Massive Language Fashions (LLMs) have revolutionized the sphere of pure language processing (NLP) by demonstrating exceptional capabilities in producing human-like textual content, answering questions, and aiding with a variety of language-related duties. On the core of those highly effective fashions lies the decoder-only transformer structure, a variant of the unique transformer structure proposed within the seminal paper “Consideration is All You Want” by Vaswani et al.

On this complete information, we’ll discover the inside workings of decoder-based LLMs, delving into the basic constructing blocks, architectural improvements, and implementation particulars which have propelled these fashions to the forefront of NLP analysis and functions.

The Transformer Structure: A Refresher

Earlier than diving into the specifics of decoder-based LLMs, it is important to revisit the transformer structure, the muse upon which these fashions are constructed. The transformer launched a novel strategy to sequence modeling, relying solely on consideration mechanisms to seize long-range dependencies within the knowledge, with out the necessity for recurrent or convolutional layers.

The unique transformer structure consists of two most important elements: an encoder and a decoder. The encoder processes the enter sequence and generates a contextualized illustration, which is then consumed by the decoder to supply the output sequence. This structure was initially designed for machine translation duties, the place the encoder processes the enter sentence within the supply language, and the decoder generates the corresponding sentence within the goal language.

Self-Consideration: The Key to Transformer’s Success

On the coronary heart of the transformer lies the self-attention mechanism, a robust method that permits the mannequin to weigh and mixture data from totally different positions within the enter sequence. In contrast to conventional sequence fashions, which course of enter tokens sequentially, self-attention allows the mannequin to seize dependencies between any pair of tokens, no matter their place within the sequence.

The self-attention operation may be damaged down into three most important steps:

1. Question, Key, and Worth Projections: The enter sequence is projected into three separate representations: queries (Q), keys (Ok), and values (V). These projections are obtained by multiplying the enter with discovered weight matrices.
2. Consideration Rating Computation: For every place within the enter sequence, consideration scores are computed by taking the dot product between the corresponding question vector and all key vectors. These scores symbolize the relevance of every place to the present place being processed.
3. Weighted Sum of Values: The eye scores are normalized utilizing a softmax perform, and the ensuing consideration weights are used to compute a weighted sum of the worth vectors, producing the output illustration for the present place.

Multi-head consideration, a variant of the self-attention mechanism, permits the mannequin to seize various kinds of relationships by computing consideration scores throughout a number of “heads” in parallel, every with its personal set of question, key, and worth projections.

Architectural Variants and Configurations

Whereas the core rules of decoder-based LLMs stay constant, researchers have explored numerous architectural variants and configurations to enhance efficiency, effectivity, and generalization capabilities. On this part, we’ll delve into the totally different architectural selections and their implications.

Structure Sorts

Decoder-based LLMs may be broadly categorized into three most important varieties: encoder-decoder, causal decoder, and prefix decoder. Every structure kind displays distinct consideration patterns, as illustrated in Determine 1.

Encoder-Decoder Structure

Based mostly on the vanilla Transformer mannequin, the encoder-decoder structure consists of two stacks: an encoder and a decoder. The encoder makes use of stacked multi-head self-attention layers to encode the enter sequence and generate latent representations. The decoder then performs cross-attention on these representations to generate the goal sequence. Whereas efficient in numerous NLP duties, few LLMs, corresponding to Flan-T5, undertake this structure.

Causal Decoder Structure

The causal decoder structure incorporates a unidirectional consideration masks, permitting every enter token to attend solely to previous tokens and itself. Each enter and output tokens are processed throughout the similar decoder. Notable fashions like GPT-1, GPT-2, and GPT-3 are constructed on this structure, with GPT-3 showcasing exceptional in-context studying capabilities. Many LLMs, together with OPT, BLOOM, and Gopher, have extensively adopted causal decoders.

Prefix Decoder Structure

Often known as the non-causal decoder, the prefix decoder structure modifies the masking mechanism of causal decoders to allow bidirectional consideration over prefix tokens and unidirectional consideration on generated tokens. Just like the encoder-decoder structure, prefix decoders can encode the prefix sequence bidirectionally and predict output tokens autoregressively utilizing shared parameters. LLMs primarily based on prefix decoders embody GLM130B and U-PaLM.

All three structure varieties may be prolonged utilizing the mixture-of-experts (MoE) scaling method, which sparsely prompts a subset of neural community weights for every enter. This strategy has been employed in fashions like Change Transformer and GLaM, with growing the variety of consultants or whole parameter dimension exhibiting important efficiency enhancements.

Decoder-Solely Transformer: Embracing the Autoregressive Nature

Whereas the unique transformer structure was designed for sequence-to-sequence duties like machine translation, many NLP duties, corresponding to language modeling and textual content technology, may be framed as autoregressive issues, the place the mannequin generates one token at a time, conditioned on the beforehand generated tokens.

Enter the decoder-only transformer, a simplified variant of the transformer structure that retains solely the decoder element. This structure is especially well-suited for autoregressive duties, because it generates output tokens one after the other, leveraging the beforehand generated tokens as enter context.

