.853 ms | 0 - 279 MB | NPU |
| VIT | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ Q-6690 | 98.808 ms | 1 - 404 MB | NPU |
| VIT | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ Q-7790 | 12.914 ms | 0 - 403 MB | NPU |
| VIT | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ Q-8750 | 3.902 ms | 0 - 259 MB | NPU |
| VIT | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ IQ-X7181 | 8.211 ms | 0 - 0 MB | NPU |
| VIT | QNN_DLC | w8a8 | Snapdragon® X2 Elite | 4.526 ms | 0 - 0 MB | NPU |
| VIT | QNN_DLC | w8a8 | Snapdragon® X Elite | 10.211 ms | 0 - 0 MB | NPU |
| VIT | QNN_DLC | w8a8 | Snapdragon® 8 Gen 3 Mobile | 6.521 ms | 0 - 306 MB | NPU |
| VIT | QNN_DLC | w8a8 | Snapdragon® 8 Gen 1 Mobile | 13.185 ms | 0 - 313 MB | NPU |
| VIT | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ QCS6490 | 62.067 ms | 0 - 2 MB | NPU |
| VIT | QNN_DLC | w8a8 | Qualcomm® QCS8275 | 28.468 ms | 0 - 202 MB | NPU |
| VIT | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 9.563 ms | 0 - 208 MB | NPU |
| VIT | QNN_DLC | w8a8 | Qualcomm® SA8775P | 8.648 ms | 0 - 202 MB | NPU |
| VIT | QNN_DLC | w8a8 | Qualcomm® SA8650P | 8.648 ms | 0 - 202 MB | NPU |
| VIT | QNN_DLC | w8a8 | Qualcomm® SA8255P | 8.648 ms | 0 - 202 MB | NPU |
| VIT | QNN_DLC | w8a8 | Qualcomm® QCS8450 | 13.185 ms | 0 - 313 MB | NPU |
| VIT | QNN_DLC | w8a8 | Snapdragon® 8 Elite Mobile | 5.369 ms | 0 - 203 MB | NPU |
| VIT | QNN_DLC | w8a8 | Snapdragon® 7 Gen 4 Mobile | 19.295 ms | 0 - 355 MB | NPU |
| VIT | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ IQ-9075 | 8.386 ms | 0 - 2 MB | NPU |
| VIT | QNN_DLC | w8a8 | Qualcomm® SA7255P | 28.468 ms | 0 - 202 MB | NPU |
| VIT | QNN_DLC | w8a8 | Qualcomm® SA8295P | 15.6 ms | 0 - 212 MB | NPU |
| VIT | QNN_DLC | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 3.55 ms | 0 - 210 MB | NPU |
| VIT | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ Q-6690 | 221.454 ms | 2 - 498 MB | NPU |
| VIT | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ Q-7790 | 19.295 ms | 0 - 355 MB | NPU |
| VIT | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ Q-8750 | 5.369 ms | 0 - 203 MB | NPU |
| VIT | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ IQ-X7181 | 10.211 ms | 0 - 0 MB | NPU |
| VIT | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 5.274 ms | 0 - 306 MB | NPU |
| VIT | TFLITE | float | Snapdragon® 8 Gen 1 Mobile | 12.929 ms | 0 - 289 MB | NPU |
| VIT | TFLITE | float | Qualcomm® QCS8275 | 34.592 ms | 0 - 168 MB | NPU |
| VIT | TFLITE | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 7.24 ms | 0 - 4 MB | NPU |
| VIT | TFLITE | float | Qualcomm® SA8775P | 10.177 ms | 0 - 184 MB | NPU |
| VIT | TFLITE | float | Qualcomm® SA8650P | 10.177 ms | 0 - 184 MB | NPU |
| VIT | TFLITE | float | Qualcomm® SA8255P | 10.177 ms | 0 - 184 MB | NPU |
| VIT | TFLITE | float | Qualcomm® QCS8450 | 12.929 ms | 0 - 289 MB | NPU |
| VIT | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 3.043 ms | 0 - 170 MB | NPU |
| VIT | TFLITE | float | Snapdragon® 8 Elite Mobile | 3.65 ms | 0 - 178 MB | NPU |
| VIT | TFLITE | float | Qualcomm® Dragonwing™ IQ-9075 | 9.839 ms | 0 - 174 MB | NPU |
| VIT | TFLITE | float | Qualcomm® SA7255P | 34.592 ms | 0 - 168 MB | NPU |
| VIT | TFLITE | float | Qualcomm® SA8295P | 13.3 ms | 0 - 156 MB | NPU |
| VIT | TFLITE | float | Qualcomm® Dragonwing™ Q-8750 | 3.65 ms | 0 - 178 MB | NPU |
| VIT | TFLITE | w8a8 | Snapdragon® 8 Gen 3 Mobile | 8.664 ms | 0 - 487 MB | NPU |
| VIT | TFLITE | w8a8 | Snapdragon® 8 Gen 1 Mobile | 20.579 ms | 0 - 446 MB | NPU |
| VIT | TFLITE | w8a8 | Qualcomm® Dragonwing™ QCS6490 | 146.212 ms | 1 - 100 MB | NPU |
| VIT | TFLITE | w8a8 | Qualcomm® QCS8275 | 35.048 ms | 0 - 385 MB | NPU |
| VIT | TFLITE | w8a8 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 12.35 ms | 0 - 3 MB | NPU |
| VIT | TFLITE | w8a8 | Qualcomm® SA8775P | 12.213 ms | 0 - 388 MB | NPU |
| VIT | TFLITE | w8a8 | Qualcomm® SA8650P | 12.213 ms | 0 - 388 MB | NPU |
| VIT | TFLITE | w8a8 | Qualcomm® SA8255P | 12.213 ms | 0 - 388 MB | NPU |
| VIT | TFLITE | w8a8 | Qualcomm® QCS8450 | 20.579 ms | 0 - 446 MB | NPU |
| VIT | TFLITE | w8a8 | Snapdragon® 8 Elite Mobile | 6.925 ms | 0 - 380 MB | NPU |
| VIT | TFLITE | w8a8 | Snapdragon® 7 Gen 4 Mobile | 29.299 ms | 0 - 213 MB | NPU |
| VIT | TFLITE | w8a8 | Qualcomm® Dragonwing™ IQ-9075 | 11.938 ms | 0 - 88 MB | NPU |
| VIT | TFLITE | w8a8 | Qualcomm® SA7255P | 35.048 ms | 0 - 385 MB | NPU |
| VIT | TFLITE | w8a8 | Qualcomm® SA8295P | 19.23 ms | 0 - 346 MB | NPU |
| VIT | TFLITE | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 4.586 ms | 0 - 384 MB | NPU |
| VIT | TFLITE | w8a8 | Qualcomm® Dragonwing™ Q-6690 | 198.429 ms | 2 - 283 MB | NPU |
| VIT | TFLITE | w8a8 | Qualcomm® Dragonwing™ Q-7790 | 29.299 ms | 0 - 213 MB | NPU |
| VIT | TFLITE | w8a8 | Qualcomm® Dragonwing™ Q-8750 | 6.925 ms | 0 - 380 MB | NPU |
License
- The license for the original implementation of VIT can be found here.
References
- An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
- 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.
