ReferenceFull Library ReferenceNemo AutomodelNemo AutomodelComponentsModelsDeepseek V4nemo_automodel.components.models.deepseek_v4.parallelization

nemo_automodel.components.models.deepseek_v4.parallelization

View as Markdown

Model-owned distributed parallelization for DeepSeek-V4.

Module Contents

Classes

NameDescription
DeepseekV4ModelParallelizerKeep DeepSeek-V4 reference-sensitive parameters in fp32 FSDP units.

Data

PARALLELIZER

__all__

API

class nemo_automodel.components.models.deepseek_v4.parallelization.DeepseekV4ModelParallelizer()

Bases: ModelParallelizer

Keep DeepSeek-V4 reference-sensitive parameters in fp32 FSDP units.

nemo_automodel.components.models.deepseek_v4.parallelization.DeepseekV4ModelParallelizer._fully_shard_module(
module,
kwargs = {}
)
nemo_automodel.components.models.deepseek_v4.parallelization.PARALLELIZER = DeepseekV4ModelParallelizer()
nemo_automodel.components.models.deepseek_v4.parallelization.__all__ = ['PARALLELIZER']