fixed the bug raised by "tmp_eval_loss += tmp_eval_loss.item()" when parallelly using multi-gpu.

examples/run_ner.py

@@ -210,6 +210,9 @@ def evaluate(args, model, tokenizer, labels, pad_token_label_id, mode, prefix=""
             outputs = model(**inputs)
             tmp_eval_loss, logits = outputs[:2]
 
+            if args.n_gpu > 1:
+                tmp_eval_loss = tmp_eval_loss.mean()  # mean() to average on multi-gpu parallel evaluating
+
             eval_loss += tmp_eval_loss.item()
         nb_eval_steps += 1
         if preds is None: