bridge.models.minicpmo.minicpmo_bridge#

MiniCPM-o 4.5 registration using Qwen3’s existing conversion rules.

Module Contents#

Classes#

MiniCPMOBridge

Convert the understanding stage while preserving MiniCPM-o identity.

API#

class bridge.models.minicpmo.minicpmo_bridge.MiniCPMOBridge#

Bases: megatron.bridge.models.qwen.qwen3_bridge.Qwen3Bridge

Convert the understanding stage while preserving MiniCPM-o identity.

The language weights are Qwen3 weights under llm. Perception weights are replicated and retain their Hugging Face names under perception. TTS is deliberately excluded from this understanding-stage model.

MODEL_CONFIG_CLASS#

None

USE_MODEL_CONFIG_FOR_CONVERSION#

True

HF_EXPORT_IGNORED_SOURCE_PREFIXES#

(‘tts.’,)

ADDITIONAL_FILE_PATTERNS#

[‘tokenizer*.json’, ‘special_tokens_map.json’, ‘chat_template.jinja’, ‘preprocessor_config.json’, ‘p…

postprocess_hf_export_artifacts(path: pathlib.Path) → None#

Keep component metadata compatible with the native processor loader.

The native processor receives its components as positional arguments. Newer Transformers also serializes them inside processor_config.json, which supplies the same arguments twice when loaded by native HF 4.51. Component settings remain in the preserved source artifact files.

hf_config_to_model_config_kwargs(
hf_config: transformers.PretrainedConfig,
) → dict[str, Any]#

Use Qwen3 language conversion and retain the perception configuration.

classmethod megatron_to_hf_config(
provider: megatron.bridge.models.minicpmo.minicpmo_config.MiniCPMOModelConfig,
) → dict[str, Any]#

Preserve perception configuration when exporting understanding weights.

mapping_registry() → megatron.bridge.models.conversion.mapping_registry.MegatronMappingRegistry#

Namespace Qwen3 mappings and add replicated perception parameters.