Forgot one: the positional encoding also changed, llama3 uses RoPE, gpt2 uses a learned embedding.