diff --git a/ludwig/distributed/deepspeed.py b/ludwig/distributed/deepspeed.py
index 12986f2afdb..9b3a04b5135 100644
--- a/ludwig/distributed/deepspeed.py
+++ b/ludwig/distributed/deepspeed.py
@@ -221,7 +221,14 @@ def load(self, save_path: str, device: Optional[torch.device] = None) -> bool:
 
         https://deepspeed.readthedocs.io/en/latest/model-checkpointing.html#loading-training-checkpoints
         """
-        _, client_state = self.model.load_checkpoint(save_path, load_lr_scheduler_states=False)
+        # NOTE(geoffrey): `load_module_strict=False` because this code path is frequently used to load models trained
+        # using adapter-based fine-tuning, where the checkpoints only contain the adapter weights, and not the full
+        # model weights. This may lead to silent, unexpected behavior for resuming full model fine-tuning,
+        # where all the model weights *must* be loaded in.
+        # TODO(geoffrey): Add a boolean arg to function to control load_module_strict behavior.
+        _, client_state = self.model.load_checkpoint(
+            save_path, load_lr_scheduler_states=False, load_module_strict=False
+        )
         self.global_step = self._get_global_step(client_state, save_path)
         if self.scheduler is not None and "scheduler_state" in client_state:
             self.scheduler.load_state_dict(client_state["scheduler_state"])