CUDA kernel Usages
CUDA kernel Coding agent
Kevin-32B
CUDA Agent
ByteDance Seed and Tsinghua University present CUDA Agent, an RL system that optimizes CUDA kernels based on real GPU profiling speed. Achieves 98.8% pass rate on KernelBench, 96.8% faster-than-compile rate vs torch.compile, and 2.11x geometric mean speedup. Publicly releases 6,000 training tasks (CUDA-Agent-Ops-6K).

Seonglae Cho
