diff --git a/src/weathergen/model/model_interface.py b/src/weathergen/model/model_interface.py
index 23334749b..528db56d9 100644
--- a/src/weathergen/model/model_interface.py
+++ b/src/weathergen/model/model_interface.py
@@ -189,13 +189,17 @@ def load_model(cf, model, device, run_id: str, mini_epoch=-1):
         maybe_sharded_sd = {}
         for param_name, full_tensor in params.items():
             sharded_meta_param = meta_sharded_sd.get(param_name)
-            sharded_tensor = distribute_tensor(
-                full_tensor,
-                sharded_meta_param.device_mesh,
-                sharded_meta_param.placements,
-            )
-            # maybe_sharded_sd[param_name.replace("module.", "")] = nn.Parameter(sharded_tensor)
-            maybe_sharded_sd[param_name] = torch.nn.Parameter(sharded_tensor)
+            if sharded_meta_param is None:
+                logger.info(f"Sharding meta parameters is None for: {param_name}")
+                maybe_sharded_sd[param_name] = torch.nn.Parameter(full_tensor)
+            else:
+                sharded_tensor = distribute_tensor(
+                    full_tensor,
+                    sharded_meta_param.device_mesh,
+                    sharded_meta_param.placements,
+                )
+                # maybe_sharded_sd[param_name.replace("module.", "")] = nn.Parameter(sharded_tensor)
+                maybe_sharded_sd[param_name] = torch.nn.Parameter(sharded_tensor)
         # choose `assign=True` for sharded model since we cannot call `copy_` on meta tensor
         mkeys, ukeys = model.load_state_dict(maybe_sharded_sd, strict=False, assign=True)