NVIDIA RAPIDS 24.10 now includes GPU-accelerated NetworkX and Polars with no code changes, improving compatibility with Python 3.12 and NumPy 2.x for better data processing. This release aims to enhance the user experience for data scientists and developers. The RAPIDS cuGraph now offers GPU-accelerated NetworkX, available in this update with NetworkX 3.4.
This upgrade allows end-to-end acceleration of graph workflows, improving performance for large datasets. To activate this feature, users can set the NX_CUGRAPH_AUTOCONFIG environment variable to True. The Polars GPU engine, powered by cuDF, is released in open beta, providing up to 13x faster workflows with no code change.
This enhancement is integrated into the Polars Lazy API, allowing users to trigger GPU computation using the engine keyword. RAPIDS v24.10 extends the capability of cuML’s UMAP algorithm to handle datasets larger than GPU memory, preventing out-of-memory errors. This is achieved through a batched approximate nearest neighbor algorithm that processes data subsets on the GPU.
Enhancements in cuDF’s pandas accelerator mode now support true NumPy arrays, improving compatibility and eliminating previous workarounds. Additionally, cuDF now supports a wider range of PyArrow versions by utilizing the Arrow C Data Interface. NVIDIA has introduced new guidelines for integrating GPUs into GitHub-based continuous integration systems, leveraging GitHub Actions’ support for hosted GPU runners.
This helps users easily integrate and test RAPIDS libraries without local GPU hardware. This release includes updates for compatibility with Python 3.12, NumPy 2.x, and other scientific computing software. However, support for Python 3.9 and older versions of NCCL is dropped. These updates in RAPIDS 24.10 continue to advance the accessibility of accelerated computing for data scientists and developers, offering enhanced performance and compatibility.
