book

Designing Large Language Model Applications

by Suhas Pai

March 2025

Intermediate to advanced

366 pages

9h 31m

English

O'Reilly Media, Inc.

Read now

Unlock full access

Includes

Includes Quizzes

Who This Book Is ForHow This Book Is StructuredWhat This Book Is Not AboutHow to Read the BookConventions Used in This BookUsing Code ExamplesO’Reilly Online LearningHow to Contact UsAcknowledgments
Defining LLMsA Brief History of LLMsEarly YearsThe Modern LLM EraThe Impact of LLMsLLM Usage in the EnterprisePromptingZero-Shot PromptingFew-Shot PromptingChain-of-Thought PromptingPrompt ChainingAdversarial PromptingAccessing LLMs Through an APIStrengths and Limitations of LLMsBuilding Your First Chatbot PrototypeFrom Prototype to ProductionSummary
Ingredients of an LLMPre-Training Data RequirementsPopular Pre-Training DatasetsSynthetic Pre-Training DataTraining Data PreprocessingData Filtering and CleaningSelecting Quality DocumentsDeduplicationRemoving Personally Identifiable InformationTraining Set DecontaminationData MixturesEffect of Pre-Training Data on Downstream TasksBias and Fairness Issues in Pre-Training DatasetsSummary
VocabularyTokenizersTokenization PipelineNormalizationPre-TokenizationTokenizationByte Pair EncodingWordPieceSpecial TokensSummary
PreliminariesRepresenting MeaningThe Transformer ArchitectureSelf-AttentionPositional EncodingFeedforward NetworksLayer NormalizationLoss FunctionsIntrinsic Model EvaluationTransformer BackbonesEncoder-Only ArchitecturesEncoder-Decoder ArchitecturesDecoder-Only ArchitecturesMixture of ExpertsLearning ObjectivesFull Language ModelingPrefix Language ModelingMasked Language ModelingWhich Learning Objectives Are Better?Pre-Training ModelsSummary
Navigating the LLM LandscapeWho Are the LLM providers?Model FlavorsOpen Source LLMsHow to Choose an LLM for Your TaskOpen Source Versus Proprietary LLMsLLM EvaluationLoading LLMsHugging Face AccelerateOllamaLLM Inference APIsDecoding StrategiesGreedy DecodingBeam SearchTop-k SamplingTop-p SamplingRunning Inference on LLMsStructured OutputsModel Debugging and InterpretabilitySummary
The Need for Fine-TuningFine-Tuning: A Full ExampleLearning Algorithms ParametersMemory Optimization ParametersRegularization ParametersBatch SizeParameter-Efficient Fine-TuningWorking with Reduced PrecisionPutting It All TogetherFine-Tuning DatasetsUtilizing Publicly Available Instruction-Tuning DatasetsLLM-Generated Instruction-Tuning DatasetsSummary
Continual Pre-TrainingReplay (Memory)Parameter ExpansionParameter-Efficient Fine-TuningAdding New ParametersSubset MethodsCombining Multiple ModelsModel EnsemblingModel FusionAdapter MergingSummary

Defining Alignment TrainingReinforcement LearningTypes of Human FeedbackRLHF ExampleHallucinationsMitigating HallucinationsSelf-ConsistencyChain-of-ActionsRecitationSampling Methods for Addressing HallucinationDecoding by Contrasting LayersIn-Context HallucinationsHallucinations Due to Irrelevant InformationReasoningDeductive ReasoningInductive ReasoningAbductive ReasoningCommon Sense ReasoningInducing Reasoning in LLMsVerifiers for Improving ReasoningInference-Time ComputationFine-Tuning for ReasoningSummary
LLM Inference ChallengesInference Optimization TechniquesTechniques for Reducing ComputeK-V CachingEarly ExitKnowledge DistillationTechniques for Accelerating DecodingSpeculative DecodingParallel DecodingTechniques for Reducing Storage NeedsSymmetric QuantizationAsymmetric QuantizationSummary
LLM Interaction ParadigmsPassive ApproachThe Explicit ApproachThe Autonomous ApproachDefining AgentsAgentic WorkflowComponents of an Agentic SystemModelsToolsData StoresAgent Loop PromptGuardrails and VerifiersAgent Orchestration SoftwareSummary
Introduction to EmbeddingsSemantic SearchSimilarity MeasuresFine-Tuning Embedding ModelsBase ModelsTraining DatasetLoss FunctionsInstruction EmbeddingsOptimizing Embedding SizeMatryoshka EmbeddingsBinary and Integer EmbeddingsProduct QuantizationChunkingSliding Window ChunkingMetadata-Aware ChunkingLayout-Aware ChunkingSemantic ChunkingLate ChunkingVector DatabasesInterpreting EmbeddingsSummary
The Need for RAGTypical RAG ScenariosDeciding When to RetrieveThe RAG PipelineRewriteRetrieveRerankRefineInsertGenerateRAG for Memory ManagementRAG for Selecting In-Context Training ExamplesRAG for Model TrainingLimitations of RAGRAG Versus Long ContextRAG Versus Fine-TuningSummary
Multi-LLM ArchitecturesLLM CascadesRoutersTask-Specialized LLMsProgramming ParadigmsDSPyLMQLSummary

Content preview from Designing Large Language Model Applications

Chapter 9. Inference Optimization

In the past few chapters, we learned several techniques for adapting and utilizing LLMs to solve specific tasks. In this chapter, we will learn how to efficiently perform inference on them for real-world usage. LLMs’ large size make deployment and inference particularly challenging, as they exert significant pressure on compute, memory, and energy requirements. This proves to be especially challenging on edge devices like mobile phones.

For the rest of the chapter, we will focus on the field of inference optimization, discussing the factors influencing LLM inference time. We will then showcase a variety of optimization techniques including caching, knowledge distillation, early exiting, quantization, parallel and speculative decoding, and more.

LLM Inference Challenges

What are the bottlenecks affecting LLM inference? As we all know, their gargantuan sizes necessitate vast computing and memory resources. Apart from that, two additional factors exacerbate the situation:

As seen in Chapter 4, contemporary LLMs are based largely on decoder-only models that operate autoregressively. This means that each token is generated one after the other, thus imposing a sequential limitation. Later in this chapter, we will discuss techniques for parallel and speculative decoding that aim to speed up the decoding process.
As the input sequence length increases, the amount of compute needed increases quadratically. Later this chapter, we will discuss techniques ...