Chapter 6: Parameter-Efficient Adaptation Techniques

Exam focus

Primary domain: Developing LLM-Based Applications (24%). Secondary: Productionizing LLM Solutions (22%).

• Full fine-tuning
• PEFT
• LoRA
• Adapter layers
• Prompt tuning
• Prefix tuning
• Low-rank adaptation
• Domain adaptation
• Model merging (conceptual)

Scope Bullet Explanations

• Full fine-tuning: Updating most or all model parameters for maximum task adaptation.
• PEFT: Parameter-efficient fine-tuning; updates only a small subset of weights.
• LoRA: Low-rank adapters inserted into target layers for efficient adaptation.
• Adapter layers: Small trainable modules added between frozen base-model layers.
• Prompt tuning: Learns virtual prompt embeddings instead of changing core model weights.
• Prefix tuning: Learns trainable prefix key/value states to steer attention behavior.
• Low-rank adaptation: General approach of approximating updates with low-rank matrices.
• Domain adaptation: Tailoring model behavior to a specific industry/task domain.
• Model merging (conceptual): Combining checkpoint deltas/models, typically requiring strong regression validation.

Chapter overview

Most teams cannot afford repeated full fine-tuning of large models. Parameter-efficient methods provide practical adaptation paths with lower cost and faster iteration. This chapter compares approaches and operational tradeoffs.

Learning objectives

• Compare full fine-tuning with PEFT methods under compute and quality constraints.
• Explain LoRA, adapters, prompt tuning, and prefix tuning mechanics.
• Select adaptation strategy based on latency, storage, and deployment complexity.
• Understand domain adaptation and model merging considerations.

6.1 Adaptation strategy landscape

Full fine-tuning

Updates most parameters and can yield strong domain fit, but it is expensive in GPU time, memory, and artifact management.

PEFT

Updates a small subset of parameters while freezing most base model weights. Benefits include:

• lower training cost,
• smaller adaptation artifacts,
• faster iteration cycles,
• easier multi-domain management.

6.2 LoRA and low-rank adaptation

LoRA inserts trainable low-rank matrices into target layers (often attention projections). Instead of modifying full weight matrices, it learns compact deltas.

Practical implications:

• reduced memory footprint,
• efficient fine-tuning on limited hardware,
• simple adapter swapping for domain variants.

6.3 Adapter layers

Adapters are small trainable modules inserted between frozen backbone layers. They add flexibility but can impact serving latency if many adapters are chained or hot-swapped frequently.

6.4 Prompt tuning and prefix tuning

Prompt tuning

Learns virtual prompt embeddings prepended to input tokens. Lightweight but may have limited adaptation capacity for deep domain shifts.

Prefix tuning

Learns trainable prefix states that influence attention behavior. Often stronger than pure prompt tuning while still lighter than full fine-tuning.

6.5 Domain adaptation decisions

Choose method by constraints:

• maximum quality needed,

• available compute budget,

• model count in production,

• deployment and governance complexity. Typical pattern:

• Start with prompt-only baseline.

• Move to LoRA/adapters when consistency is insufficient.

• Use full fine-tuning only when adaptation gap justifies cost.

6.6 Model merging (conceptual)

Model merging combines weight deltas or checkpoints. Risks include incompatible updates, regression in safety behavior, and unpredictable interaction effects. Always require strong regression testing.

6.7 Operational workflow

1. Define target behavior and acceptance tests.
2. Train PEFT variant with fixed dataset version.
3. Compare against baseline prompt-only and full-FT reference.
4. Evaluate quality, latency, and memory.
5. Promote only variants meeting objective and safety gates.

6.8 Failure modes

• Overestimating PEFT gains on severe domain shift.
• Ignoring serving overhead for many adapter variants.
• Merging deltas without safety/evaluation regression suite.
• Lack of artifact versioning for adapter checkpoints.

Chapter summary

Parameter-efficient adaptation is the default production path for many organizations. The right technique depends on adaptation depth needed and operational simplicity required.

Mini-lab: adaptation strategy matrix

Goal: choose adaptation method for three business scenarios.

1. Define three scenarios: low-risk FAQ, regulated support, high-specialization technical assistant.
2. Score prompt-only, LoRA, adapters, and full-FT across quality, cost, latency, and maintenance.
3. Select recommended method per scenario.
4. Define required evaluation gates before deployment. Deliverable in Notion:

• Decision matrix with justification and rollout plan.

Review questions

1. Why is LoRA widely used in enterprise adaptation?
2. When should full fine-tuning still be considered?
3. How do prompt tuning and prefix tuning differ?
4. What deployment complexity comes with adapter-per-domain patterns?
5. Why must PEFT artifacts be versioned like models?
6. What risks are introduced by model merging?
7. Which method is best under strict GPU constraints?
8. How do you measure whether PEFT quality is sufficient?
9. Why is baseline prompt-only comparison still necessary?
10. What governance controls apply to adaptation artifacts?

Key terms

PEFT, LoRA, adapters, prompt tuning, prefix tuning, low-rank adaptation, domain adaptation, model merging.

Exam traps

• Assuming PEFT always matches full fine-tuning quality.
• Ignoring inference-time complexity in adaptation strategy choice.
• Treating merged models as production-ready without validation.