fix(mlx): keep the PEFT export path free of torch imports

Exporting an embedding LoRA without base_weights_source logged through the
shared logger, which resolves on first use by importing a module that imports
torch. saving_utils is otherwise importable and usable without torch, so that
export raised ModuleNotFoundError on a torch-free Apple install. Drop the
informational log; the behaviour it described is unchanged.

Author: Lyxot

unsloth_zoo/saving_utils.py

@@ -6475,13 +6475,10 @@ def _unnest_map(mapping, what):
                     "source does not match the adapter's base model."
                 )
             out[f"{_PEFT_PREFIX}{_p}.base_layer.weight"] = _w
-    elif _unsourced_emb:
-        logger.info(
-            "Unsloth MLX: exporting embedding LoRA for "
-            f"{_unsourced_emb} without base embedding weights (no "
-            "base_weights_source); PEFT loads such adapters cleanly and "
-            "takes the embeddings from the base model."
-        )
+    # An embedding adapter exported without base_weights_source simply omits
+    # those weights; peft resolves them from the base model. Nothing is logged
+    # here on purpose -- this module stays importable and usable without torch,
+    # and the shared logger pulls in a torch-importing module on first use.
     # Same stage-then-atomic-rename destination contract as
     # convert_peft_dir_to_mlx; see the note there.
     dst = os.fspath(dst)