My understanding is that @percyliang
's prefix tuning approach prepended activations to *every* layer, but yours only prepends to the first layer — is that right?

→ View original post on X — @jeremyphoward