add-jit-kernel

To add a JIT CUDA kernel in sglang, you implement lightweight element-wise operations using provided scaffolding, validate tensors with TensorMatcher, and execute deterministic launches via LaunchKernel for safe integration.

The best way to validate tensors before launching a custom GPU kernel is using TensorMatcher, which ensures input tensors meet the required conditions for safe, deterministic execution via LaunchKernel.

To perform 128-bit aligned memory operations in a CUDA kernel, you use AlignedVector within the JIT kernel scaffolding, ensuring efficient vectorized memory patterns for element-wise operations.

Yes, you can use JIT CUDA kernels for simple element-wise operations by integrating runtime checks with TensorMatcher and utilizing LaunchKernel to ensure deterministic and safe execution.

Yes, sglang's jit_kernel module supports deterministic kernel launches through the provided LaunchKernel utility, which automates safe execution for rapidly prototyped JIT CUDA kernels.

You should not use a JIT CUDA kernel when your task requires complex operations beyond simple element-wise calculations, or when tensor validation constraints provided by TensorMatcher cannot be satisfied.