There are basically two ways to interpret "temperature" in this context. You may see that there is an temperature for each electrode in a device calculation (and in fact also one for regular bulk calculations). These refer to the electron temperature, or more specifically the width of the Fermi distribution used to compute the occupation of eigenstates. It is easy to change this in a device to e.g. consider thermoelectric effects.
This, finite, temperature is to some extent a numerical trick to be able to use fewer k-points. If the Fermi distribution were a sharp step function, it would be very hard to reach convergence in the self-consistent cycle without a huge number of k-points. Formally this means that every calculation should be extrapolated to zero electron temperature by running successive calculations at lower and lower electron temperature, restarting from the higher temperature in each step to help convergence. In practice this is usually overlooked, and typically not that important :-)
Note however that the default electron temperature is QuantumATK is quite high (1000 K) to help convergence, in particular for metals, but it often leads to a Fermi level that is a bit offset from the center of the band gap where it should be in our definition of zero energy. So for insulators and semiconductors it is strongly recommended to lower the default to 200-300 K to avoid incorrect results.
Separately, if you want to consider the lattice temperature, things get more complex fast. Now you have to introduce phonons, or lattice vibrations, to move the atoms around from their equilibrium positions. There is a special technique in QuantumATK for this, see
https://docs.quantumatk.com/manual/Types/SpecialThermalDisplacement/SpecialThermalDisplacement.html, which perhaps is a bit unknown, but can be very powerful. It has been used to compute e.g. the dielectric constant or band gap as a function of temperature. A straightforward but more time-consuming approach is to use molecular dynamics (MD) to evolve the geometry at finite temperature over time, and compute the property of interest for a sequence of time-steps, and average. The result should be equivalent to the STD method, as has been verified for e.g. temperature-dependent conductance of metal nanowires.
