CuPy¶

CuPy uses Advect's single-device Array API path. Pass CuPy arrays to a transform and write numerical code against the namespace returned by x.__array_namespace__(); this does not intercept arbitrary cupy.* calls.

The generated Array API table defines the callable set for this path. CuPy exercises that shared contract through __array_namespace__(); Advect does not maintain a second CuPy inventory.

Install¶

Install the CuPy wheel family matching the CUDA generation available to the host. Advect does not select the CUDA package, and no Advect extra or registration import is required.

Boundaries¶

• One process and one device; multi-GPU placement and stream control are not claimed.
• Arrays attached to another autodiff system (requires_grad=True) are rejected instead of composing two derivative tapes.