add-sgl-kernel

Adding a custom CUDA kernel to sgl-kernel requires placing the C++ source files, wiring CMake, registering the op, creating Python bindings, and implementing standardized tests and benchmarks against PyTorch baselines.

Integrating AOT C++ kernels into the sgl-kernel ecosystem involves a guided end-to-end workflow that automates csrc implementation, include declarations, torch extension wiring, and build validation to prevent boilerplate errors.

Setting up GPU kernel benchmarks against PyTorch baselines is achieved through the standardized benchmarking harness included in the kernel integration workflow, which automates performance validation for new CUDA kernels.

Yes, the workflow handles CMake wiring and torch extension registration automatically, enforcing the required steps for op registration and Python bindings to ensure the GPU kernel integrates correctly into the sgl-kernel ecosystem.

This workflow is specifically designed for heavyweight AOT CUDA/C++ kernel integration into sgl-kernel. It manages complex end-to-end tasks like CMake wiring, op registration, and benchmark harnessing required for substantial kernel development.