Complete the pre-norm transformer block recipe

The first blank is **attention** (or any modern variant: MHA, GQA, MLA). The second blank is the **feed-forward network**, commonly written FFN or MLP. Every transformer block has exactly two sublayers and they always appear in this order: 1. Attention: lets each token in the sequence look at every other token and gather context. 2. Feed-forward network: processes each token independently through a two step (or three step, for SwiGLU) MLP. The two blanks together describe the entire block. Around each sublayer, two things wrap it: a layer norm (or RMSNorm) that runs first on the input, and a residual connection that adds the sublayer's output to the original input. That is the entire recipe, repeated 32 times in Llama 3.1 8B, or 80 times in Llama 3.1 70B.

**Attention in the first slot, FFN in the second.** The full block code is: ``` x = x + Attention(LN(x)) x = x + FFN(LN(x)) ``` The first sublayer is attention. In a modern open weight LLM that means **GQA** (grouped-query attention) or **MLA** (multi-head latent attention); the original 2017 paper used full MHA. The job of this sublayer is cross-sequence mixing: every token gets to gather information from every other token (with causal masking in decoder-only models). The second sublayer is the **feed-forward network**, also called the MLP. In modern recipes it is a SwiGLU gated unit with three weight matrices and a hidden dimension of roughly `2.67 * d_model` to `3.5 * d_model`. The job of this sublayer is per-position computation: each token is processed independently through the same MLP, with no information flow between positions. The two sublayers do complementary jobs. Attention mixes information across the sequence. FFN computes per-token features. Alternating them is what gives transformers their expressive power; neither one alone is sufficient.