KernelBlaster: Continual Cross-Task CUDA Optimization via Memory-Augmented In-Context Reinforcement Learning

TL;DR

KernelBlaster introduces a memory-augmented reinforcement learning framework that enhances CUDA code optimization across GPU generations by learning from past experience, significantly improving performance over traditional methods.

Contribution

It proposes a novel MAIC-RL framework with a persistent knowledge base for CUDA optimization, enabling systematic exploration and transfer of knowledge across GPU architectures.

Findings

Achieves up to 2.50x speedup on KernelBench levels

Outperforms baseline by 1.43x on average

Provides open-source framework for reproducible CUDA optimization

Abstract

Optimizing CUDA code across multiple generations of GPU architectures is challenging, as achieving peak performance requires an extensive exploration of an increasingly complex, hardware-specific optimization space. Traditional compilers are constrained by fixed heuristics, whereas finetuning Large Language Models (LLMs) can be expensive. However, agentic workflows for CUDA code optimization have limited ability to aggregate knowledge from prior exploration, leading to biased sampling and suboptimal solutions. We propose KernelBlaster, a Memory-Augmented In-context Reinforcement Learning (MAIC-RL) framework designed to improve CUDA optimization search capabilities of LLM-based GPU coding agents. KernelBlaster enables agents to learn from experience and make systematically informed decisions on future tasks by accumulating knowledge into a retrievable Persistent CUDA Knowledge Base. We propose a novel profile-guided, textual-gradient-based agentic flow for CUDA generation and optimization to achieve high performance across generations of GPU architectures. KernelBlaster guides LLM agents to systematically explore high-potential optimization strategies beyond naive rewrites. Compared to the PyTorch baseline, our method achieves geometric mean speedups of 1.43x, 2.50x, and 1.50x on KernelBench Levels 1, 2, and 3, respectively. We release KernelBlaster as an open-source agentic framework, accompanied by a test harness, verification components, and a reproducible evaluation pipeline.