add-sgl-kernel

To add a new CUDA kernel to sgl-kernel, write the C++ kernel code, register it as a Torch extension, expose the Python API, and integrate it into the build pipeline for compilation and testing.

Integrating a C++ kernel involves declaring and registering the operation so it works with PyTorch, exposing the functionality through a Python API, and ensuring the build system compiles the extension correctly.

You test and benchmark a custom CUDA kernel by adding unit tests for correctness and performance benchmarks to verify speed, enforcing device validation and compilation checks within the CI pipeline.

Yes, sgl-kernel enforces device validation for new PyTorch extensions. The workflow includes compilation checks and performance verification within CI to ensure operations run correctly across CUDA devices.

The best way to deploy a heavyweight C++ kernel is using an end-to-end integration workflow that covers C++ code, Torch extension registration, Python API exposure, and automated build, test, and benchmark pipelines.

Your Torch extension may fail compilation checks if the C++ kernel code is not properly registered or if device validation fails. The workflow enforces these checks within CI to prevent integration errors.