add-sgl-kernel

To integrate a custom CUDA kernel into a PyTorch extension, you must implement the C++ code, declare the functions, register the PyTorch extension, and update the CMake build system to expose the Python API.

Heavyweight AOT compilation for CUTLASS kernels involves writing the CUDA kernel, integrating it into the existing C++ project's CMake build system, and exposing it as a PyTorch op to ensure proper compilation and execution.

Writing tests and benchmarks for new CUDA kernels requires creating specific unit tests to validate functionality and performance benchmarks to measure execution speed, ensuring the kernel operates correctly within the library.

Yes, you can use sgl-kernel to add highly optimized matrix multiplication kernels by following the step-by-step guide from writing the initial CUDA code to making it fully available as a tested and benchmarked PyTorch op.

Yes, integrating custom C++ kernels requires configuring the CMake build system to properly declare, register, and build the PyTorch extension components alongside the existing library framework.