[NN] Refine the API of GraphormerLayer (#5565)

Co-authored-by: rudongyu <ru_dongyu@outlook.com>

python/dgl/nn/pytorch/gt/graphormer.py

@@ -30,6 +30,8 @@ class GraphormerLayer(nn.Module):
         afterwards. Default: False.
     dropout : float, optional
         Dropout probability. Default: 0.1.
+    attn_dropout : float, optional
+        Attention dropout probability. Default: 0.1.
     activation : callable activation layer, optional
         Activation function. Default: nn.ReLU().
 
@@ -60,6 +62,7 @@ class GraphormerLayer(nn.Module):
         attn_bias_type="add",
         norm_first=False,
         dropout=0.1,
+        attn_dropout=0.1,
         activation=nn.ReLU(),
     ):
         super().__init__()
@@ -70,7 +73,7 @@ class GraphormerLayer(nn.Module):
             feat_size=feat_size,
             num_heads=num_heads,
             attn_bias_type=attn_bias_type,
-            attn_drop=dropout,
+            attn_drop=attn_dropout,
         )
         self.ffn = nn.Sequential(
             nn.Linear(feat_size, hidden_size),

tests/python/pytorch/nn/test_nn.py

@@ -2322,6 +2322,7 @@ def test_GraphormerLayer(attn_bias_type, norm_first):
         attn_bias_type=attn_bias_type,
         norm_first=norm_first,
         dropout=0.1,
+        attn_dropout=0.1,
         activation=th.nn.ReLU(),
     )
     out = net(nfeat, attn_bias, attn_mask)