krnopt-cuda-domain-optimization

Optimizing CUDA kernels for MoE workloads involves classifying the kernel into routing, dispatch, or expert MLP fusion domains. The planner then applies domain-driven patterns like grouped GEMM to generate a concrete optimization plan targeting specific hardware architectures.

Domain-guided CUDA kernel optimization is a planning approach that uses workload-domain insights, such as decode or prefill regimes, instead of generic memory coalescing advice. It identifies domain boundaries like grouped GEMM to prescribe specific hardware-aware improvements and validation criteria.

Planning boundary changes for a grouped GEMM CUDA kernel requires identifying its workload shape and applying domain references to prescribe specific boundary modifications. The resulting plan includes target architectures, build targets, and correctness gates for validation.

Generic CUDA optimization advice is insufficient for MoE kernel design because it ignores domain boundaries. MoE kernel design requires domain-driven patterns that address routing, dispatch, and expert MLP fusion specific to decode and prefill workload regimes.

A CUDA kernel optimization plan should include architecture-specific constraints, boundary changes, build targets, correctness gates, and timing criteria. These elements ensure the domain-specific improvements align with targeted hardware regimes like decode and prefill.