chore: import upstream snapshot with attribution

decoding/GAD/fairseq/modules/fp32_group_norm.py

@@ -0,0 +1,25 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+"""
+Layer norm done in fp32 (for fp16 training)
+"""
+
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class Fp32GroupNorm(nn.GroupNorm):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def forward(self, input):
+        output = F.group_norm(
+            input.float(),
+            self.num_groups,
+            self.weight.float() if self.weight is not None else None,
+            self.bias.float() if self.bias is not None else None,
+            self.eps,
+        )
+        return output.type_as(input)