Modular PEFT Ref. Architecture

• Modular PEFT Reference Architecture is a design blueprint that incorporates lightweight, parameter-efficient modules into pre-trained models for specialized task adaptation.
• It specifies precise insertion points, such as embedding and attention slots, ensuring systematic integration and interoperability of diverse adaptation modules.
• The architecture enables effective trade-off analysis between adaptation power, resource efficiency, and training speed, supporting multi-domain and multimodal applications.

A Modular Parameter-Efficient Fine-Tuning (PEFT) Reference Architecture is a general structural blueprint for integrating compact, specialized adaptation modules into large pre-trained models (PLMs) to enable efficient adaptation across tasks, domains, and modalities. This reference design specifies plug-and-play insertion of well-scoped PEFT modules at defined locations in the model, supporting compositionality, extensibility, and organized trade-off analysis between adaptation power and resource efficiency. The architecture underpins most state-of-the-art PEFT systems, including those tailored for multi-domain, multi-modal, federated, and mixture-of-experts (MoE) transformer frameworks (Seo et al., 9 Mar 2025, Sabry et al., 2023, Prottasha et al., 19 Apr 2025, Liu et al., 4 Aug 2025, Patel et al., 24 Jan 2025).

1. Core Principles and Architectural Scope

The modular PEFT reference architecture is defined by the following principles:

• Separation of Backbone and Adaptation Modules: The pretrained backbone (PLM) remains frozen or partially frozen, providing the foundation for reasoning, while task- or domain-specific adaptation modules (“adapters”/“PEMs”) supply lightweight, trainable correction or control pathways (Sabry et al., 2023, Patel et al., 24 Jan 2025).
• Named Insertion Points and Interfaces: Adaptation modules are inserted at precisely specified slots such as input embedding, attention projections (Q/K/V/O), feedforward sublayers (FFN), or post-layer normalization junctions. Each module exhibits a documented interface specifying inputs, outputs, and parameter convention (Sabry et al., 2023, Seo et al., 9 Mar 2025, Belanec et al., 2 Dec 2025).
• Composition and Extensibility: Multiple PEFT modules may coexist in parallel or sequentially, supporting additive, multiplicative, or even router/composed interaction. Reuse and recombination across tasks or domains is a first-class guarantee (Patel et al., 24 Jan 2025, Sabry et al., 2023).
• Parameter and Efficiency Accounting: The architecture enables a priori analysis of parameter count, memory, throughput, and efficiency trade-offs for any compatible PEFT method (Prottasha et al., 19 Apr 2025, Belanec et al., 2 Dec 2025).

A schematic of the modular architecture is given in Table 1.

Block Type	PEFT Module Examples	Typical Slot
Prompt/Prefix	Soft Prompt, Prefix Tuning	Embedding, Attention
Adapter	Houlsby, Compacter	MLP after attention/FFN
Reparameterize	LoRA, IA³	Linear projections
MoE	MoFE, PERFT-Adapters	FFN/MoE block

2. Modular Components and Embedding Strategies

A modular PEFT system includes the following types of components:

• Base Model: Frozen or partially trainable stack of layers (transformer encoder/decoder), responsible for all “standard” model computations (token embeddings, position encoding, MHSA, FFN blocks, residual connections, layer normalization). e.g., TinyLlama, BERT, ViT (Seo et al., 9 Mar 2025, Prottasha et al., 19 Apr 2025).
• PEFT Modules: Small residual or multiplicative circuits (adapters, LoRA/IA³ projections, prompt vectors) inserted at predefined locations, uniquely identified by “module type” and potentially by domain/context label. Each PEFT module m comprises parameters φ of dimensionality orders of magnitude smaller than the backbone’s θ (Sabry et al., 2023, Hadji-Kyriacou et al., 2023).
• Gating/Router Systems (optional): For MoE-style or multi-module setups, a lightweight gating network or router computes dynamic mixture weights for each module or expert at inference time (Seo et al., 9 Mar 2025, Liu et al., 4 Aug 2025).

