Implementation:Alibaba ROLL AutoConfig

Knowledge Sources	Alibaba_ROLL
Domains	Configuration, Model_Architecture
Last Updated	2026-02-07 20:00 GMT

Overview

Auto-configuration registry and factory for loading and instantiating model configurations by model type in the mcore_adapter framework.

Description

This module implements a registry pattern for model configurations, allowing dynamic registration and lookup of McaModelConfig subclasses by model type string. The global CONFIG_MAPPING (an OrderedDict) serves as the central registry. The register_config function acts as both a decorator and a direct registration call, associating a model type string (e.g., "qwen3_omni_moe") with its configuration class. If a model type is re-registered, a warning is logged but the new class overrides the previous one.

The AutoConfig class provides two factory class methods:

• for_model: Instantiates a config class directly from a model type string and optional constructor arguments
• from_pretrained: Loads a config from a checkpoint directory by first checking for an mca_config.json file (MCA format), then falling back to the standard HuggingFace config.json. It reads the hf_model_type field to determine which registered config class to use, then delegates to that class's from_pretrained method.

Usage

Use register_config as a decorator on new McaModelConfig subclasses to register them with the auto-config system. Use AutoConfig.from_pretrained to load model configurations from checkpoint directories in a model-type-agnostic manner. Use AutoConfig.for_model when you know the model type and want to create a fresh config instance.

Code Reference

Source Location

• Repository: Alibaba_ROLL
• File: mcore_adapter/src/mcore_adapter/models/auto/config_auto.py
• Lines: 1-63

Signature

CONFIG_MAPPING: OrderedDict  # Global registry

def register_config(model_type: str, cls=None): ...

def get_config_cls(model_type: str) -> Type[McaModelConfig]: ...

class AutoConfig:
    @classmethod
    def for_model(cls, model_type: str, *args, **kwargs) -> McaModelConfig: ...

    @classmethod
    def from_pretrained(cls, model_name_or_path: str, *args, **kwargs) -> McaModelConfig: ...

Import

from mcore_adapter.models.auto.config_auto import AutoConfig, register_config, get_config_cls, CONFIG_MAPPING

I/O Contract

Inputs

Name	Type	Required	Description
model_type	str	Yes	String identifier for the model type (e.g., "qwen3_omni_moe", "qwen2")
cls	type or None	No	Config class to register; if None, returns a decorator
model_name_or_path	str	Yes (for from_pretrained)	Path to a checkpoint directory containing mca_config.json or config.json

Outputs

Name	Type	Description
(register_config)	type or Callable	The registered class, or a decorator if cls is None
(get_config_cls)	Type[McaModelConfig]	The registered config class for the given model type, or McaModelConfig as fallback
(AutoConfig.for_model)	McaModelConfig	A fresh config instance for the specified model type
(AutoConfig.from_pretrained)	McaModelConfig	A config instance loaded from a checkpoint directory

Usage Examples

from dataclasses import dataclass
from mcore_adapter.models.auto.config_auto import AutoConfig, register_config
from mcore_adapter.models.model_config import McaModelConfig

# Register a new model config class
@register_config("my_model")
@dataclass
class MyModelConfig(McaModelConfig):
    custom_param: int = 42

# Create config by model type
config = AutoConfig.for_model("my_model", custom_param=100)

# Load config from a pretrained checkpoint directory
config = AutoConfig.from_pretrained("/path/to/checkpoint")
print(config.hidden_size)

Related Pages

• Environment:Alibaba_ROLL_Python_Runtime_Environment