Rupliu Liu developed a performance-focused feature for the vllm-project/tpu-inference repository, implementing Key-Value Quantization within the GptOssAttention module. Using Python and leveraging deep learning frameworks such as JAX and TensorFlow, Rupliu designed and delivered an approach that reduces memory usage during inference and increases throughput for large attention workloads. This technical solution enables more scalable deployments in memory-constrained environments by optimizing how attention data is stored and processed on TPUs. The work demonstrated a clear understanding of both the underlying machine learning concepts and practical engineering, resulting in merge-ready code that addressed a real bottleneck in inference scalability.