During May 2025, Ziheng contributed to the pytorch/FBGEMM repository by implementing writeback support for SplitTableBatchedEmbeddingBagsCodegen, enabling direct value assignment to TBE tables and laying groundwork for more memory-efficient embedding learning. Ziheng introduced a configuration option and a backward pre-hook for EXACT_SGD, allowing precise management of gradient updates during writeback. The work involved modifying both configuration and code generation components, as well as developing targeted tests to validate the new functionality. Leveraging expertise in Deep Learning, GPU Computing, and PyTorch, Ziheng delivered a technically deep feature that enhances embedding workflows and supports future optimization efforts.