Expand description
Device selection for the candle-backed inner loops.
Picks CUDA when the crate was built with the cuda feature and a
GPU is actually present at runtime; falls back to CPU otherwise. The
choice is made once on first call and cached for the rest of the
process — every subsequent tensor allocation lands on the same
device, so there are no surprise host↔device transfers in the
middle of a tight loop.
Callers shouldn’t need to think about devices in normal use:
default_device() returns whatever’s right for this build.
Functions§
- default_
device - Return the device docbert-plaid uses for all tensor compute.