Worked on optimizing memory management for the vllm-project/tpu-inference repository, focusing on stabilizing TPU inference workloads. Addressed a memory bottleneck by implementing a CPU-first loading strategy, which prevents high-bandwidth memory spikes during FP8 inference. This approach reduced memory pressure and improved throughput predictability in production environments. The solution involved modifying the weight loading process so that data is initially loaded on the CPU before being transferred to the TPU, ensuring more efficient resource utilization. Leveraged expertise in JAX, TPU optimization, and machine learning, and contributed a targeted bug fix that enhanced the reliability and scalability of inference pipelines.