Transformer architecture

As explained in LLMs in Production

A transformer architecture is a neural-network architecture that models sequence relationships primarily through attention rather than recurrence or convolution.

For a task such as translation, a model must process the input before generating an output. A full transformer combines an encoder, which processes input, with a decoder, which generates output. The decoder receives the encoder’s representation through a multihead attention block. Attention replaces recurrence as the architecture’s central mechanism, allowing inputs to be processed in parallel rather than sequentially as in an LSTM. section 2.3.3 section 2.3

Choose the arrangement and training plan together:

• Use the full encoder–decoder arrangement when both input processing and generation matter, as in translation or summarization. T5 is an example. The full arrangement is bulkier and slower than either half alone. section 2.3.3
• Expect substantial training demands: transformers are memory-intensive and require large amounts of data and VRAM. Parallelism makes large-scale computation practical, but does not remove those costs. section 2.3.3 section 5.2
• When training from scratch, choose the layer arrangement for the task, dataset and available compute; then prepare and tokenize diverse text before optimizing the model. section 5.2.1

Figure 2.7 A full transformer visualized. A full transformer combines the encoder and the decoder and does well on all of the tasks of each, as well as conditional generation tasks such as summarization and translation. Because transformers are bulkier and slower than each of their halves, researchers and businesses have generally opted to use those halves over the whole transformer.

The diagram shows where encoder output enters the decoder and why the combined arrangement suits conditional generation. section 2.3.3

Strong language-model performance is not the same as complete linguistic understanding. Even state-of-the-art transformers are described here as making semantic approximations; typical uses do not model pragmatics or phonetics. section 2.3.3

How LLMs in Production covers it

How deep it goes

Explains encoder–decoder roles, parallelism and training trade-offs, but does not derive attention calculations in the supplied passages.

Examples

Uses translation, summarization and T5 to illustrate the full architecture, alongside a PyTorch training-from-scratch listing.

What it assumes

Introduces tokenization as conversion of text to numerical tokens while placing transformers after earlier language-model approaches.

What you can do after

Evaluate/understand architecture choices against task fit, parallelism, data needs and training memory.

How it connects

Connects attention and multihead attention to encoders, decoders, LSTMs, tokenization and language-model training.

liveBookmanning.com

Read LLMs in Production on liveBook Get the book at manning.com

Also explained in

• 

  liveBookmanning.com
  Learn Generative AI with PyTorchThe internal design and structural organization of a Transformer model built around decoder blocks or self-attention components.
• 

  liveBookmanning.com
  Natural Language Processing in Action, Second EditionA transformer is a neural network architecture that uses attention mechanisms to process sequences and model relationships among their elements.
• 

  liveBookmanning.com
  Sutskever's ListA Transformer is a neural network architecture based primarily on attention mechanisms for modeling relationships among sequence elements.
• 

  liveBookmanning.com
  Deep Learning with Python, Third EditionA neural-network architecture based on attention mechanisms, used here as the core architecture for large language models.
• 

  liveBookmanning.com
  Build a Text-to-Image Generator (from Scratch)A transformer architecture is a neural-network architecture that uses attention mechanisms to process relationships among elements in a sequence or other structured input.
• 

  liveBookmanning.com
  Deep Learning with R, Third EditionA Transformer is a neural-network architecture designed to process sets or sequences using attention-based contextual representations.
• 

  liveBookmanning.com
  How Large Language Models WorkA transformer is a neural-network architecture that uses attention mechanisms to process relationships among elements in a sequence.
• 

  liveBookmanning.com
  AI-Powered SearchA transformer is a neural-network architecture that models relationships among sequence elements using attention and can encode and decode representations.
• 

  liveBookmanning.com
  Transformers in ActionA transformer is a neural network architecture that processes sequential and other structured data using attention mechanisms.
• 

  liveBookmanning.com
  Deep Learning with PyTorch, Second EditionA neural network architecture for sequence modeling and content understanding based on attention rather than recurrence or convolution.
• 

