DLPack

lastmod 2026-11-10created 2024-07-122 edits

DLPack is a common in-memory tensor structure specification for exchanging tensor data between deep learning frameworks without copying. It defines a C struct holding the data pointer, device information, data type, shape, and strides, and is supported by PyTorch, CuPy, TensorFlow, JAX, and others. Because frameworks share the same GPU memory, there is no conversion cost.

common in-memory tensor structure

See also

• cupy
• pytorch

Favorite site

• Github - dmlc/dlpack