fix for MCore dist ckpt loading (#14229)

Signed-off-by: He Huang (Steve) <105218074+stevehuang52@users.noreply.github.com>

Author: stevehuang52

nemo/collections/speechlm/models/speech_to_text_llm_model.py

@@ -195,7 +195,7 @@ def _freeze_module(self, module: Optional[nn.Module] = None) -> None:
         for param in module.parameters():
             param.requires_grad = False
 
-    def _maybe_load_pretrained_llm(self, model: MCoreGPTModel) -> MCoreGPTModel:
+    def _maybe_load_pretrained_llm(self, model: MCoreGPTModel, strict: bool = False) -> MCoreGPTModel:
         if not self.language_model_from_pretrained:
             return model
 
@@ -224,6 +224,7 @@ def _maybe_load_pretrained_llm(self, model: MCoreGPTModel) -> MCoreGPTModel:
             sharded_state_dict=sharded_state_dict,
             checkpoint_dir=ckpt_to_weights_subdir(ckpt_path, is_saving=False),
             validate_access_integrity=False,
+            **({"strict": "log_all"} if not strict else {}),
         )
         loaded_state_dict = {k.removeprefix("module."): v for k, v in loaded_state_dict["state_dict"].items()}
         model.load_state_dict(loaded_state_dict)