Inside a Transformer Block: An Operational Anatomy

Exact ordering differs across implementations, and the difference is not cosmetic. Where the layer normalization sits decides whether learning-rate warm-up is needed at all. Xiong and nine colleagues showed this in 2020 with a mean-field analysis. Put the normalization after the sublayer — the Post-LN arrangement — and the expected gradients near the output layer are large at initialization. Put it inside the residual block instead and, as they write, “On the other hand, our theory also shows that if the layer normalization is put inside the residual blocks (recently proposed as Pre-LN Transformer), the gradients are well-behaved at initialization.” With Pre-LN the warm-up stage can be removed. The ordering is a documented engineering choice with a measured training consequence, not an implementation detail.

The recurring components keep stable roles whichever ordering is used. A normalization scheme controls scale before or after a sublayer. Attention then routes information: queries score keys and combine the corresponding values under a mask. The sublayer update is added to the incoming representation along the residual path rather than replacing it. A feed-forward network mixes features position-wise using parameters shared across positions. The next block receives states that now contain both local and routed information.

Each token state is projected into query, key, and value vectors. A query-key score determines how strongly one position uses a value from another allowed position. These names describe computational roles. A query vector is not a conscious question, and a key is not a symbolic database key with guaranteed meaning.

The vocabulary is portable enough that the same words have been applied to a different sublayer entirely. The position-wise feed-forward sublayers themselves behave as key-value memories, with their outputs refined across layers through the residual connections. Geva and colleagues argued that in 2021, and their abstract opens: “Feed-forward layers constitute two-thirds of a transformer model's parameters, yet their role in the network remains under-explored.” Two consequences follow. First, "key" and "value" are labels for what a computation does in a given place, not fixed objects with a single meaning across the architecture. Second, if two-thirds of the parameters live in the mixing step, then studying only the routing step means studying the smaller share of the block.

Attention cannot use a position that the mask or architecture makes unavailable. The causal mask is the clearest case, and it is worth seeing how literally it is implemented. The 2017 paper describes it in two sentences: “We need to prevent leftward information flow in the decoder to preserve the auto-regressive property. We implement this inside of scaled dot-product attention by masking out (setting to −∞) all values in the input of the softmax which correspond to illegal connections.” The restriction is arithmetic, not structural. A value of −∞ entering the softmax leaves a weight of zero. It supports autoregressive generation and prevents direct leakage from future target tokens, while leaving all permitted earlier context available. It guarantees nothing about sensible time reasoning.

A padding mask does a different job. Variable-length batches need artificial padding positions excluded from useful attention so that padding does not contaminate states. It depends on correct sequence boundaries, and it does not solve truncation.

A structured mask restricts interactions by role or region, and the best-measured instance is BigBird: local windows, random links and O(1) global tokens, which reduces the quadratic dependency on sequence length to linear. Zaheer and colleagues published it in 2020, and the abstract states: “The proposed sparse attention can handle sequences of length up to 8x of what was previously possible using similar hardware.” That is what a structured mask buys. The cost comes with it: a pattern chosen in advance can block useful evidence if designed poorly, so it requires task-specific evaluation.

A residual connection adds a sublayer update to its input instead of replacing the state entirely. That helps the optimizer, and it lets later layers refine information while retaining a direct path. The idea did not originate in the Transformer. He and colleagues introduced it for image recognition in 2016: “We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions.” Vaswani et al. cite that paper for the residual connection placed around each sublayer.

The payoff was depth that had not previously been trainable: networks up to 152 layers deep. An ensemble of them reached 3.57% error on the ImageNet test set and won ILSVRC 2015. That is the evidence behind "helps the optimizer" — a specific depth and a specific error rate, not a general reassurance.

Residuals still do not guarantee that every useful feature survives. Scale, normalization, depth, and learned transformations all affect which information remains accessible further up the stack.

These tools are evidence about what the block computed. None is a complete explanation by itself, and the interventions carry more weight than the pictures.

How much more is measurable. Take a WMT English–French Transformer with 6 layers × 16 heads, that is 96 encoder self-attention heads, and remove heads at test time. Only 8 of the 96 caused a statistically significant change in performance — and half of those changes were a higher BLEU score. Iterative pruning went further: “We observe that this approach allows us to prune up to 20% and 40% of heads from WMT and BERT (respectively), without incurring any noticeable negative impact.” Michel and colleagues ran that experiment in 2019, under the title "Are Sixteen Heads Really Better than One?". A head that looks busy in a heatmap can be one of the ninety-odd whose removal changes nothing measurable.

Use a short sequence so every tensor and mask can be inspected. Five passes are enough. Each one turns a claim from this lesson into something you can read off your own screen.

First, record the tensor shapes: batch, sequence, head, and feature dimensions. On the base configuration of "Attention Is All You Need" those last two are eight heads of 64 inside a model dimension of 512. Checking that 8 × 64 = 512 on your own tensors confirms the heads partition the block rather than widening it. Second, display the mask and verify exactly which positions can interact. Before the softmax the disallowed entries should be the −∞ the original paper describes, and after it, zeros. Third, inspect the attention scores for one head, comparing pre-softmax and normalized values. Fourth, measure the residual updates: compare the norm of each sublayer update with the norm of the incoming state. That is what "adds rather than replaces" looks like numerically. Fifth, perturb one token and test whether routing and output change in a plausible direction. For the stronger version of that test, delete the head entirely and measure the behavior change, the way the pruning experiments did.