AutomatedCUDA/A100-CUDA-RL

KernelForge-OpenEnv: A100 CUDA Kernel RL

Interactive demo for the KernelForge reinforcement learning environment that trains LLMs to write optimized CUDA kernels targeting NVIDIA A100 GPUs.

Features

• —Live Optimization — CUDA kernel editor with real-time metrics (compilation, correctness, speedup)
• —Training Monitor — Reward progress, success rates, and model performance tracking
• —Hardware Telemetry — A100/H100 SM utilization, memory throughput, and feature availability
• —PAC Verification — Mathematical correctness verification for graph kernels (WCC)

Architecture

This demo showcases the OpenEnv-compatible RL environment:

• —Discrete milestone reward: {-1, 1, 2, 3} based on correctness and speedup tiers
• —Multi-turn episodes: 3 turns per episode (model sees errors, iterates)
• —Anti-hack checks: 5 Dr. Kernel-inspired runtime checks
• —Task pool: Ops-6K operators + doubleGraph topology-aware tasks

Links

• —GitHub: OCWC22/A100-CUDA-RL
• —Papers: CUDA Agent, Dr. Kernel, KernelBench