tenseleyflow/documentlanguagemodel / 8cb8a8c

         *,

         adapter_name: str | None = None,

     ) -> None:

         from dlm.inference.loader import resolve_adapter_path

         adapter_path = resolve_adapter_path(store, adapter_name=adapter_name)

         if not adapter_path.exists():

             raise AdapterNotFoundError(f"mlx backend: adapter dir {adapter_path} does not exist")

         # Stage both tensors + adapter_config.json into a scratch dir.

         # `stage_mlx_adapter_dir` performs the preflight PEFT-shape

         # check + translates PEFT config into mlx-lm's schema + writes

         with force_mps():

             caps = probe()

         with caplog.at_level(logging.WARNING, logger="dlm.hardware.plan"):  # type: ignore[attr-defined]

         assert plan.precision == "fp16"

         # The caller must see the risk explicitly — silent fp16 on MPS

         # is what caused the original bug.

             "system", "user", max_new_tokens=21, temperature=0.5, top_p=0.8, seed=11

         )

         assert out == "hf output"

         assert seen["runner"][3:] == (21, 0.5, 0.8, 11)