bridge.models.gemma_vl.gemma4_vl_provider

Gemma 4 Vision-Language model providers.

Gemma4VLModelProvider: MoE Vision-Language provider (extends Gemma4ModelProvider). Gemma4DenseVLProvider: Dense Vision-Language provider (extends Gemma4DenseProvider).

Text-only providers (Gemma4DenseProvider, Gemma4ModelProvider) live in: megatron.bridge.models.gemma.gemma4_provider

Module Contents

Classes

Gemma4VLModelProvider	Model provider for Gemma 4 MoE Vision-Language models.
Gemma4DenseVLProvider	Model provider for Dense Gemma 4 Vision-Language checkpoints.

API

class bridge.models.gemma_vl.gemma4_vl_provider.Gemma4VLModelProvider

Bases: megatron.bridge.models.gemma.gemma4_provider.Gemma4ModelProvider

Model provider for Gemma 4 MoE Vision-Language models.

scatter_embedding_sequence_parallel: bool

False

vision_config: Any

None

text_config: Any

None

audio_config: Any

None

vision_soft_tokens_per_image: int

280

bos_token_id: int

2

eos_token_id: int

1

image_token_id: int

258880

video_token_id: int

258884

audio_token_id: int

258881

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,

) → megatron.bridge.models.gemma_vl.modeling_gemma4_vl.Gemma4VLModel

provide_language_model(

pre_process=None,

post_process=None,

vp_stage=None,

) → megatron.core.models.gpt.GPTModel

class bridge.models.gemma_vl.gemma4_vl_provider.Gemma4DenseVLProvider

Bases: megatron.bridge.models.gemma.gemma4_provider.Gemma4DenseProvider

Model provider for Dense Gemma 4 Vision-Language checkpoints.

scatter_embedding_sequence_parallel: bool

False

vision_config: Any

None

text_config: Any

None

audio_config: Any

None

vision_soft_tokens_per_image: int

280

bos_token_id: int

2

eos_token_id: int

1

image_token_id: int

258880

video_token_id: int

258884

audio_token_id: int

258881

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,

) → megatron.bridge.models.gemma_vl.modeling_gemma4_vl.Gemma4VLModel

provide_language_model(

pre_process=None,

post_process=None,

vp_stage=None,

) → megatron.core.models.gpt.GPTModel