add-sgl-kernel

To add a CUDA kernel to sgl-kernel, you create a .cu file with proper dispatch, register it with the Torch extension, expose the Python API, and add unit tests and benchmarking scripts to validate GPU acceleration and performance gains against PyTorch.

Integrating an AOT C++ kernel into sgl-kernel involves implementing the kernel with CUDA stream usage, adding C++ declarations, registering with the Torch extension, exposing the Python API, and configuring build integration alongside unit tests and performance benchmarks.

You benchmark a PyTorch operation against a custom CUDA kernel by adding a benchmark script to the sgl-kernel workflow, which rigorously compares the GPU-accelerated kernel execution against the baseline PyTorch operation to ensure performance parity or gains.

You need to expose a Python API when extending sgl-kernel, because the workflow requires adding C++ declarations and registering the kernel with the Torch extension to make the new GPU-accelerated operation accessible and testable from Python.

You should add a heavyweight kernel to sgl-kernel when you need to implement a new elementwise operation or optimize an existing one for GPU acceleration, requiring rigorous validation and performance benchmarking that standard PyTorch operations do not provide.

sgl-kernel supports build integration for AOT CUDA extensions by automating the placement of .cu files, Torch extension registration, and Python API exposure, enabling seamless GPU-accelerated workflow extensions with proper testing.