Cinder: Design and Implementation of a Compact CUDA Tensor Library

Cinder exposes one Python value type over a C++ and CUDA implementation, currently focused on float32 values, runtime-rank shapes, row-major storage, explicit host-device transfer, and eager materialized operations. The report separates mathematical tensors, finite arrays, storage layout, and execution plans instead of collapsing those meanings into one ambiguous object.

This is an auditable engineering-artifact record rather than a performance-competition claim. It documents the public API, native ownership, compact device metadata, host-side kernel planning, current capability boundaries, and the risks involved in evolving toward more dtypes, views, stronger tests, and optimization representations.