February 2021 Summaries

This new release of Ray 1.2 introduces two major features: native support for spilling to external storage and integration with libraries from the Python data processing ecosystem, including PySpark, Dask, Modin, and Mars. The goal is to enable developers to work with distributed data processing more directly within their Python applications, reducing the need for separate clusters or services. This should improve system efficiency and reduce operational overheads, allowing users to focus on writing more expressive code without worrying about setting up and maintaining complex infrastructure.

Huggingface Transformers has integrated the Retrieval Augmented Generation (RAG) model with Ray, a library for building scalable applications, to improve the scalability of RAG distributed fine-tuning. This integration speeds up retrieval calls by 2x and improves the overall performance of RAG on knowledge-intensive tasks. The new implementation uses Ray's stateful actor abstractions to load the index and handle retrieval queries, overcoming limitations of previous implementations. With this integration, users can leverage RAG for retrieval-based generation on their own knowledge-intensive tasks, and also take advantage of hyperparameter tuning with Ray Tune library.

Scikit-learn is a Python library for machine learning that can be optimized by changing the solver, which allows better use of hardware with more efficient algorithms. To further speed up model building, hyperparameter tuning techniques such as grid search, random search, and Bayesian optimization can be used, with libraries like Tune-sklearn providing cutting-edge techniques and consistency with the scikit-learn API. Parallelizing or distributing training using joblib and Ray can also significantly increase model building speed, especially for models with high parallelism like random forests. By utilizing these approaches, developers can create the best scikit-learn model in the least amount of time.