CUDA-L1: Improving CUDA Optimization via Contrastive Reinforcement Learning Trains a DeepSeek-v3-671B model to optimize CUDA kernels using only execution-time speedup as reward. Pipeline: - SFT: Finetuned on 2.1K correct, executable CUDA variants from 6 LLMs across 250 KernelBench tasks - Self-Supervised: Iterative REINFORCE updates on successful code (exec correctness) - Contrastive-RL: Prompts include prior variants speedup scores; model compares, improves, and updates via GRPO Results (KernelBench, 250 tasks): - 17.7× avg speedup (A100), 449× max - 100% success on complex ML workloads (Level 3), 50.8× mean - Generalizes across GPUs: 19.0× (3090), 17.8× (H100), 13.9× (H20) Findings: - Discovers non-trivial strategies that multiply performance - Identifies gatekeeper techniques (e.g. stream mgmt → unlocks CUDA graphs) - Learns optimization dependencies (e.g., certain techniques must precede others) - Outperforms both GRPO-only and evolutionary LLMs - No reward hacking (via prompt constraints)