remove unwanted code (#13145)

src/transformers/models/deberta_v2/modeling_deberta_v2.py

@@ -753,8 +753,6 @@ class DisentangledSelfAttention(nn.Module):
                 r_pos = relative_pos
 
             p2c_pos = torch.clamp(-r_pos + att_span, 0, att_span * 2 - 1)
-            if query_layer.size(-2) != key_layer.size(-2):
-                pos_index = relative_pos[:, :, :, 0].unsqueeze(-1)
 
         if "p2c" in self.pos_att_type:
             p2c_att = torch.bmm(key_layer, pos_query_layer.transpose(-1, -2))
@@ -763,12 +761,6 @@ class DisentangledSelfAttention(nn.Module):
                 dim=-1,
                 index=p2c_pos.squeeze(0).expand([query_layer.size(0), key_layer.size(-2), key_layer.size(-2)]),
             ).transpose(-1, -2)
-            if query_layer.size(-2) != key_layer.size(-2):
-                p2c_att = torch.gather(
-                    p2c_att,
-                    dim=-2,
-                    index=pos_index.expand(p2c_att.size()[:2] + (pos_index.size(-2), key_layer.size(-2))),
-                )
             score += p2c_att / scale
 
         # position->position
@@ -776,12 +768,6 @@ class DisentangledSelfAttention(nn.Module):
             pos_query = pos_query_layer[:, :, att_span:, :]
             p2p_att = torch.matmul(pos_query, pos_key_layer.transpose(-1, -2))
             p2p_att = p2p_att.expand(query_layer.size()[:2] + p2p_att.size()[2:])
-            if query_layer.size(-2) != key_layer.size(-2):
-                p2p_att = torch.gather(
-                    p2p_att,
-                    dim=-2,
-                    index=pos_index.expand(query_layer.size()[:2] + (pos_index.size(-2), p2p_att.size(-1))),
-                )
             p2p_att = torch.gather(
                 p2p_att,
                 dim=-1,