Don't cache reinit_modules

CHANGELOG.md

@@ -26,6 +26,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Removed a spurious error message "'torch.cuda' has no attribute '_check_driver'" that would be appear in the logs
   when a `ConfigurationError` for missing GPU was raised.
 - Load model on CPU post training to save GPU memory.
+- Don't cache models with `cached_transformers` when `reinit_modules` is not `None`.
 - Fixed a bug in `ShouldValidateCallback` that leads to valuation occuring after the first epoch regardless of `validation_start` value.
 - Fixed a bug in `ShouldValidateCallback` that leads to valuation occuring every `validation_interval + 1` epochs, instead of every `validation_interval` epochs.
 

allennlp/common/cached_transformers.py

@@ -14,7 +14,6 @@ class TransformerSpec(NamedTuple):
     model_name: str
     override_weights_file: Optional[str] = None
     override_weights_strip_prefix: Optional[str] = None
-    reinit_modules: Optional[Union[int, Tuple[int, ...], Tuple[str, ...]]] = None
 
 
 _model_cache: Dict[TransformerSpec, transformers.PreTrainedModel] = {}
@@ -66,9 +65,8 @@ def get(
         model_name,
         override_weights_file,
         override_weights_strip_prefix,
-        reinit_modules,
     )
-    transformer = _model_cache.get(spec, None)
+    transformer = None if reinit_modules is not None else _model_cache.get(spec, None)
     if transformer is None:
         if not load_weights:
             if override_weights_file is not None: