Deepspeed example should use gather_for_metrics (#821)

* Deepspeed example should use gather_for_metrics

I believe this example should be using gather_for_metrics here instead of gather.

* Update deepspeed_with_config_support.py

examples/by_feature/deepspeed_with_config_support.py

@@ -285,10 +285,9 @@ def evaluate(args, model, eval_dataloader, accelerator, eval_dataset):
             outputs = model(**batch)
 
         loss = outputs.loss
-        losses.append(accelerator.gather(loss.repeat(args.per_device_eval_batch_size)))
+        losses.append(accelerator.gather_for_metrics(loss.repeat(args.per_device_eval_batch_size)))
 
     losses = torch.cat(losses)
-    losses = losses[: len(eval_dataset)]
     try:
         eval_loss = torch.mean(losses)
         perplexity = math.exp(eval_loss)