[Big tensor] fix nan in cross_entropy

Author: hxzd5568

python/paddle/nn/functional/loss.py

@@ -3074,6 +3074,10 @@ def cross_entropy(
             #     numerator: loss's weighted sum
             #     denominator: cal the sum of weight where the sample's class_index!=ignore_index
             if ignore_index >= 0:  # ignore label
+                out_type = out.dtype
+                if out_type == paddle.float16:
+                    out = paddle.cast(out, dtype=paddle.float32)
+
                 out_sum = _C_ops.sum(out, [], None, False)
                 # for each label[i],set 1 or 0, according to ignore_index
                 # mask[i]=0, if label[i]==ignore_index
@@ -3093,6 +3097,7 @@ def cross_entropy(
                         weight_sum
                         + (weight_sum == 0.0).astype(weight_sum.dtype)
                     )
+                ret = paddle.cast(ret, dtype=out_type)
                 return ret
             elif weight is not None:
                 out_sum = _C_ops.sum(out, [], None, False)