Warp-NN: CUDA Graphable Neural Networks for NVIDIA Warp#

Warp-NN is a Warp-native Python library for building and training neural networks for Physical AI workflows. It is designed for compact neural network components that run directly within Warp-based simulation, robotics, control, and differentiable computing pipelines. It is not intended to be a general-purpose replacement for PyTorch, JAX, or other ML frameworks.

Disclaimer

Warp-NN is not part of the warp-lang package, and it is not maintained by the NVIDIA Warp core team. It is a Warp ecosystem library maintained by @Toni-SM / NVIDIA Isaac Sim. Issues, releases, roadmap, and support are managed by the maintainers of this repository.

Note

Although the library strives to offer high-performance implementations, they may not always outperform other highly optimized solutions for NVIDIA GPUs that use cuBLAS / cuDNN.

Main features:

  • Clean, simple, and readable code.

  • Configurable CUDA kernels.

  • CUDA Graphable implementation.

Quickstart#

The easiest way to install Warp-NN is from PyPI. Refer to the Installation section for more details.

pip install warp-nn

Sections#