Owlite

Latest version: v0.0.8

Safety actively analyzes 682334 Python packages for vulnerabilities to keep your Python projects secure.

Scan your dependencies

Page 1 of 2

2.2.0

![image](https://cdn.owlite.ai/release240926/220_1.png)

Key Updates:

- Removed Dependencies: The dependency on onnx-simplifier and onnx-graphsurgeon for ONNX-level optimization has been removed.

- Expanded Compatibility: Added support for PyTorch versions 2.1, 2.2, 2.3 and 2.4 and Python versions 3.10, 3.11 and 3.12.

- TensorRT Update: Starting with this version, we strongly recommend using the latest TensorRT version (currently 10.2) for the runner to ensure compatibility and optimal performance.

Minor Changes:

- Calibration in no_grad Mode: Calibration now runs in no_grad mode by default. If benchmarking without no_grad is required, adjust the setting by modifying OWLITE_CALIBRATION_ENABLE_GRAD.

- Minor bug fixes and UX improvements.

Impact on Users:

- Please use pre-built wheel for installation to avoid potential issues with dependencies and setup.

2.1.0

**Enhanced Latency Bottleneck Visualization - Node Breakdown**


<figure><img src=https://cdn.owlite.ai/release240612/2.1.0_1.png alt=><figcaption></figcaption></figure>

The Node Breakdown feature has been significantly improved to help you easily identify latency bottlenecks. Enhancing the visualization of the matching TensorRT benchmark results on ONNX graphs makes pinpointing high-latency areas based on operator colors simpler. The stronger the color, the higher the latency.



**Addition of New Data Type for Optimization: FP8**



<figure><img src=https://cdn.owlite.ai/release240612/2.1.0_2.png alt=><figcaption></figcaption></figure>

A new data type, FP8, has been added for . This aims to achieve possible accuracy gains by reducing quantization error. You can now simulate quantization using the fp8\_e4m3 data type. The FP8 quantization supports STE as QAT backward and Minmax and percentile as PTQ calibration.

* The **precision** parameter has been replaced with the **data type** parameter.
* The former **precision : 8** is now updated to **data type : int8**.
* The **unsigned toggle** has been removed.
* A new **data type uint8 (unsigned int8)** has been added.

**Note:** This feature is currently highly unstable and may yield unexpected results. Also, please note that FP8 benchmarks are not supported on the free plan.



**Updated OwLite Support Channels**



The OwLite support system has been updated for a smoother experience. You can now search through the OwLite documentation or help center if you have any inquiries. The system has been improved to provide a more seamless inquiry and response experience.

2.0.0

Hello! In this update, we are excited to introduce collaborative features, the ability to skip benchmark polling, and output node selection. Check out the details below to learn more about these new features.



**1. Collaborative Features**

***

<figure><img src="https://cdn.owlite.ai/release240523/project_share.png" alt=""><figcaption></figcaption></figure>

Work together with your colleagues more efficiently. With our collaborative features, you can now share the compression process and achieve the best results faster. When creating a project or editing settings, you can invite users within the same workgroup by entering their username.

* **Invite Colleagues:** Enter the username to invite users within your workgroup when setting up a project.
* **Collaborative Optimization:** Work together on the same project and compare optimization results in one place.
* **Note:** You cannot edit experiments owned by others, but you can duplicate them and start a new experiment once the collaborator finishes their work.
* **Availability:** This feature is available for users on the Lite plan and above.



**2. Skip Benchmark Polling**

***

Optimal optimization requires numerous trials and errors. Multiple tests and benchmarks need to be performed to achieve the desired optimization results, but waiting for code execution each time can be time-consuming. To solve this, you can add specific content to the benchmark function, allowing it to run on the server while you perform other tasks.

python
owl.benchmark(download_engine=False)


* **Feature Addition:** Add content to the benchmark function to run it on the server.
* **Note:** To receive the TensorRT engine, you will need to remove the added content and rerun the code.
* **Usage Scenario:** Perform multiple configs consecutively and rerun only the optimal experiment to receive the TensorRT engine.



**3. Output Node Selection**

***

<figure><img src="https://cdn.owlite.ai/release240523/output_quantization.png" alt=""><figcaption></figcaption></figure>

You can now quantize the output node of the AI model if needed. While the recommended settings are sufficient for general use cases, we offer various quantization options to help you find the perfect optimization for your model, reflecting our philosophy of providing tailored solutions.

* **Feature Description:** Quantize the output node of your AI model if needed.



We hope these new features enhance your workflow and make your tasks more efficient and convenient. If you have any additional needs or suggestions, please feel free to let us know. Thank you!

1.2.4

Updates
- Please note that there are no significant changes to the Python package in this update.
- For a detailed overview of the updates across the entire OwLite service, please refer to the link provided below. Thank you!
- https://squeezebits.gitbook.io/owlite/updates/1.2.4-update-22nd-apr-2024

1.2.3

Error logging for user convenience
* Forwards TensorRT engine build failure messages to users
* Explicitly warns users when an experiment is created on a different device from the baseline
* Raises an exception when an experiment has a different input signature from the baseline

1.2.2

[HOTFIX] Version 1.2.2 Update (19th Mar 2024)

Bugfix
- Fixed issues with torch.nn.DataParallel compatibility
- Fixed issues with Dropout which lead to TensorRT engine compilation failure

Page 1 of 2

© 2024 Safety CLI Cybersecurity Inc. All Rights Reserved.