add-sgl-kernel

To add custom CUDA kernels to the sgl-kernel library, implement the kernel in C++, declare it, register it in the Torch extension, update the CMake build system, expose the Python API, and create tests and benchmarks.

Integrating heavyweight AOT C++ operations into sgl-kernel involves writing the kernel, updating C++ declarations, registering the operation in common_extension.cc, modifying CMakeLists.txt, and exposing the functionality through a Python API.

Yes, sgl-kernel supports benchmarking custom PyTorch CUDA extensions against Triton, providing guidance on creating performance benchmarks in the benchmark directory to evaluate your kernel's efficiency.

You test newly added CUDA kernels in the sgl-kernel ecosystem by creating robust unit tests using pytest within the tests directory to ensure the correctness of your custom operations.

Yes, you can use sgl-kernel to optimize specific model inference tasks by adding highly optimized custom matrix multiplication kernels to accelerate your specific workloads.

When writing element-wise operations for sgl-kernel, you must follow the strict integration workflow, which requires manual C++ declarations, Torch extension registration, and CMake build updates to function properly.