bridge.models.nemotron_vl.nemotron_vl_provider

Module Contents

Classes

NemotronNano12Bv2VLModelProvider

Configuration provider for Nemotron-VL models.

API

class bridge.models.nemotron_vl.nemotron_vl_provider.NemotronNano12Bv2VLModelProvider

Bases: megatron.bridge.models.nemotronh.nemotron_h_provider.NemotronNano12Bv2Provider

Configuration provider for Nemotron-VL models.

scatter_embedding_sequence_parallel: bool

False

attention_softmax_in_fp32: bool

True

vision_model_type: str

‘radio’

language_model_type: str

‘nemotron5-hybrid-12b’

generation_config: Optional[Any]

None

freeze_language_model: bool

False

freeze_vision_model: bool

False

freeze_vision_projection: bool

False

provide(pre_process=None, post_process=None, vp_stage=None)

Assemble a full :class:~megatron.core.models.multimodal.llava_model.LLaVAModel and wrap it.

This is a very trimmed-down version of the assembly code used in pretrain_vlm.py – it relies only on parameters already stored in the provider so that it works in any script (no Megatron-training CLI required).

provide_language_model(

pre_process=None,

post_process=None,

vp_stage=None,

)