Integration points include parallel (additive to hidden state or linear map), sequential (input/output chaining), or contextual (per-token context-dependent adapters (Hadji-Kyriacou et al., 2023)). Modules may be loaded, replaced, or composed at runtime, and their states versioned and indexed for task/domain management (Patel et al., 24 Jan 2025).

3. Formal Parameterization and Forward Pass Semantics

Let x denote the input, θ parameters of the (frozen) base model, and {φ_i} adaptation modules plugged at slots S_i:

• Standard Forward Layer:

For PLM layer ℓ: $$h_\ell = \mathrm{LayerNorm}(h_{\ell-1} + a_\ell + f_\ell)$$ with $a_\ell$ (attention), $f_\ell$ (FFN).

• PEFT-Enhanced Layer (residual additive):

$h_\ell = G_\ell(h_{\ell-1};\theta_\ell, \phi_\ell)$, with $$h_\ell = \mathrm{LayerNorm}(h_{\ell-1} + a_\ell + f_\ell + \Delta h_\ell)$$, $\Delta h_\ell = m_\ell(\cdot;\phi_\ell)$, and $\phi_\ell=\varnothing$ if no module inserted (Sabry et al., 2023, Seo et al., 9 Mar 2025).

• MoE and Router Example (MoFE):

Mixture over K frozen experts $E_i$, router computes gate weights $g_i$: $y(h) = \sum_{i=1}^K g_i(h) \cdot E_i(h)$, with $g = \operatorname{SoftmaxTop}_m(Vh)$, $V\in\mathbb{R}^{K\times d}$ (Seo et al., 9 Mar 2025).

For composition, multiple modules may be summed at each slot: $$\Delta h_\ell = m^{(1)}_\ell(\cdot; \phi^{(1)}_\ell) + m^{(2)}_\ell(\cdot; \phi^{(2)}_\ell)$$ or chained: $$h_\ell = h_\ell + m^{(1)}_\ell(h_\ell) \to h_\ell = h_\ell + m^{(2)}_\ell(h_\ell)$$ (Sabry et al., 2023, Patel et al., 24 Jan 2025).

Table 2 compares typical parameter costs.

| Method | Formula for Δ|θ| (Added Params) | Location | | -------------- | ------------------------------ | ------------------ | | Prompt Tuning | $n\cdot d$ | Input embedding | | Prefix Tuning | $L\cdot 2nd$ | Attention | | LoRA | $2rd$ | Linear projections | | Adapter | $L\cdot 2d d_h$ | FFN/Attention | | MoFE (K exp) | $K\cdot P_{ex}$ (frozen) | FFN/MoE block | | PERFT | $M\cdot 2Dr + D\cdot M$ | MoE parallel |

4. Module Composition, Reusability, and Domain Generalization

A key property is composability: modular PEFT architectures support merging, interpolation, and weighted gating of independently fine-tuned modules:

• Module Summation & Convex Combination:

For N adaptation modules (e.g., domains), $\theta_C = \sum_{i=1}^N \lambda_i\,\theta_i$, $\sum_i \lambda_i=1$ (Patel et al., 24 Jan 2025).

• Block-wise, Element-wise Gating:

$\theta_C[j] = \sum_{i=1}^N g_i[j] \theta_i[j]$, $g_i[j] \in [0,1]$

• Plug-and-Play Multi-domain Assembly:

Modular repositories track PEMs/versioned adapters by domain, base model checkpoint, and method. PEMs are dynamically composed at inference for composite tasks (Patel et al., 24 Jan 2025, Hadji-Kyriacou et al., 2023).

The compositional design allows the same backbone to power distinct tasks, multi-domain generalization, or federated updates via local adapters and central aggregation (Chua et al., 2023). The shared subspace structure enables summing without additional fine-tuning and preserves directional biases.

