EGG Framework Boosts GPU Kernel Generation with Expert Guidance.

Developing high-performance GPU kernels is crucial for managing the escalating computational demands of large language models (LLMs), but this process traditionally relies heavily on manual optimization by specialized experts. While recent LLM-based approaches have shown potential for automating kernel generation, they often struggle to achieve both correctness and optimal performance due to a lack of domain-specific optimization guidance. The EGG framework addresses this by integrating expert optimization principles directly into the LLM's decision-making process. It structures kernel generation into two hierarchical stages: first, designing the algorithmic structure to establish a high-quality computational foundation, and second, performing hardware-specific tuning through techniques like parallel mapping, tensor tiling, and memory optimization. This staged decomposition provides clear optimization objectives and a structured design space for progressive refinement. EGG employs a stage-aware multi-agent collaboration mechanism to manage context across and within stages, ensuring stable optimization trajectories. Experiments on KernelBench and real-world workloads demonstrate that EGG achieves an average 2.13x speedup over PyTorch, outperforming other agent-based and reinforcement learning approaches.