  liveBookmanning.com
  Grokking AI ApplicationsA transformer is a neural-network architecture for sequential data that uses attention to relate elements of an input sequence.
• 

  liveBookmanning.com
  The AI Pocket BookA neural network architecture that contextualizes each token using surrounding tokens rather than compressing the whole prompt into one summary vector.
• 

  liveBookmanning.com
  Time Series Forecasting Using Foundation ModelsA neural-network architecture built from stacked attention and feed-forward processing components, typically organized into encoder and decoder parts.
• 

  liveBookmanning.com
  Machine Learning Algorithms in DepthA transformer model is a neural network architecture that uses attention-based mechanisms to model contextual relationships in structured sequences or other data.
• 

  liveBookmanning.com
  Inside Deep Learning: Math, Algorithms, ModelsA sequence model built from positional encoding and multiheaded attention, used here as the standard original design for NLP and other tasks.
• 

  liveBookmanning.com
  Grokking AI Algorithms, Second EditionThe Transformer is a neural-network architecture based primarily on attention mechanisms for processing sequences and modeling context.
• 

  liveBookmanning.com
  Hugging Face in ActionThe Transformer architecture is a neural-network architecture based on attention mechanisms for processing sequences and other structured inputs.
• 

  liveBookmanning.com
  Domain-Specific Small Language ModelsA neural network architecture for sequence modeling built around self-attention and parallel processing, underlying both large and small language models.
• 

  liveBookmanning.com
  Introduction to Generative AI, Second EditionA transformer is a neural-network architecture for sequence modeling that uses attention, especially self-attention, rather than recurrence.
• 

  liveBookmanning.com
  Build AI-Enhanced Web AppsA transformer model is a neural-network architecture that uses attention mechanisms to model contextual relationships in sequence data.
• 

  liveBookmanning.com
  Build a Large Language Model (From Scratch)a neural network architecture based on self-attention mechanisms that enables processing of sequential data and underpins many large language models.
• 

  liveBookmanning.com
  Build a Website with ChatGPTA transformer is a neural-network architecture that processes sequences using attention mechanisms to model relationships among their components.
• 

  liveBookmanning.com
  Data Analysis with LLMsThe Transformer is a neural-network architecture designed to model relationships in sequences, especially sequences with variable-length inputs or outputs.
• 

  liveBookmanning.com
  Generative AI in ActionThe Transformer architecture is a neural-network architecture based primarily on attention mechanisms for processing sequences.
• 

  liveBookmanning.com
  Coding with AIA Transformer is a neural-network architecture that processes sequence elements using attention and can process them in parallel.
• 

  liveBookmanning.com
  Optimization Algorithms: AI techniques for design, planning, and control problemsA transformer architecture is a neural-network architecture that processes sequences primarily through attention mechanisms rather than recurrence.
• 

  liveBookmanning.com
  Essential GraphRAGA transformer architecture is a neural-network architecture that processes sequences using attention-based representations and supports efficient language modeling.

Compare how 28 books explain transformer architecture →

Related concepts

• attentionuses · Attention is a neural-network mechanism that computes input-dependent weights to combine information according to its relevance to a computation.
• encoderhas part · An encoder is a component that transforms input data into an encoded representation for subsequent processing, reconstruction, comparison, or transmission.
• decodehas part · Decoding is the process of transforming an encoded or internal representation into an interpretable or target-form output.
• Vision Transformerincludes · A Vision Transformer is a Transformer-based neural-network architecture that processes images as sequences of tokenized regions using attention.
• model architecturekind of · The structural organization of a computational system’s components, relationships, and operations that determines how it processes information or behaves.
• GPT-2includes · GPT-2 is a family of autoregressive, decoder-only Transformer language models developed by OpenAI that generate text by predicting successive tokens from preceding context.

This explainer was generated by Manning's knowledge graph from the text of LLMs in Production. The section links point to the parts of the book it draws on.