DICE: Diffusion Large Language Models Excel at Generating CUDA Kernels

TL;DR

This paper introduces DICE, a diffusion-based large language model tailored for CUDA kernel generation, leveraging a new dataset and a bi-phase reinforcement learning framework to outperform existing models.

Contribution

The paper presents DICE, a novel diffusion LLM for CUDA kernels, along with CuKe dataset and BiC-RL training framework, advancing code generation capabilities.

Findings

DICE outperforms comparable autoregressive and diffusion models on KernelBench.

The model achieves state-of-the-art results in CUDA kernel generation.

The approach demonstrates the effectiveness of diffusion models in code synthesis.

Abstract

Diffusion large language models (dLLMs) have emerged as a compelling alternative to autoregressive (AR) LLMs, owing to their capacity for parallel token generation. This paradigm is particularly well-suited for code generation, where holistic structural planning and non-sequential refinement are critical. Despite this potential, tailoring dLLMs for CUDA kernel generation remains challenging, obstructed not only by the high specialization but also by the severe lack of high-quality training data. To address these challenges, we construct CuKe, an augmented supervised fine-tuning dataset optimized for high-performance CUDA kernels. On top of it, we propose a bi-phase curated reinforcement learning (BiC-RL) framework consisting of a CUDA kernel infilling stage and an end-to-end CUDA kernel generation stage. Leveraging this training framework, we introduce DICE, a series of diffusion large language models designed for CUDA kernel generation, spanning three parameter scales, 1.7B, 4B, and 8B. Extensive experiments on KernelBench demonstrate that DICE significantly outperforms both autoregressive and diffusion LLMs of comparable scale, establishing a new state-of-the-art for CUDA kernel generation.