add-sgl-kernel

Adding custom CUDA kernels to a PyTorch C++ extension involves writing the kernel in C++, registering it via Torch extension, updating CMake build files, exposing the Python API, and creating unit tests and performance benchmarks.

Integrating AOT compiled GPU operations into sgl-kernel requires updating C++ declarations, configuring the CMake build system, registering the Torch extension, and exposing the functionality through Python APIs for immediate use.

Yes, sgl-kernel supports dtype dispatching for FP16, BF16, and FP32 in CUDA programming, allowing custom kernels to integrate seamlessly with PyTorch's CUDA backend across multiple precision formats.

The best way to write benchmarks for custom CUDA kernels in a PyTorch library is to follow the sgl-kernel integration workflow, which guides you through creating robust performance benchmarks alongside unit tests for new operations.

Yes, you need to update CMake build files when adding new C++ kernels to sgl-kernel, as the build system must be configured to compile and link the new AOT CUDA operations with the PyTorch backend.

Yes, you can use CUTLASS for element-wise tensor scaling operations in sgl-kernel, leveraging it for maximum performance while following the guided workflow from kernel implementation to testing and benchmarking.