@@ -269,6 +269,38 @@ class HuggingFaceDifferentialBackend: |
| 269 | 269 | finally: |
| 270 | 270 | self._exit() |
| 271 | 271 | |
| 272 | + @contextmanager |
| 273 | + def as_scaled_adapter(self, lam: float) -> Iterator[_HFView]: |
| 274 | + """Temporarily multiply every LoRA layer's scaling factor by ``lam``. |
| 275 | + |
| 276 | + Works by walking the PEFT module tree and mutating each |
| 277 | + ``LoraLayer.scaling[adapter_name]`` in place. The original |
| 278 | + scalings are restored when the context exits — or when an |
| 279 | + exception propagates, to keep the model in a sane state. |
| 280 | + """ |
| 281 | + self._enter(f"scaled({lam})") |
| 282 | + saved: list[tuple[object, str, float]] = [] |
| 283 | + try: |
| 284 | + import peft # noqa: PLC0415 — already a hard dep of this backend |
| 285 | + |
| 286 | + lora_cls = getattr(peft.tuners.lora, "LoraLayer", None) |
| 287 | + if lora_cls is None: |
| 288 | + raise RuntimeError("peft.tuners.lora.LoraLayer not found; check peft>=0.13 pin") |
| 289 | + for module in self._peft_model.modules(): |
| 290 | + if not isinstance(module, lora_cls): |
| 291 | + continue |
| 292 | + scaling = getattr(module, "scaling", None) |
| 293 | + if not isinstance(scaling, dict): |
| 294 | + continue |
| 295 | + for key, original in scaling.items(): |
| 296 | + saved.append((module, key, float(original))) |
| 297 | + scaling[key] = float(original) * lam |
| 298 | + yield self._make_view(f"scaled_{lam:.2f}") |
| 299 | + finally: |
| 300 | + for module, key, original in saved: |
| 301 | + module.scaling[key] = original # type: ignore[attr-defined] |
| 302 | + self._exit() |
| 303 | + |
| 272 | 304 | def close(self) -> None: |
| 273 | 305 | """Release GPU memory. Safe to call more than once.""" |
| 274 | 306 | if getattr(self, "_peft_model", None) is not None: |
