Worked on the vllm-project/tpu-inference repository to deliver key-value (KV) offloading for attention data parallelism, targeting improved memory efficiency during large-scale model training and inference on TPUs. Developed new functions for gathering and updating KV caches using pure JAX operations, and implemented dynamic KV cache metadata management to optimize distributed training performance. This feature enables handling of larger models by reducing memory pressure and improving scalability and throughput across multiple TPU devices. The work demonstrated expertise in distributed systems, machine learning, and TPU programming, with all code-level changes traceable and focused on enhancing memory footprint and training efficiency.