Summary

This repository serves as a comprehensive survey of LLM-driven kernel generation, addressing the critical shift from manual, expert-dependent high-performance computing (HPC) kernel engineering to scalable, automated workflows. It targets researchers, engineers, and practitioners seeking to understand and leverage Large Language Models (LLMs) for optimizing GPU kernels and other computational tasks. The survey offers a structured overview of the rapidly evolving landscape, categorizing advancements and providing pointers to key resources.

How It Works

The field is broadly categorized into two main streams: LLM4Kernel and Agent4Kernel. LLM4Kernel focuses on applying LLMs directly to kernel synthesis, primarily through supervised fine-tuning (SFT) and reinforcement learning (RL) methodologies. These approaches aim to imbue LLMs with the capability to generate correct and performance-sensitive code across diverse programming abstractions. Agent4Kernel introduces more sophisticated, autonomous, closed-loop paradigms where LLM agents iteratively plan, utilize external tools, and refine generated code based on feedback. This agentic approach enables long-horizon optimization beyond the reach of manual or single-pass methods, often integrating hardware profiling and external memory management for enhanced capabilities and iterative improvement.

Quick Start & Requirements

This repository is a curated collection of research papers, datasets, code repositories, and benchmarks related to LLM-driven kernel generation. It does not provide a direct installation or execution environment for a specific tool. Users are directed to individual linked projects for their respective setup instructions, prerequisites (e.g., specific hardware, CUDA versions), and estimated resource footprints.

Highlighted Details

• LLM4Kernel Methodologies: Explores supervised fine-tuning and reinforcement learning techniques for direct LLM application in kernel synthesis, addressing challenges in correctness and performance.
• Agent4Kernel Architectures: Details agentic systems employing iterative planning, tool use, and feedback loops for complex optimization tasks, often incorporating learning mechanisms, external memory, hardware profiling, and multi-agent orchestration.
• Resource Landscape: Compiles a wide array of structured datasets (e.g., The Stack v2, HPC-Instruct), source code repositories (e.g., CUTLASS, FlashAttention, Triton, xFormers), frameworks (PyTorch, vLLM, TensorRT-LLM), domain-specific languages (Triton, TileLang), and knowledge bases.
• Benchmarking: Surveys numerous benchmarking frameworks designed to evaluate LLM-generated code quality and performance, covering aspects from correctness and verification to hardware-specific optimization and real-world inference workloads.

Maintenance & Community

The repository encourages community contributions via pull requests or issues to update the survey with new papers, projects, or feedback, reflecting its nature as a living document in a fast-moving research area. Specific maintainer details, sponsorships, or dedicated community channels (like Discord/Slack) are not provided within the README.

Licensing & Compatibility

No specific license is mentioned for the repository itself. Individual linked projects will have their own licenses, which users must consult for compatibility, especially for commercial use or closed-source linking.

Limitations & Caveats

As a survey, this repository's primary limitation is its scope, which is dependent on the authors' curation and the rapid pace of research in LLM-driven kernel generation. It does not offer a unified tool or framework but rather a comprehensive map of the existing ecosystem. Users must evaluate individual projects for their maturity, stability, and specific limitations, such as unsupported platforms or alpha status.