KTO: decorate MLXKTOConfig @DataClass(init=False)

Bare @DataClass regenerated __init__, bypassing MLXTrainingConfig.__init__ and
its setup (e.g. _unsloth_mlx_warmup_steps_explicit). init=False inherits the
parent init while fields() still registers the KTO-specific fields, which the
parent __init__ reads via fields(type(self)). Mirrors unslothai#830's
MLXORPOConfig/MLXDPOConfig. Regression test asserts the inherited init and intact
KTO fields.

Author: BardiaKoopah

tests/test_mlx_kto_loss.py

@@ -324,3 +324,16 @@ def trainable_parameters(self):
     monkeypatch.setattr(T, "_kto_model_has_non_lora_trainable_params", lambda m: True)
     with pytest.raises(ValueError, match="structural limit"):
         tr.train()
+
+
+def test_kto_config_inherits_parent_init_not_a_generated_one():
+    # Bare @dataclass regenerates __init__ and bypasses MLXTrainingConfig.__init__
+    # (dropping e.g. _unsloth_mlx_warmup_steps_explicit). @dataclass(init=False)
+    # inherits the parent init while keeping the KTO-specific fields.
+    from unsloth_zoo.mlx.trainer import MLXKTOConfig, MLXTrainingConfig
+    assert MLXKTOConfig.__init__ is MLXTrainingConfig.__init__
+    c = MLXKTOConfig()
+    assert hasattr(c, "_unsloth_mlx_warmup_steps_explicit")  # parent init ran
+    assert c.beta == 0.1 and c.loss_type == "kto"  # KTO fields intact
+    c2 = MLXKTOConfig(beta=0.5, desirable_weight=2.0)
+    assert c2.beta == 0.5 and c2.desirable_weight == 2.0

unsloth_zoo/mlx/trainer.py

@@ -8155,10 +8155,16 @@ def train_on_responses_only(
 # ============================================================================
 
 
-@dataclass
+@dataclass(init=False)
 class MLXKTOConfig(MLXTrainingConfig):
     """KTO configuration mirroring TRL's KTOConfig field names, on top of the
-    shared MLX training knobs (optimizer, schedule, clipping, save)."""
+    shared MLX training knobs (optimizer, schedule, clipping, save).
+
+    init=False (not a bare @dataclass) so the subclass inherits
+    MLXTrainingConfig.__init__ rather than a generated one that would bypass the
+    parent's setup (e.g. the warmup-steps-explicit initialization). Mirrors
+    #830's MLXORPOConfig/MLXDPOConfig. fields() still registers the KTO-specific
+    fields below, which the inherited __init__ reads via fields(type(self))."""
 
     beta: float = 0.1
     desirable_weight: float = 1.0