pycc.device

Device / precision manager for PyCC wavefunctions.

A single object that owns the device (‘CPU’/’GPU’) and precision (‘SP’/’DP’) policy: it validates the kwargs, resolves the storage (device0) and compute (device1) handles, builds the contraction backend, and applies the dtype/placement cast policy through seed_compute() / seed_store().

This centralizes logic that was previously duplicated across ccwfn.__init__ (the SP real-casts and the GPU torch casts) and the contraction backend’s __init__ (a third copy of the device resolution). In the 2026-06 refactor the manager is created in ccwfn.__init__; it moves to the Wavefunction base in Phase 3.

Notes

GPU tensors are seeded real (float32/64). Complex arithmetic appears only transiently during RT-CC propagation; ContractionBackend and utils.dot upcast the real operands to complex per-contraction, since torch (unlike NumPy) does not promote real<->complex in einsum/matmul/dot.

Classes

ContractionBackend([device, device1])	A device-aware wrapper around opt_einsum.contract.
DeviceManager([device, precision])	Owns the device/precision policy and the contraction backend.