OnnxSlim is a toolkit designed to optimize ONNX models by reducing operator count, aiming to improve inference speed while preserving accuracy. It targets developers and researchers seeking to enhance the performance of their ONNX models for deployment across various platforms.

How It Works

OnnxSlim employs graph simplification techniques to reduce the complexity of ONNX models. Its core approach involves identifying and eliminating redundant or equivalent operators within the model graph. This process results in a more streamlined model representation, which directly translates to faster inference times without sacrificing the model's predictive accuracy.

Quick Start & Requirements

• Installation:
  • Prebuilt: pip install onnxslim
  • From Source: pip install git+https://github.com/inisis/OnnxSlim@main
• Usage:
  • CLI: onnxslim <your_onnx_model> <slimmed_onnx_model>
  • Python API: Load model using onnx.load, apply onnxslim.slim, and save using onnx.save.
• Prerequisites: Python and ONNX library. No specific versions or hardware requirements (like GPU/CUDA) are mentioned.

Highlighted Details

• Achieved 1st place in AICAS 2024 and 2025 LLM inference optimization challenges.
• Reached 1 million downloads as of January 2025.
• Key integrations include NVIDIA TensorRT-Model-Optimizer, HuggingFace Optimum, ultralytics, and transformers.js.
• Reported a 5% performance increase when merged into MNN-LLM.

Maintenance & Community

• Contact channels include Discord (https://discord.gg/nRw2Fd3VUS) and QQ Group (873569894).
• Significant recent activity involves merging into major projects, suggesting a transition from standalone maintenance to integrated development.

Licensing & Compatibility

• The project's license is not specified in the README. This lack of clarity presents a significant barrier for adoption, particularly in commercial contexts.

Limitations & Caveats

• The project's primary development focus appears to have shifted towards its integration into larger frameworks like TensorRT-Model-Optimizer and HuggingFace Optimum. Standalone maintenance status is unclear.
• The absence of a specified license creates ambiguity regarding usage rights and potential restrictions, especially for commercial use.
• Detailed benchmarks or specific accuracy retention figures beyond challenge results are not provided.