Update dmx_perplexity.py

supporting model parallel for gpt

dmx_perplexity.py

@@ -105,6 +105,13 @@ class DmxPerplexity(evaluate.Metric):
             target_ids = input_ids.clone()
             target_ids[:, :-trg_len] = -100
 
+            # Setting device for labels if mdodel has device mapping
+            if hasattr(model, "hf_device_map"):
+                last_device = "cuda:" + str(
+                    max(model.hf_device_map.values())
+                )
+                target_ids = target_ids.to(last_device)
+
             with torch.no_grad():
                 outputs = model(input_ids, labels=target_ids)
                 if isinstance(outputs, Dict):