A Few Key Features in 10.3
This TensorRT release includes the following enhancements
- Enhances cross-platform compatibility for TensorRT engines, allows for executing TensorRT engine on a runtime platform that is different from the one it was built on. [experimental]
- Improved engine build time when GEMMs have large weight constants.
- Improved the performance of Normalization and FP8 quantization fusion.
See the full release notes on the NVIDIA documentation page.