DeepYardDeepYard
K

KernelSkill

Multi-agent framework that systematically optimizes GPU kernels using expert rules over LLM heuristics

Open SourceFree

About

KernelSkill is a research framework that applies multi-agent collaboration to GPU kernel optimization. Instead of relying on implicit LLM trial-and-error, it uses explicit expert optimization rules to systematically improve kernel efficiency. Designed for researchers and HPC developers working on GPU performance optimization, it offers interpretable strategies that make optimization decisions transparent and reproducible.

Details

Typecoding-agent
Deploymentself-hosted
Supported Models

Tags

multi-agentautonomousopen-sourceresearchorchestration