Description
This is not a bug report, I’m trying to understand why recent versions of TensorRT seem to require both libnvinfer_builder_resource and libnvinfer_builder_resource_win libraries.
We’re developing software that requires optimizing and running models at runtime, using TensorRT as a dependency. We package TensorRT into our installer for ease of use. libnvinfer_builder_resource and libnvinfer_builder_resource_win take 2GB each and are by far the largest libraries. This is quite a size bump from previous versions, even when we used cuDNN on older versions.
My questions are:
- Do we actually need both libraries + the other (nvinfer, lean, etc) to build and run engines? Could we have a breakdown of what are they meant to be used for, like this but extended
- Out of curiosity what do they contain?
- Why the “win” name? It seems to confuse some people on Linux Engine building failure of TensorRT 10.2.0 (pip install) when building a custom diffusion model on RTX 4090 · Issue #3983 · NVIDIA/TensorRT · GitHub
Environment
TensorRT Version: 10.9
Platform: Ubuntu Desktop (x86_64)