Automated CUDA Kernel Generation via RL on LLMs

Overview

An attempt to use GRPO as a low memory RL technique to improve LLM capabilities at CUDA kernel generation. Due to severe hardware/budget constraints, we were limited to using DeepSeek-R1-Distill-Qwen1.5B and training with LoRA (low-rank adaptation). We were able to improve the abilities of the LLM to generate valid CUDA kernels, but were limited in the ability to generate more efficient CUDA kernels.

Poster

View the PDF version here.

Source Code

Soon to be open source on GitHub.