Principle:Predibase Lorax Adapter Inventory Configuration
| Knowledge Sources | |
|---|---|
| Domains | Parameter_Efficient_Finetuning, Model_Serving |
| Last Updated | 2026-02-08 02:00 GMT |
Overview
A configuration pattern for specifying multiple LoRA adapters with per-adapter blending weights for runtime merging into a single composite adapter.
Description
Adapter Inventory Configuration solves the problem of combining capabilities from multiple fine-tuned adapters at inference time. Instead of deploying a separate model for each adapter, users specify a list of adapter IDs with corresponding blending weights. The system validates that:
- At least one adapter ID is provided
- Weights array matches the length of the IDs array
- Adapter source types are valid ("hub", "local", "s3", "pbase")
- The merge configuration is mutually exclusive with single adapter_id
This enables task-specific ensemble creation without retraining.
Usage
Use this principle when you want to combine the strengths of multiple fine-tuned adapters. For example, merging a domain-specific adapter with a style adapter to get both domain knowledge and a particular output style.
Theoretical Basis
The adapter inventory defines the inputs to the merge operation:
Pseudo-code:
# Adapter inventory specification
adapters = {
"ids": ["adapter-A", "adapter-B", "adapter-C"],
"weights": [0.5, 0.3, 0.2], # Must sum to desired scale
"merge_strategy": "linear", # How to combine
"density": 1.0, # Pruning density
}
# Validation: len(ids) == len(weights), ids not empty