bridge.models.kimi_vl.kimi_k25_vl_provider

Module Contents

Classes

KimiK25VLModelProvider

Model provider for Kimi K2.5 VL (Vision-Language) Models.

API

class bridge.models.kimi_vl.kimi_k25_vl_provider.KimiK25VLModelProvider

Bases: megatron.bridge.models.mla_provider.MLAModelProvider

Model provider for Kimi K2.5 VL (Vision-Language) Models.

Inherits language model configuration from MLAModelProvider. Core architecture parameters (num_layers, hidden_size, MoE config, MLA config, etc.) are populated by provider_bridge in the bridge class via hf_config_to_provider_kwargs.

Only VLM-specific fields (vision config, token IDs, freeze options, etc.) that are NOT part of the language model are defined here.

The vision component (MoonViT3d + PatchMergerMLP) is dynamically loaded from the HuggingFace model repository at runtime via trust_remote_code.

scatter_embedding_sequence_parallel: bool

False

vision_config: Any

None

hf_model_path: Optional[str]

None

bos_token_id: int

163584

eos_token_id: int

163585

image_token_id: int

163605

media_placeholder_token_id: int

163605

pad_token_id: int

163839

ignore_index: int

None

freeze_language_model: bool

False

freeze_vision_model: bool

False

freeze_vision_projection: bool

False

generation_config: Any | None

None

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

Provide a KimiK25VL model instance with vision and language components.

provide_language_model(

pre_process=None,

post_process=None,

vp_stage=None,

) → megatron.core.models.gpt.GPTModel

Provide just the language model component (MoE with MLA) without vision.