Principle:Unslothai Unsloth Model Registry
| Knowledge Sources | |
|---|---|
| Domains | Model_Management, Registry |
| Last Updated | 2026-02-07 08:40 GMT |
Overview
Pattern for cataloging and managing quantized model variants through a structured registration system with HuggingFace Hub-compatible naming conventions.
Description
The Model Registry principle defines how ML frameworks maintain a catalog of supported model variants across different quantization types, sizes, and instruction-tuning configurations. Each model family is described by a metadata template that specifies the combinatorial space of variants (sizes x instruct tags x quantization types). Registration produces HuggingFace-compatible model identifiers that can be directly used for model loading and Hub operations.
Usage
Apply this principle when building an ML framework that needs to support multiple model families across different quantization and variant configurations, with programmatic discovery and validation of supported models.
Theoretical Basis
The registry follows a hierarchical naming convention:
Model identity: Failed to parse (syntax error): {\displaystyle \text{path} = \text{org}/\text{base\_name}\text{-}\text{version}\text{-}\text{size}[\text{-}\text{instruct}][\text{-}\text{quant}]}
Registration:
# Abstract registration algorithm
for size in model_sizes:
for tag in instruct_tags:
for quant in quant_types:
name = construct_name(base, version, size, tag, quant)
registry[f"{org}/{name}"] = ModelInfo(...)
Key properties:
- Uniqueness: Each org/name combination must be unique
- Combinatorial: Families define the Cartesian product of variants
- Hub-compatible: Names map directly to HuggingFace Hub repository paths