input attention seems using wrong projection matrix

[sentient-codebot]

        if not share_query_proj:
            # giving each query vector different projection matrix (one-to-one)
            self.query = GroupLinearLayer(
                hidden_size, kdim * num_heads, num_hidden)
        else:
            # giving each query vector different projection matrix (one-to-one)
            self.query = SharedGroupLinearLayer(
                hidden_size, kdim * num_heads, num_hidden)
            # all query share the same projection *proj*

SharedGroupLinearLayer is not what's intended.