[LLM] fix llama precision on custom devices (#7895)

paddlenlp/transformers/llama/modeling.py

@@ -1489,7 +1489,7 @@ def forward(self, prediction_scores, masked_lm_labels):
                 _hcg = fleet.get_hybrid_communicate_group()
                 masked_lm_loss = ConcatSePMaskedLoss.apply(masked_lm_loss, axis=1, group=_hcg.get_sep_parallel_group())
             # skip ignore_index which loss == 0
-            masked_lm_loss = masked_lm_loss[masked_lm_loss > 0].astype("float32")
+            masked_lm_loss = masked_lm_loss[masked_lm_loss > 0]
             loss = paddle.mean(masked_lm_loss)
 
         return loss