kernel-benchmarks

LLM4LLM introduces a deployment-aware closed-loop optimization framework to bridge kernel benchmarks and real LLM inference, achieving up to 6.98x speedups on H100 GPUs.