The important thing distinction between the decoder-only transformer and the unique transformer decoder lies within the self-attention mechanism. Within the decoder-only setting, the self-attention operation is modified to forestall the mannequin from attending to future tokens, a property referred to as causality. That is achieved via a way known as “masked self-attention,” the place consideration scores comparable to future positions are set to unfavorable infinity, successfully masking them out through the softmax normalization step.

Architectural Elements of Decoder-Based mostly LLMs

Whereas the core rules of self-attention and masked self-attention stay the identical, fashionable decoder-based LLMs have launched a number of architectural improvements to enhance efficiency, effectivity, and generalization capabilities. Let’s discover a few of the key elements and strategies employed in state-of-the-art LLMs.

Enter Illustration

Earlier than processing the enter sequence, decoder-based LLMs make use of tokenization and embedding strategies to transform the uncooked textual content right into a numerical illustration appropriate for the mannequin.

Tokenization: The tokenization course of converts the enter textual content right into a sequence of tokens, which may be phrases, subwords, and even particular person characters, relying on the tokenization technique employed. Well-liked tokenization strategies for LLMs embody Byte-Pair Encoding (BPE), SentencePiece, and WordPiece. These strategies intention to strike a stability between vocabulary dimension and illustration granularity, permitting the mannequin to deal with uncommon or out-of-vocabulary phrases successfully.

Token Embeddings: After tokenization, every token is mapped to a dense vector illustration known as a token embedding. These embeddings are discovered through the coaching course of and seize semantic and syntactic relationships between tokens.

Positional Embeddings: Transformer fashions course of the complete enter sequence concurrently, missing the inherent notion of token positions current in recurrent fashions. To include positional data, positional embeddings are added to the token embeddings, permitting the mannequin to differentiate between tokens primarily based on their positions within the sequence. Early LLMs used mounted positional embeddings primarily based on sinusoidal capabilities, whereas more moderen fashions have explored learnable positional embeddings or different positional encoding strategies like rotary positional embeddings.

Multi-Head Consideration Blocks

The core constructing blocks of decoder-based LLMs are multi-head consideration layers, which carry out the masked self-attention operation described earlier. These layers are stacked a number of instances, with every layer attending to the output of the earlier layer, permitting the mannequin to seize more and more complicated dependencies and representations.

Consideration Heads: Every multi-head consideration layer consists of a number of “consideration heads,” every with its personal set of question, key, and worth projections. This enables the mannequin to take care of totally different facets of the enter concurrently, capturing various relationships and patterns.

Residual Connections and Layer Normalization: To facilitate the coaching of deep networks and mitigate the vanishing gradient downside, decoder-based LLMs make use of residual connections and layer normalization strategies. Residual connections add the enter of a layer to its output, permitting gradients to movement extra simply throughout backpropagation. Layer normalization helps to stabilize the activations and gradients, additional enhancing coaching stability and efficiency.

Feed-Ahead Layers

Along with multi-head consideration layers, decoder-based LLMs incorporate feed-forward layers, which apply a easy feed-forward neural community to every place within the sequence. These layers introduce non-linearities and allow the mannequin to study extra complicated representations.

Activation Capabilities: The selection of activation perform within the feed-forward layers can considerably influence the mannequin’s efficiency. Whereas earlier LLMs relied on the widely-used ReLU activation, more moderen fashions have adopted extra subtle activation capabilities just like the Gaussian Error Linear Unit (GELU) or the SwiGLU activation, which have proven improved efficiency.

Sparse Consideration and Environment friendly Transformers

Whereas the self-attention mechanism is highly effective, it comes with a quadratic computational complexity with respect to the sequence size, making it computationally costly for lengthy sequences. To handle this problem, a number of strategies have been proposed to scale back the computational and reminiscence necessities of self-attention, enabling environment friendly processing of longer sequences.

Sparse Consideration: Sparse consideration strategies, such because the one employed within the GPT-3 mannequin, selectively attend to a subset of positions within the enter sequence, moderately than computing consideration scores for all positions. This could considerably scale back the computational complexity whereas sustaining affordable efficiency.

Sliding Window Consideration: Launched within the Mistral 7B mannequin , sliding window consideration (SWA) is an easy but efficient method that restricts the eye span of every token to a hard and fast window dimension. This strategy leverages the flexibility of transformer layers to transmit data throughout a number of layers, successfully growing the eye span with out the quadratic complexity of full self-attention.

Rolling Buffer Cache: To additional scale back reminiscence necessities, particularly for lengthy sequences, the Mistral 7B mannequin employs a rolling buffer cache. This method shops and reuses the computed key and worth vectors for a hard and fast window dimension, avoiding redundant computations and minimizing reminiscence utilization.

Grouped Question Consideration: Launched within the LLaMA 2 mannequin, grouped question consideration (GQA) is a variant of the multi-query consideration mechanism that divides consideration heads into teams, every group sharing a standard key and worth matrix. This strategy strikes a stability between the effectivity of multi-query consideration and the efficiency of ordinary self-attention, offering improved inference instances whereas sustaining high-quality outcomes.