add-sgl-kernel

To integrate a CUDA kernel into sgl-kernel, you need C++ declarations, torch operator registration, CMake source inclusion, and Python API exposure. This ensures the kernel is properly built into the AOT wheel and callable from Python.

AOT compilation in sgl-kernel means heavyweight CUDA or C++ kernels are pre-compiled into the wheel, avoiding runtime compilation overhead. This approach is required for complex operations like those dependent on CUTLASS.

Writing pytest tests for a custom CUDA kernel involves adding correctness checks and error-condition coverage. You should validate PyTorch tensor dtypes, device placement, and kernel launch errors to ensure robust integration.

The best way to benchmark a CUDA kernel against PyTorch is using Triton benchmarking. This allows you to measure performance gains and validate that the AOT kernel provides a speedup over native PyTorch operations.

Yes, you can use CUTLASS-dependent operations with sgl-kernel. The integration workflow specifically supports these heavyweight C++ kernels by handling the necessary CMake configuration and torch operator registration.

Required PyTorch tensor validations for a new CUDA kernel include checking tensor dtypes, verifying device placement, and handling stream execution. These checks prevent runtime errors during kernel launches.