Worked on optimizing memory management in distributed systems by delivering a core feature to the pytorch/pytorch repository, focusing on NCCL Symmetric Memory. Developed a first-level cache for tensor-to-allocation lookups, combined with a two-level lookup mechanism that uses both cache and cuMemGetAddressRange, with a safe fallback path. This approach, implemented in C++ and CUDA, reduced lookup overhead in the rendezvous path and achieved a dramatic speedup for large allocations on multi-GPU hardware. The work was validated through targeted tests and benchmarks, directly improving latency and scalability for large-scale distributed training and enhancing NCCL memory resource utilization in production environments.