Fix sft_trainer crash when adapter_file is None (#908)

[H-A-Khan]

What

sft_trainer.train (and the analogous orpo_trainer.train_orpo) crash on the first checkpoint or final save if args.adapter_file is None:

File "mlx_vlm/trainer/sft_trainer.py", line 471, in train
    save_adapter(model, args.adapter_file)
File "mlx_vlm/trainer/utils.py", line 235, in save_adapter
    path = Path(adapter_file)
TypeError: argument should be a str or an os.PathLike object
where __fspath__ returns a str, not 'NoneType'

This happens when callers construct TrainingArgs (or pass a SimpleNamespace) without explicitly setting adapter_file. The dataclass default is "adapters.safetensors", but if a caller overrides that to None — or if the field is shadowed during serialisation — the train loop runs to the first save, then crashes and loses the work.

Closes #908.

Fix

Defend at the top of train / train_orpo. If args.adapter_file is None, set it to the dataclass default (adapters.safetensors) and emit a warning on rank 0 so the substitution is visible. Covers all three call sites (steps_per_save checkpoint, indexed checkpoint filename, final save) with one guard per trainer.

Changes

• mlx_vlm/trainer/sft_trainer.py: defensive guard at top of train().
• mlx_vlm/trainer/orpo_trainer.py: same guard at top of train_orpo() (the same bug exists in the ORPO trainer).
• mlx_vlm/tests/test_trainer.py: regression test test_train_handles_none_adapter_file that constructs TrainingArgs(...), sets adapter_file = None, runs a single-iteration training loop, and asserts both that the default is applied and that save_safetensors is still invoked.

Tests

python -m pytest mlx_vlm/tests/test_trainer.py -v

test_dataset_getitem PASSED
test_dataset_initialization PASSED
test_train_handles_none_adapter_file PASSED
test_train_smoke PASSED
test_trainer_initialization PASSED
======================== 5 passed in 3.84s =========================