enums.Precision

enums.Precision()

The floating-point width a descent runs in.

A property of the run, not of the objective: a solver may scope double precision around its own forward pass (:func:cadjoint.flow.precision.double_precision does), but :func:jax.grad runs the transposed pass after that scope has closed, so the flag has to be held for the whole loop or the backward pass meets float64 intermediates it cannot materialise. Declaring it on the optimization rather than at a scene’s module scope is what keeps the scene float32 for the WGSL shader, which has no f64.

Attributes

Name	Description
DOUBLE	jax_enable_x64 for the duration of the run, restored after.
SINGLE	Whatever the process is already in (the default).