bridge.models.muse_spark_open.muse_spark_open_bridge

HF config and parameter mappings for the Muse Spark Open Hybrid integration.

Module Contents

Classes

MuseSparkBridge	Convert the supplied multimodal HF checkpoint to a native Hybrid stack.
MuseSparkTextBridge	Use the same Hybrid decoder and mappings for the standalone language model.

API

class bridge.models.muse_spark_open.muse_spark_open_bridge.MuseSparkBridge

Bases: megatron.bridge.models.conversion.model_bridge.MegatronModelBridge

Convert the supplied multimodal HF checkpoint to a native Hybrid stack.

MODEL_CONFIG_CLASS

None

USE_MODEL_CONFIG_FOR_CONVERSION

True

text_only_pretrained(

hf_pretrained: megatron.bridge.models.hf_pretrained.PreTrainedCausalLM,

) → megatron.bridge.models.hf_pretrained.PreTrainedCausalLM

Select language weights and a standalone HF text configuration.

postprocess_hf_export_artifacts(path: pathlib.Path) → None

Preserve the VL head computation in standalone HF text exports.

stream_weights_megatron_to_hf(

megatron_model: megatron.bridge.models.muse_spark_open.modeling_muse_spark_open.model.MuseSparkModel | list[megatron.bridge.models.muse_spark_open.modeling_muse_spark_open.model.MuseSparkModel],

hf_pretrained: Any,

cpu: bool = True,

show_progress: bool = True,

conversion_tasks: list[megatron.bridge.models.conversion.model_bridge.WeightConversionTask] | None = None,

merge_adapter_weights: bool = True,

weight_dtype: torch.dtype | None = None,

with_megatron_names: bool = False,

) → collections.abc.Iterable[megatron.bridge.models.conversion.model_bridge.HFWeightTuple | megatron.bridge.models.conversion.model_bridge.HFSourcedWeightTuple]

Emit small tensors first so interleaved HF shards can release expert weights.

hf_config_to_model_config(

hf_config: transformers.PretrainedConfig,

) → megatron.bridge.models.muse_spark_open.muse_spark_open_config.MuseSparkModelConfig

Read all architecture sizes from HF and expand schedules to Hybrid positions.

classmethod megatron_to_hf_config(

model_config: megatron.bridge.models.muse_spark_open.muse_spark_open_config.MuseSparkModelConfig,

) → dict[str, Any]

Export current language settings and retain the original HF perception config.

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

Map HF blocks to separate attention and MoE Hybrid positions.

class bridge.models.muse_spark_open.muse_spark_open_bridge.MuseSparkTextBridge

Bases: bridge.models.muse_spark_open.muse_spark_open_bridge.MuseSparkBridge

Use the same Hybrid decoder and mappings for the standalone language model.