bridge.models.minicpmo.minicpmo_config#

Builder configuration for MiniCPM-o’s understanding stage.

Module Contents#

Classes#

MiniCPMOModelConfig

Qwen3 language config plus the original MiniCPM-o perception config.

API#

class bridge.models.minicpmo.minicpmo_config.MiniCPMOModelConfig#

Bases: megatron.bridge.models.gpt.model_config.BridgeGPTModelConfig

Qwen3 language config plus the original MiniCPM-o perception config.

Remote checkpoint code must be explicitly trusted before model construction. Encoder and projector freezing are independent of language-model training.

builder: ClassVar[str]#

‘megatron.bridge.models.minicpmo.minicpmo_builder.MiniCPMOModelBuilder’

minicpmo_config: dict[str, Any]#

‘field(…)’

trust_remote_code: bool#

False

freeze_language_model: bool#

False

freeze_vision_model: bool#

True

freeze_audio_model: bool#

True

freeze_vision_projection: bool#

True

freeze_audio_projection: bool#

True