5. Efficiency Analysis and Trade-off Practices

The architecture robustly characterizes memory, parameter, and compute efficiency:

• Parameter Count:

$$P_{total} = P_{base} + \sum_{\ell \in S} |\phi_\ell|$$

• Memory Footprint:

$\Delta M \approx \Delta|θ| \times$ wordsize (bytes, e.g., 2 for FP16) (Sabry et al., 2023, Prottasha et al., 19 Apr 2025).

• Training Speed:

Time ∼ O(forward/backward FLOPs in PEFT modules) Many methods reduce training time by >50–70% compared to full fine-tuning with only ∼2–3 point accuracy drop; specific MoFE results capture this directly (Seo et al., 9 Mar 2025).

• Composite Efficiency Metric:

TPME = weighted norm of {train-time, parameter, GPU memory} (Fu et al., 2024).

Best-practice guidelines:

• Use LoRA or IA³ for lowest-overhead tasks; adapters for deeper influence; compacters to enable sharing; prompt/prefix for simple, memory-constrained scenarios (Sabry et al., 2023, Prottasha et al., 19 Apr 2025).
• Modularize via uniform PEFT-branch interface (Prottasha et al., 19 Apr 2025, Belanec et al., 2 Dec 2025).
• Employ load-balancing loss or router reset if MoE gating degenerates (Seo et al., 9 Mar 2025, Liu et al., 4 Aug 2025).

6. Mixture-of-Experts, Contextual, Multimodal, and Federated Extensions

Advanced modular PEFT architectures generalize to:

• MoE and Sparse Routing:

MoFE and PERFT instantiate mixtures of frozen (domain) experts, routed by parameter-efficient gates. Adapter mixtures (PERFT) increase efficiency in MoE LLMs over MoE-agnostic LoRA, especially with token-wise soft top-$K$ selection (Seo et al., 9 Mar 2025, Liu et al., 4 Aug 2025, Liu et al., 2024).

• Context-Aware and Multi-Modal PEFT:

Context-PEFT injects parallel context-specific adapters for each token-domain (modality, task, semantic role), replacing single-module updates with C-way selection (Hadji-Kyriacou et al., 2023).

• Multi-modal/Decoupled Frameworks:

IISAN decouples adaptation into separate intra-modal and inter-modal towers, drastically reducing GPU/memory cost compared to embedded fusion (Fu et al., 2024).

• Federated/Privacy-Preserving Patterns:

FedPEAT combines centralized backbone, distributed adapter fine-tuning, and optional emulation (distilled or compressed base models), orchestrated by RL-informed resource control (Chua et al., 2023).

These modular extensions share the backbone-PEFT interface, router/gating schema, and compositional layer, enabling seamless scaling across distributed, heterogeneous, or multi-functional environments.

7. Implementation and Benchmarking Frameworks

Modern modular reference architectures (e.g., PEFT-Factory) formalize interfaces and workflows:

• Core Modules: PEFT methods registry, dataset loaders, base model loader, metrics/evaluators (Belanec et al., 2 Dec 2025).
• Interface Standards:
  • PeftConfig (hyperparameter dataclass)
  • BaseTuner (module/adapter instantiation and forward logic)
  • Registry/Plugin architecture for custom method addition
  • Command-line/YAML configuration for instantiation and reproducibility
• Parameter/Memory Formulas:
  • Overhead: $N_{peft}/N_{total}\times 100\%$
  • Memory estimate: $M_{total}+M_{peft}\approx 4(N_{total}+N_{peft})$ bytes (FP32)
• Evaluation:
  • Metrics: BLEU, ROUGE, token acc., F1, PEFT-specific efficiency metrics (e.g., PSCP, TPME)
  • Results are logged and versioned for comparisons (Belanec et al., 2 Dec 2025, Fu et al., 2024).

The modular PEFT reference design thus guarantees, through clear definition of slots, module contracts, and efficiency metrics, a scalable and extensible substrate for future PEFT research and applications across evolving model and domain frontiers.