Параллельный ввод-вывод cuPyNumeric-массивов в HDF5

★ 7.6 · content

cupynumeric-hdf5 is a Claude Code skill that enables parallel, distributed HDF5 I/O for cuPyNumeric arrays via the legate.io.hdf5 module. It exposes three operations: to_file writes an array to an .h5 file as a virtual dataset where every rank writes its own tile independently, from_file loads an HDF5 dataset into a distributed cuPyNumeric array, and from_file_batched streams a large file in chunks without pulling it fully into memory. The skill requires cuPyNumeric and Legate 26.01 or newer, plus h5py installed via conda-forge — the module imports it at load time, so the import fails without it. Use it to save large arrays to a single .h5/.hdf5 file, feed an HPC post-processing pipeline, or accelerate disk I/O with GPUDirect Storage; it does not handle Parquet, Zarr, .npz archives, or plain h5py workflows outside Legate.