add-jit-kernel

To add custom CUDA kernels to SGLang, use its JIT compilation system to implement C++ abstractions for tensor validation, vectorized memory access, and kernel launching, alongside Python wrapper integration for custom tensor operations.

JIT kernel compilation in SGLang enables lightweight, on-the-fly compilation of custom CUDA kernels during runtime, whereas AOT involves ahead-of-time compilation, with the Skill specifically detailing the JIT integration strategy for high-performance operations.

To implement and benchmark element-wise operations, follow the step-by-step tutorial for creating a JIT CUDA kernel, which includes writing the kernel logic, wrapping it in Python, and running the provided benchmarking procedures to evaluate performance.

Yes, you can use SGLang's provided C++ and CUDA abstractions for safe kernel development, which includes built-in tensor validation, device utilities, and vectorized memory access for your custom JIT CUDA kernels.

Yes, implementing Python wrappers is a required step in the SGLang JIT kernel integration workflow, ensuring your custom CUDA operations are callable within the framework alongside completing tests and benchmarks.