drop device setting for already parallel models

dmx_perplexity.py

@@ -40,6 +40,7 @@ Examples:
         46.05925369262695
 """
 
+
 @evaluate.utils.file_utils.add_start_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)
 class DmxPerplexity(evaluate.Metric):
     def _info(self):
@@ -89,9 +90,13 @@ class DmxPerplexity(evaluate.Metric):
             max_seq_len = model.config.n_positions
         else:
             max_seq_len = 2048
-            
-        if not hasattr(model, "hf_device_map"):
+
+        if not hasattr(model, "hf_device_map") and (
+            not hasattr(model, "model_parallel") or not model.model_parallel
+        ):
             model = model.to(device)
+
+        model.eval()
         encodings = tokenizer("\n\n".join(references), return_tensors="pt")
 
         stride = max_seq_len