Summary

sqlmodel_validate (which backs SQLModel.model_validate and the deprecated from_orm/parse_obj) sets relationships with attribute access:

for key in new_obj.__sqlmodel_relationships__:
    value = getattr(use_obj, key, Undefined)
    if value is not Undefined:
        setattr(new_obj, key, value)

But use_obj is a plain dict in the common case: it is the input obj when a dict is passed, or {**obj, **update} when update= is given. getattr(some_dict, "team", Undefined) never reads dict keys, so it always returns Undefined and the relationship is silently dropped, no error.

This makes model_validate inconsistent with both the constructor and model_validate(object):

team = Team(name="Avengers")
Hero(name="IronMan", team=team).team is team                         # True
Hero.model_validate({"name": "Thor", "team": team}).team is team     # False (dropped)
Hero.model_validate(SimpleNamespace(name="Loki", team=team)).team    # works (object -> getattr)
Hero.model_validate({"name": "Hulk"}, update={"team": team}).team    # False (merged dict, dropped)

The sibling sqlmodel_table_construct already handles this correctly with values.get(key, Undefined).

Fix

Read the relationship via dict.get when use_obj is a dict, falling back to getattr otherwise (mirroring sqlmodel_table_construct). This also fixes the update= path.

Test

Added test_validate_dict_sets_relationship: a relationship passed in the dict to model_validate (and via update=) must be set, matching the constructor. It fails before the change and passes after. The existing validation tests only pass scalar fields, so this path was uncovered.