diff --git a/torchscale/component/multiway_network.py b/torchscale/component/multiway_network.py
index a44a699e..63f3cffc 100644
--- a/torchscale/component/multiway_network.py
+++ b/torchscale/component/multiway_network.py
@@ -45,11 +45,14 @@ def forward(self, x, **kwargs):
         return torch.cat([y1, y2], dim=self.dim)
 
 
-class MutliwayEmbedding(MultiwayNetwork):
+class MultiwayEmbedding(MultiwayNetwork):
     def __init__(self, modules, dim=1):
         super(MultiwayNetwork, self).__init__()
         self.dim = dim
         assert len(modules) == 2
         self.A = modules[0]
         self.B = modules[1]
-        self.split_position = -1
\ No newline at end of file
+        self.split_position = -1
+
+# Backwards-compatible alias for the previous misspelling.
+MutliwayEmbedding = MultiwayEmbedding
diff --git a/torchscale/model/BEiT3.py b/torchscale/model/BEiT3.py
index 92737a21..eaefb7ab 100644
--- a/torchscale/model/BEiT3.py
+++ b/torchscale/model/BEiT3.py
@@ -10,7 +10,7 @@
     TextEmbedding,
     VisionEmbedding,
 )
-from torchscale.component.multiway_network import MutliwayEmbedding
+from torchscale.component.multiway_network import MultiwayEmbedding
 
 
 class BEiT3(nn.Module):
@@ -30,7 +30,7 @@ def __init__(self, args, **kwargs):
             prepend_cls_token=True,
         )
         # being consistent with Fairseq, which starts from 2 for position embedding
-        embed_positions = MutliwayEmbedding(
+        embed_positions = MultiwayEmbedding(
             modules=[
                 PositionalEmbedding(self.vision_embed.num_position_embeddings() + 2, args.encoder_embed_dim),
                 PositionalEmbedding(args.max_source_positions, args.encoder_embed_dim),