Параллельная загрузка шардированных данных в cuPyNumeric

★ 7.6 · content

cupynumeric-parallel-data-load is a Claude Code skill that loads sharded on-disk datasets — sharded .npy, Parquet/Arrow, HDF5, raw binary, and custom layouts — into a distributed cuPyNumeric ndarray using a manual partition plus leaf @task launch with CPU/OMP/GPU variants. Legate Python tasks handle per-shard reads in parallel across all available processors; per-shard row counts may differ across files as long as dtype and trailing axes match. Use this skill when no single-call loader fits, such as replacing sequential np.concatenate calls with parallel GPU reads across multi-file sharded data. The reference implementation lives at assets/examples/parallel_npy_load.py and supports single-node (legate --gpus 4) and multi-node Slurm launches.