The workaround for GPU inference failure allocates new fields

[Sbozzolo]

ClimaLand.jl/src/standalone/Soil/energy_hydrology.jl

Lines 648 to 650 in cd785c8

	# Wrap hydrology and earth parameters in one struct to avoid type inference failure
	hydrology_earth_params =
	ClimaLand.Soil.HydrologyEarthParameters.(hydrology_cm, earth_param_set)

This line is allocating a new field at every step (on GPUs), which should be avoided, if possible.

[kmdeck]

@juliasloan25, since hydrology_cm and earth_param_set are both part of the model parameters, could we refactor this to pass model parameters instead and unpack them as needed? If the problem is passing two structs in broadcasted expression

[juliasloan25]

@juliasloan25, since hydrology_cm and earth_param_set are both part of the model parameters, could we refactor this to pass model parameters instead and unpack them as needed? If the problem is passing two structs in broadcasted expression

I think that should work! Then we'll just modify phase_change_source to take in the model parameters as its last arg and access hydrology_cm and earth_param_set from there rather than from hydrology_earth_params as is done now. I'll open a PR drafting the changes