add-cuda-kernel

To add a custom CUDA kernel to FlashInfer, you place code under include/flashinfer, csrc, and flashinfer/jit, then wire it through the JIT and TVM-FFI bindings for runtime execution.

FlashInfer JIT compilation integrates with TVM-FFI bindings by wiring custom CUDA kernels through the csrc and jit binding layers, ensuring proper data flow and runtime execution.

Yes, you can implement custom GPU operations for production workloads in FlashInfer by adding a CUDA kernel and exposing it through the Python API with robust input validation and error reporting.

The best way to wire a CUDA launcher and Python API in FlashInfer is to implement the launcher within the jit and csrc layers, applying robust input validation and error reporting for production workloads.

You need to configure the build system for a CUDA kernel in FlashInfer when adding custom GPU operations, ensuring proper integration with the existing csrc and jit binding layers for runtime execution.