add-sgl-kernel

Adding a custom CUDA kernel to a torch extension involves implementing the C++ kernel source, updating CMake and headers, registering the op, and exposing Python APIs for end-to-end integration.

An end-to-end CUDA C++ kernel workflow includes writing the kernel source, updating CMake and headers, registering the op, exposing Python APIs, and adding tests and benchmarks for validation.

Yes, updating CMake and header declarations is required when integrating a heavyweight CUDA kernel to properly compile the project and register the custom ops through the torch extension.

You benchmark a new custom op in sgl-kernel by adding benchmarking scripts alongside your tests, allowing you to verify both correctness and performance gains of the CUDA kernel implementation.

No, exposing Python APIs for an AOT CUDA kernel requires updating CMake configurations to compile the torch extension and properly register the custom op for Python accessibility.

The best way to verify correctness for a newly added C++ kernel is to implement dedicated tests within the project structure, ensuring the custom op behaves as expected before benchmarking performance.