add-sgl-kernel

To add a new CUDA kernel, you implement the operation in C++/CUDA, expose it via Python API, register it in the library, and wire up CMake. This ensures your custom op integrates with PyTorch and the AOT workflow.

A complete CUDA kernel integration requires C++/CUDA implementation, public declarations, library registration, Python API exposure, unit tests, and benchmarks. This validates both correctness and performance within the ML pipeline.

You benchmark a custom PyTorch CUDA kernel by wiring up a Triton or Torch-backed benchmark alongside the CMake build. This validates the performance and correctness of your new operation within the SGL kernel ecosystem.

Yes, you can extend the SGL kernel by writing a new operation in C++/CUDA and registering it in the library. This exposes the custom op to PyTorch and integrates it into the AOT workflow.

The best way to test a new CUDA kernel is to wire up unit tests and benchmarks using CMake. This validates the correctness and performance of your custom operation within the SGL kernel ecosystem.

Registering your CUDA kernel in the library exposes the C++/CUDA implementation to the Python API. This enables PyTorch integration and ensures the custom op functions within the AOT workflow.