Nilearn glm when stimulus onsets are not locked to TRs

How will nilearn.glm.first_level.FirstLevelModel handle an events file that is not locked to the TR? There are many different opinions on what to do in this case and I could not find documentation for a) whether this is supported in nilearn and b) what nilearn does when events occur at times that are not locked to the TR.

Thanks to @Jeanette_Mumford for this explanation (which is now clear if you go down multiple levels in the documentation).
By default, NiLearn creates the design matrix at a higher temporal resolution than the TR (default is sub sampling by 50). The convolutions are done at this higher temporal resolution and then it is resampled to the TR.
Regressors are created in compute_regressor and the oversampling is used in _sample_condition.