RAPIDS Introduces GPU Polars Streaming and Unified GNN API Enhancements
[ad_1]
Tony Kim
Jul 04, 2025 21:25
NVIDIA’s RAPIDS suite model 25.06 unveils new options including GPU Polars streaming, a unified GNN API, and zero-code ML speedups, enhancing Python knowledge science capabilities.
NVIDIA has introduced the newest model 25.06 of its RAPIDS suite, a assortment of CUDA-X libraries for Python knowledge science. This launch introduces several groundbreaking options designed to improve computational effectivity and knowledge processing capabilities, according to NVIDIA.
Polars GPU Engine Enhancements
The new launch brings important updates to the Polars GPU engine, initially launched in September 2024. One of the key options is the experimental streaming executor, which permits execution on datasets bigger than the accessible VRAM through knowledge partitioning and parallel processing. This improvement is essential for accelerating analytics operations on extraordinarily massive datasets, scaling from a whole lot of gigabytes to terabytes. Additionally, the replace introduces a shuffle mechanism to facilitate knowledge redistribution between units and help multi-GPU execution.
Another enhancement contains help for rolling aggregations and expanded column manipulation capabilities, which are notably helpful for time collection knowledge evaluation. The GPU engine now also helps a wider vary of expressions for datetime column manipulation, such as .strftime() and .cast_time_unit().
Unified API for Graph Neural Networks (GNNs)
The integration of CompleteGraph into NVIDIA’s cuGraph-PyG has led to the creation of a Unified API, which accelerates function fetching for GNNs. This API permits customers to seamlessly transition from a single GPU to multi-GPU or multi-node workflows with out modifying their scripts. The acquainted torchrun command from PyTorch is used to handle processes, facilitating ease of use for PyTorch customers.
Zero-Code Change ML Enhancements
The RAPIDS 25.06 launch expands its zero-code-change acceleration for machine studying by including help vector machines (SVMs) in the cuML library. This permits current scikit-learn workflows utilizing SVMs to profit from GPU acceleration with out any code modifications. The replace improves compatibility with scikit-learn, enhancing parameter validation and error dealing with.
Additional Platform and Compatibility Updates
The launch also contains upgrades to the RAPIDS Memory Manager (RMM), which now helps the hardware-based decompression engine on NVIDIA Blackwell GPUs. This function guarantees efficiency enhancements in IO-intensive workflows. Furthermore, the platform now helps Python 3.13, marking the last launch to help CUDA 11.
Overall, the RAPIDS 25.06 launch delivers important developments for knowledge scientists and builders, focusing on enhanced efficiency and ease of use for GPU-accelerated knowledge processing duties.
Image supply: Shutterstock
[ad_2]
