Release Notes¶

June 23, 2025 - LEIP Optimize 5.0¶

License Key Format Changes: Users can now place their LEIP license keys in ~/.leip/license.key or a .leip/license.key file in their project directory. If you prefer to use an environment variable, set it using LEIP_LICENSE_KEY.
TVM-TRT Deprecation: The TVM-TRT backend is now deprecated. Users targeting NVIDIA TensorRT devices should use the ONNX backend.
TensorRT 10 Support: LEIP Optimize now supports NVIDIA TensorRT 10, providing broader operator coverage and improved performance.
Jetpack 6.2 Support: LEIP Optimize users can now target NVIDIA Jetson devices running Jetpack 6.2. Support for Jetpack 5.1 and 6.0 has now been deprecated. Users running older versions of Jetpack should use LEIP Optimize 4.2.
Raspberry Pi 5 Support: LEIP Optimize now targets Raspberry Pi 5 devices.
Expanded Python Support: Added support for Python 3.12 and 3.13.
Usage and Diagnostic Reporting: LEIP Optimize may collect usage metrics and diagnostic reports to help improve product quality and performance.

Jetpack 6.0 Support: LEIP Optimize users can now target NVIDIA Jetson devices with Jetpack 6.0.

Android Support: LEIP Optimize users can now target Android CPU devices. The included Docker image comes preconfigured with the Android cross compiler.
Pre-quantized Model Ingestion: Users can now ingest and compile pre-quantized models.
New Tag Search API: Added forge.search_target_tags(), enabling users to search for target tags containing the specified query string (case-insensitive).

ONNX Dynamic Quantization: Resolved an issue where dynamic quantization incorrectly required calibration.

ONNX export: Fixed an issue with TensorRT ONNX export.
ONNX Model Precision: Ensured correct model precision settings for ONNX-exported models.

Compiler Framework is now LEIP Optimize.
Introduces Forge: A powerful tool that allows you to easily edit, manipulate, and introspect model compute graphs in an Intermediate Representation. Forge is capable of ingesting models from a wide variety of sources and compiling and optimizing them for a diverse set of target hardware.
ONNX Integration: Users can now optimize ONNX models with the option of maintaining the ONNX graph representation.
Model Support: LEIP Optimize now supports a much broader set of models; for example, it now supports 90% of the top 50 computer vision models from Hugging Face.
Pip Install: LEIP Optimize is now available for installation via pip.