Adapter

A core production use case is multi-adapter serving, where a single frozen base model (e.g., Llama 3, GPT) hosts multiple task-specific adapters. An inference server dynamically performs adapter switching based on request metadata (e.g., task_id). This allows one GPU-resident model to handle classification, summarization, and translation for different tenants without redundant model copies.

• Key Benefit: Drastically reduces memory footprint and management overhead compared to serving separate fine-tuned models for each task.
• Example: A customer support platform uses one base model with adapters for sentiment analysis, intent classification, and response generation, routing queries appropriately.

Adapters facilitate quick adaptation of a general-purpose LLM to a specialized vertical (e.g., legal, medical, finance) with minimal labeled data. Instead of full fine-tuning, a small adapter is trained on domain-specific corpora and instructions.

• Process: The base model's world knowledge remains intact while the adapter learns domain-specific terminology, formatting, and reasoning patterns.
• Efficiency: This is far faster and cheaper than full fine-tuning, enabling rapid prototyping and A/B testing of specialized models.
• Example: A financial services firm adapts a base model with a legal adapter for contract review and a separate SEC-filing adapter for earnings call analysis.

Adapters provide a controlled mechanism for continuous model learning. When new data or user feedback indicates performance drift, a new adapter can be trained on the recent data while the old adapter is preserved. This enables safe deployment strategies:

• Canary Deployment: Route 5% of traffic to the model with the new adapter, monitoring for regressions.
• Shadow Mode: Run inferences with the new adapter in parallel, logging outputs without affecting users, to compare against the production adapter.
• Rollback Safety: If the new adapter fails, simply revert the routing logic; the base model and previous adapters are unaffected, avoiding a full model rollback.

A seminal application of adapters is in cross-lingual NLP. A massively multilingual pre-trained model (e.g., mBERT, XLM-R) is frozen, and language-specific adapters are trained for each target language. This approach is highly parameter-efficient compared to fine-tuning the entire model per language.

• Mechanism: The adapter learns to project language-specific features into the shared multilingual semantic space of the base model.
• Advantage: Enables support for dozens of languages by simply adding small adapter files (~10MB each) rather than multi-gigabyte model copies.
• Example: A content moderation system uses a base model with separate adapters for Spanish, Japanese, and Arabic to classify toxic content, all sharing the core moderation knowledge.

Adapters can be composed or stacked to combine capabilities. This modularity allows for building complex behaviors from simpler, pre-trained components.

• AdapterFusion: A technique where a second, fusion adapter is trained to dynamically combine the outputs of multiple pre-trained task adapters (e.g., sentiment + fact-checking) for a new, composite task.
• Sequential Stacking: Adapters can be inserted sequentially within layers to learn hierarchical representations (e.g., a syntax adapter followed by a semantics adapter).
• Use Case: An agentic system might use a base model with a tool-use adapter and a safety adapter stacked together to ensure capable yet constrained behavior.

The small size of adapters (often <1% of base model parameters) makes them ideal for edge AI and on-device learning. A large model can be quantized and compiled for a device, while a lightweight adapter is trained locally on private user data.

• Privacy: User data never leaves the device; only the small adapter update could be federated.
• Efficiency: Avoids the infeasible compute of full on-device fine-tuning. Only the adapter's gradients are calculated and updated.
• Example: A smartphone keyboard's language model personalizes to a user's writing style by training a local adapter, improving predictions without compromising the core model or user privacy.

A collection of techniques for adapting large pre-trained models to new tasks by updating only a small, targeted subset of the model's parameters. This drastically reduces computational and memory costs compared to full fine-tuning.

• Core principle: Freeze the vast majority of the base model's weights.
• Methods include: Adapters, Low-Rank Adaptation (LoRA), and prompt tuning.
• Primary benefit: Enables task-specific adaptation with a fraction of the storage and GPU memory, making it feasible to fine-tune and serve very large models.

A dominant PEFT method that freezes pre-trained model weights and injects trainable rank decomposition matrices into transformer layers.

• Mechanism: Represents weight updates (ΔW) as the product of two low-rank matrices: ΔW = BA, where B and A are the trainable adapters.
• Efficiency: The rank r of these matrices is typically very small (e.g., 8, 16), leading to a minimal number of trainable parameters.
• Inference: Trained LoRA weights can be merged with the base model for a standalone checkpoint or served separately for dynamic composition.

High-throughput, open-source inference servers optimized for large language models, critical for serving adapted models efficiently.

• vLLM: Notable for its PagedAttention algorithm, which optimizes memory management for the Key-Value (KV) Cache, drastically improving throughput.
• Text Generation Inference (TGI): Hugging Face's toolkit featuring optimized transformers code, token streaming, and continuous batching.
• PEFT Support: Both systems support serving models with LoRA adapters, enabling efficient multi-adapter serving scenarios.