nemo_automodel.components.models.llava_onevision.state_dict_adapter

State dict adapter for LLaVA-OneVision-1.5.

HF on-disk safetensors layout (from lmms-lab/LLaVA-OneVision-1.5-): visual.{patch_embed,class_embedding,class_pos_emb,pre_layernorm,blocks.,merger.} model.{embed_tokens,layers.,norm} lm_head.weight

NeMo in-memory module layout: model.visual.* model.language_model.* (transformers.Qwen3Model; keys match the Qwen3 prefix) lm_head.*

Applies the same regex rename HF does via _checkpoint_conversion_mapping: ^visual -> model.visual ^model(?!.(language_model|visual)) -> model.language_model

Module Contents

Classes

LlavaOneVisionStateDictAdapter

Functions

_rename

Data

_HF_TO_NEMO_RULES
_NEMO_TO_HF_RULES

API

nemo_automodel.components.models.llava_onevision.state_dict_adapter._HF_TO_NEMO_RULES

[(), ()]

nemo_automodel.components.models.llava_onevision.state_dict_adapter._NEMO_TO_HF_RULES

[(), ()]

class nemo_automodel.components.models.llava_onevision.state_dict_adapter.LlavaOneVisionStateDictAdapter(config: Any = None, **kwargs)

Bases: nemo_automodel.components.checkpoint.state_dict_adapter.StateDictAdapter

from_hf(

hf_state_dict: dict[str, Any],

**kwargs,

) → dict[str, Any]

to_hf(

state_dict: dict[str, Any],

exclude_key_regex: Optional[str] = None,

**kwargs,

) → dict[str, Any]

convert_single_tensor_to_hf(

fqn: str,

tensor: Any,

**kwargs,

) → list[tuple[str, Any]]

nemo_automodel.components.models.llava_onevision.state_dict_adapter._rename(key: str, rules) → str