cuDF — ускорение pandas на GPU

★ 7.6 · content

accelerated-computing-cudf is a Claude Code skill that delivers official NVIDIA-authored guidance for working with cuDF GPU DataFrames, accelerating pandas workflows on GPU, and building high-throughput ETL pipelines. The skill covers three distinct paths: cudf.pandas for broad compatibility with silent CPU fallback, the explicit cuDF API for hot-path ETL optimization, named DataFrame migrations, and parity-sensitive operations, and dask-cuDF for datasets that exceed GPU memory using LocalCUDACluster with per-GPU workers. Key operations include join, groupby, read_csv/read_parquet, nullable semantics, rolling windows, reshape, and memory management via RMM pool allocators and spill-to-host. Targeting cuDF release 26.04 on CUDA 12/13 with Python 3.11–3.14, the skill is most effective for datasets above 100K rows, where GPU transfer overhead stops outweighing the speedup.