`498b85e`

backends/hf: drop duplicate adapter-path normalization, lean on ModelSpec validator (B22)

Authored by

espadonne 3 weeks ago

Status	File	+	-
M	`src/dlm_sway/backends/hf.py`	6	1

src/dlm_sway/backends/hf.pymodified

          torch, transformers, peft = _require_hf()
          self._torch = torch
          self._spec = base_spec
 -        self._adapter_path = Path(adapter_path).expanduser().resolve()
 +        # Path normalization lives in ``ModelSpec.adapter`` (B22). When
 +        # the backend is constructed via ``backends.build``, the value
 +        # is already absolute. Direct constructions (some tests) may
 +        # pass a relative path, so ``Path(...).resolve()`` stays as a
 +        # cheap idempotent fallback.
 +        self._adapter_path = Path(adapter_path).resolve()
          device_str: Device = (
              _detect_device() if base_spec.device == "auto" else base_spec.device  # type: ignore[assignment]