Developed Dask Arrays support for the AnnTorchDataset component in the scverse/scvi-tools repository, enabling efficient on-demand computation for datasets exceeding available memory. This work involved updating dependencies to ensure seamless compatibility with Dask-based workflows and integrating comprehensive tests to validate the new functionality within the PyTorch data loading pipeline. By leveraging Python, Dask, and PyTorch, the implementation reduced memory usage and improved scalability for large-scale machine learning experiments. The changes enhanced the data handling capabilities of the training pipeline, allowing users to process and train models on much larger datasets without compromising performance or reliability in production environments.