The Journey of a Token: What Actually Occurs Inside a Transformer

On this article, you’ll learn the way a transformer converts enter tokens into context-aware representations and, finally, next-token chances.

Matters we’ll cowl embody:

• How tokenization, embeddings, and positional data put together inputs
• What multi-headed consideration and feed-forward networks contribute inside every layer
• How the ultimate projection and softmax produce next-token chances

Let’s get our journey underway.

The Journey of a Token: What Actually Occurs Inside a Transformer (click on to enlarge)
Picture by Editor

The Journey Begins

Giant language fashions (LLMs) are based mostly on the transformer structure, a posh deep neural community whose enter is a sequence of token embeddings. After a deep course of — that appears like a parade of quite a few stacked consideration and feed-forward transformations — it outputs a likelihood distribution that signifies which token must be generated subsequent as a part of the mannequin’s response. However how can this journey from inputs to outputs be defined for a single token within the enter sequence?

On this article, you’ll be taught what occurs inside a transformer mannequin — the structure behind LLMs — on the token stage. In different phrases, we’ll see how enter tokens or components of an enter textual content sequence flip into generated textual content outputs, and the rationale behind the adjustments and transformations that happen contained in the transformer.

The outline of this journey by a transformer mannequin will likely be guided by the above diagram that exhibits a generic transformer structure and the way data flows and evolves by it.

Getting into the Transformer: From Uncooked Enter Textual content to Enter Embedding

Earlier than getting into the depths of the transformer mannequin, a couple of transformations already occur to the textual content enter, primarily so it’s represented in a type that’s absolutely comprehensible by the inner layers of the transformer.

Tokenization

The tokenizer is an algorithmic element usually working in symbiosis with the LLM’s transformer mannequin. It takes the uncooked textual content sequence, e.g. the consumer immediate, and splits it into discrete tokens (typically subword items or bytes, generally complete phrases), with every token within the supply language being mapped to an identifier i.

Token Embeddings

There’s a realized embedding desk E with form |V| × d (vocabulary measurement by embedding dimension). Trying up the identifiers for a sequence of size n yields an embedding matrix X with form n × d. That’s, every token identifier is mapped to a d-dimensional embedding vector that varieties one row of X. Two embedding vectors will likely be related to one another if they’re related to tokens which have related meanings, e.g. king and emperor, or vice versa. Importantly, at this stage, every token embedding carries semantic and lexical data for that single token, with out incorporating details about the remainder of the sequence (at the least not but).

Positional Encoding

Earlier than absolutely getting into the core components of the transformer, it’s essential to inject inside every token embedding vector — i.e. inside every row of the embedding matrix X — details about the place of that token within the sequence. That is additionally referred to as injecting positional data, and it’s usually accomplished with trigonometric features like sine and cosine, though there are methods based mostly on realized positional embeddings as effectively. An almost-residual element is summed to the earlier embedding vector e_t related to a token, as follows:

[
x_t^{(0)} = e_t + p_{text{pos}}(t)
]

with p_pos(t) usually being a trigonometric-based perform of the token place t within the sequence. In consequence, an embedding vector that previously encoded “what a token is” solely now encodes “what the token is and the place within the sequence it sits”. That is equal to the “enter embedding” block within the above diagram.

Now, time to enter the depths of the transformer and see what occurs inside!

Deep Contained in the Transformer: From Enter Embedding to Output Chances

Let’s clarify what occurs to every “enriched” single-token embedding vector because it goes by one transformer layer, after which zoom out to explain what occurs throughout all the stack of layers.

The components

[
h_t^{(0)} = x_t^{(0)}
]

is used to indicate a token’s illustration at layer 0 (the primary layer), whereas extra generically we’ll use ht(l) to indicate the token’s embedding illustration at layer l.

Multi-headed Consideration

The primary main element inside every replicated layer of the transformer is the multi-headed consideration. That is arguably essentially the most influential element in all the structure with regards to figuring out and incorporating into every token’s illustration quite a lot of significant details about its position in all the sequence and its relationships with different tokens within the textual content, be it syntactic, semantic, or another type of linguistic relationship. A number of heads on this so-called consideration mechanism are every specialised in capturing totally different linguistic elements and patterns within the token and all the sequence it belongs to concurrently.

The results of having a token illustration ht(l) (with positional data injected a priori, don’t neglect!) touring by this multi-headed consideration inside a layer is a context-enriched or context-aware token illustration. Through the use of residual connections and layer normalizations throughout the transformer layer, newly generated vectors turn out to be stabilized blends of their very own earlier representations and the multi-headed consideration output. This helps enhance coherence all through all the course of, which is utilized repeatedly throughout layers.

Feed-forward Neural Community

Subsequent comes one thing comparatively much less advanced: a couple of feed-forward neural community (FFN) layers. As an example, these may be per-token multilayer perceptrons (MLPs) whose purpose is to additional rework and refine the token options which can be steadily being realized.

The primary distinction between the eye stage and this one is that focus mixes and incorporates, in every token illustration, contextual data from throughout all tokens, however the FFN step is utilized independently on every token, refining the contextual patterns already built-in to yield helpful “data” from them. These layers are additionally supplemented with residual connections and layer normalizations, and because of this course of, we’ve got on the finish of a transformer layer an up to date illustration ht(l+1) that may turn out to be the enter to the subsequent transformer layer, thereby getting into one other multi-headed consideration block.

The entire course of is repeated as many occasions because the variety of stacked layers outlined in our structure, thus progressively enriching the token embedding with increasingly more higher-level, summary, and long-range linguistic data behind these seemingly indecipherable numbers.

Ultimate Vacation spot

So, what occurs on the very finish? On the high of the stack, after going by the final replicated transformer layer, we get hold of a remaining token illustration ht*(L) (the place t* denotes the present prediction place) that’s projected by a linear output layer adopted by a softmax.

The linear layer produces unnormalized scores referred to as logits, and the softmax converts these logits into next-token chances.

Logits computation:

[
text{logits}_j = W_{text{vocab}, j} cdot h_{t^*}^{(L)} + b_j
]

Making use of softmax to calculate normalized chances:

[
text{softmax}(text{logits})_j = frac{exp(text{logits}_j)}{sum_{k} exp(text{logits}_k)}
]

Utilizing softmax outputs as next-token chances:

[
P(text{token} = j) = text{softmax}(text{logits})_j
]

These chances are calculated for all attainable tokens within the vocabulary. The following token to be generated by the LLM is then chosen — typically the one with the best likelihood, although sampling-based decoding methods are additionally widespread.

Journey’s Finish

This text took a journey, with a mild stage of technical element, by the transformer structure to offer a common understanding of what occurs to the textual content that’s supplied to an LLM — essentially the most distinguished mannequin based mostly on a transformer structure — and the way this textual content is processed and reworked contained in the mannequin on the token stage to lastly flip right into a mannequin’s output: the subsequent phrase to generate.

We hope you’ve loved our travels collectively, and we sit up for the chance to embark upon one other within the close to future.