flagos-ai/awesome-LLM-driven-kernel-generation

What it solves

Kernel engineering—writing low-level code that controls how GPUs or other accelerators execute computations—has traditionally required deep expertise and manual tuning. This curated collection addresses the challenge of keeping up with a fast-growing research area that applies large language models (LLMs) and autonomous agents to make kernel generation and optimization more scalable and automated.

How it works

The repository organizes a list of papers, tools, datasets, and benchmarks according to a taxonomy proposed in an accompanying survey. The main categories are:

• LLM4Kernel – techniques for training or fine-tuning LLMs specifically for kernel synthesis, split into supervised fine-tuning (SFT) and reinforcement learning (RL) approaches.
• Agent4Kernel – agentic systems that iteratively plan, use tools, and refine kernels, with sub-categories for learning mechanisms, external memory or skill management, hardware profiling integration, and multi-agent orchestration.
• Datasets – structured datasets and source code repositories that provide training data and reference implementations for kernel-related tasks.
• Benchmarks – resources for evaluating kernel generation and optimization performance (the README lists this as a section, though detailed entries are not shown in the provided excerpt).

Each entry typically includes a link to the paper, code, or dataset repository, along with publication metadata.

Who it’s for

Researchers, engineers, and students working on high-performance computing, LLM-based code generation, or AI accelerator optimization. It is also useful for anyone tracking the state of the art in automating GPU kernel development.

Highlights

• Papers span from 2024 to 2026, covering both foundational LLM training and advanced agentic workflows.
• Includes concrete resources like KernelLLM, AutoTriton, CUDA-L1/L2, KernelBench, and CUTLASS, each linked to their source or paper.
• Categorization follows a structured survey taxonomy, making it easy to navigate by approach or hardware target.
• Active community involvement is visible through GitHub stars and frequent updates listed on many entries.