Add SoftmaxCrossEntropyLoss to mixed-precision-transformer. (microsoft#3760)

Author: codemzs

orttraining/orttraining/core/graph/mixed_precision_transformer.cc

@@ -33,7 +33,9 @@ namespace training {
 // continue to use 32-bit precision. Others will used reduced precision.
 static const std::unordered_set<std::string> FP32_Nodes = {
     "SparseSoftmaxCrossEntropy",
-    "SparseSoftmaxCrossEntropyGrad"};
+    "SparseSoftmaxCrossEntropyGrad",
+    "SoftmaxCrossEntropyLoss",
+    "SoftmaxCrossEntropyLossGrad"};
 
 bool IsFP32Node(const Node* node) {
   return FP32_Nodes.find(node->OpType()) != FP32_Nodes.cend();
@@ -54,6 +56,8 @@ static const std::unordered_map<std::string, std::vector<int>> stage2_fp32_node_
     {"DropoutGrad", {2}},
     {"SparseSoftmaxCrossEntropy", {0, 2}},
     {"SparseSoftmaxCrossEntropyGrad", {0, 1, 3}},
+    {"SoftmaxCrossEntropyLoss", {0, 2}},
+    {"SoftmaxCrossEntropyLossGrad", {0, 1, 3}},
 };
 
 bool IsFP32(const std::unordered_map<std::string, std::vector<int>>& map, std::string opname, int argnum) {