Worked on the ROCm/FBGEMM repository to address a precision consistency issue in the SplitTableBatchedEmbeddingBagsCodegen module. Implemented a robustness fix by ensuring that cache_precision defaults to weights_precision when not explicitly set, aligning the precision used for embedding cache and weights. This adjustment prevents discrepancies that could arise from unset precision values, thereby improving numerical stability and reliability during production deployments. The solution was developed using Python and leveraged expertise in Deep Learning, GPU Computing, and PyTorch. The code was reviewed and integrated upstream, contributing to more stable and accurate embedding operations in GPU-accelerated environments.