KernelBlaster: Continual Cross-Task CUDA Optimization via Memory-Augmented In-Context Reinforcement Learning
This addresses the problem of inefficient CUDA optimization for GPU programming, offering a novel agentic framework that is incremental in enhancing existing LLM-based methods.
The paper tackles the challenge of optimizing CUDA code across GPU architectures by proposing KernelBlaster, a memory-augmented in-context reinforcement learning framework that accumulates knowledge to improve search capabilities, achieving geometric mean speedups of 1.43x to 2.50x on benchmarks compared to a PyTorch baseline.
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.