Real-time LLM Inference on Standard GPUs: 3k tokens/s per request