Surpassing vLLM with a Generated Inference Stack