state-management cuda high-throughput memory-efficiency rocm gpu-optimization long-term-memory inference-optimization transformer-architecture lossless-compression kv-cache large-language-models vllm llm-inference context-window enterprise-ai real-time-inference vram-optimization linear-serialization state-inference