Distributed Trainer: 2 little fixes (#7461)

* reset model.config

* Update src/transformers/trainer.py

* use lower case tensor

* Just tensor change

src/transformers/trainer_utils.py

@@ -203,7 +203,7 @@ def distributed_broadcast_scalars(
 ) -> "torch.Tensor":
     if is_torch_available():
         try:
-            tensorized_scalar = torch.Tensor(scalars).cuda()
+            tensorized_scalar = torch.tensor(scalars).cuda()
             output_tensors = [tensorized_scalar.clone() for _ in range(torch.distributed.get_world_size())]
             torch.distributed.all_gather(output_tensors, tensorized_scalar)
             concat = torch.cat(output_tensors, dim=0)