Implementation:Unslothai Unsloth Model Registry

Knowledge Sources	Unsloth
Domains	Model_Management, Registry
Last Updated	2026-02-07 08:40 GMT

Overview

Concrete tool for managing quantized model metadata registration and lookup via a global model registry.

Description

The registry module defines the core data structures and registration logic for Unsloth's model registry. QuantType enumerates supported quantization types (BNB, UNSLOTH, GGUF, NONE, BF16). ModelInfo stores metadata for individual model variants including organization, name, version, size, and quantization type, auto-constructing HuggingFace-compatible model paths. ModelMeta templates bulk registration of model families. The global MODEL_REGISTRY dictionary and register_model / _register_models functions manage variant registration with duplicate detection.

Usage

Import ModelInfo and register_model when adding new model families to the Unsloth registry, or access MODEL_REGISTRY to look up supported model variants.

Code Reference

Source Location

• Repository: Unslothai_Unsloth
• File: unsloth/registry/registry.py
• Lines: 1-191

Signature

class QuantType(Enum):
    BNB = "bnb"
    UNSLOTH = "unsloth"
    GGUF = "GGUF"
    NONE = "none"
    BF16 = "bf16"

@dataclass
class ModelInfo:
    org: str
    base_name: str
    version: str
    size: int
    name: str = None
    is_multimodal: bool = False
    instruct_tag: str = None
    quant_type: QuantType = None
    description: str = None

@dataclass
class ModelMeta:
    org: str
    base_name: str
    model_version: str
    model_info_cls: type[ModelInfo]
    model_sizes: list[str] = field(default_factory=list)
    instruct_tags: list[str] = field(default_factory=list)
    quant_types: list[QuantType] = field(default_factory=list)
    is_multimodal: bool = False

MODEL_REGISTRY: dict[str, ModelInfo] = {}

def register_model(
    model_info_cls: ModelInfo, org: str, base_name: str,
    version: str, size: int, instruct_tag: str = None,
    quant_type: QuantType = None, is_multimodal: bool = False,
    name: str = None,
) -> None:
    """Register a single model variant in the global registry."""

def _register_models(
    model_meta: ModelMeta, include_original_model: bool = False,
) -> None:
    """Bulk register model family variants from a ModelMeta template."""

Import

from unsloth.registry.registry import (
    ModelInfo, ModelMeta, QuantType,
    MODEL_REGISTRY, register_model,
)

I/O Contract

Inputs (register_model)

Name	Type	Required	Description
model_info_cls	type	Yes	ModelInfo class or subclass
org	str	Yes	Organization name (e.g., "unsloth")
base_name	str	Yes	Base model name (e.g., "Llama-3.1")
version	str	Yes	Model version (e.g., "3.1")
size	int	Yes	Model size parameter count
instruct_tag	str	No	Instruction-tuned variant tag
quant_type	QuantType	No	Quantization type
is_multimodal	bool	No	Whether model is multimodal (default: False)

Outputs

Name	Type	Description
MODEL_REGISTRY	dict[str, ModelInfo]	Global dict mapping "org/name" to ModelInfo
ModelInfo.model_path	str	HuggingFace model path "org/name"

Usage Examples

Register a Model Family

from unsloth.registry.registry import (
    ModelMeta, ModelInfo, QuantType, _register_models,
)

meta = ModelMeta(
    org="meta-llama",
    base_name="Llama-3.1",
    model_version="3.1",
    model_info_cls=ModelInfo,
    model_sizes=["1B", "3B", "8B", "70B"],
    instruct_tags=["Instruct"],
    quant_types=[QuantType.BNB, QuantType.UNSLOTH],
)

_register_models(meta, include_original_model=True)

Look Up a Model

from unsloth.registry.registry import MODEL_REGISTRY

info = MODEL_REGISTRY.get("unsloth/Llama-3.1-8B-Instruct-bnb-4bit")
if info:
    print(f"Model: {info.model_path}")
    print(f"Quant: {info.quant_type}")

Related Pages

• Principle:Unslothai_Unsloth_Model_Registry