NVIDIA/TensorRT

Provides modular ONNX parser and plugin ecosystem for custom operator support, enabling developers to extend inference capabilities beyond built-in operations. The SDK performs graph optimization, kernel auto-tuning, and precision calibration (transitioning from implicit to explicit quantization) to compile trained models into optimized GPU engines. Integrates with TensorFlow, PyTorch, and ONNX through conversion tools and sample applications, with multi-device inference support via NCCL.

12,784 stars and 213,033 monthly downloads. Used by 2 other packages. Actively maintained with 2 commits in the last 30 days. Available on PyPI.