nemo_automodel.components.models.ministral_bidirectional.model

Custom bidirectional Ministral3 model for explicit and legacy use.

This module provides a modified Ministral3Model that uses bidirectional (non-causal) attention, suitable for generating embeddings where each token should attend to all other tokens in the sequence. Standard ministral3 embedding checkpoints use the stock HuggingFace model with is_causal=False; this custom architecture remains available for ministral3_bidirec checkpoints.

Module Contents

Classes

Name	Description
`Ministral3BidirectionalConfig`	Configuration for Ministral3BidirectionalModel with pooling and temperature settings.
`Ministral3BidirectionalModel`	Ministral3Model modified to use bidirectional (non-causal) attention.

Functions

Name	Description
`_register_with_hf_auto_classes`	Register bidirectional Ministral3 with HuggingFace Auto classes.

Data

ModelClass

__all__

logger

API

class nemo_automodel.components.models.ministral_bidirectional.model.Ministral3BidirectionalConfig(
    pooling: str = 'avg',
    temperature: float = 1.0,
    kwargs = {}
)

Bases: Ministral3Config

Configuration for Ministral3BidirectionalModel with pooling and temperature settings.

model_type

= 'ministral3_bidirec'

class nemo_automodel.components.models.ministral_bidirectional.model.Ministral3BidirectionalModel(
    config
)

Bases: Ministral3Model

Ministral3Model modified to use bidirectional (non-causal) attention.

This class is not selected automatically for standard ministral3 embedding checkpoints. It remains registered for explicit and legacy use.

In causal Ministral3, each token can only attend to previous tokens (causal attention). This model removes that restriction, allowing each token to attend to all tokens in the sequence, which is useful for embedding tasks.

Loading a Mistral3 VLM checkpoint (e.g. mistralai/Ministral-3-3B-Base-2512 or mistralai/Ministral-3-3B-Instruct-2512) requires extracting the language tower; this is driven by the recipe YAML via extract_submodel: language_model and handled by :func:nemo_automodel._transformers.retrieval.build_encoder_backbone.

Text-only checkpoints (e.g. mistralai/Ministral-3B-Instruct) load directly via the standard from_pretrained path with no extraction needed.

nemo_automodel.components.models.ministral_bidirectional.model.Ministral3BidirectionalModel.forward(
    input_ids: torch.LongTensor | None = None,
    attention_mask: torch.Tensor | None = None,
    position_ids: torch.LongTensor | None = None,
    past_key_values: transformers.cache_utils.Cache | None = None,
    inputs_embeds: torch.FloatTensor | None = None,
    use_cache: bool | None = None,
    cache_position: torch.LongTensor | None = None,
    kwargs = {}
) -> transformers.modeling_outputs.BaseModelOutputWithPast

Forward pass with bidirectional attention.

Identical to Ministral3Model.forward() except the causal mask is replaced with a bidirectional mask, allowing all tokens to attend to each other.

nemo_automodel.components.models.ministral_bidirectional.model._register_with_hf_auto_classes() -> None

Needed so AutoModel.from_config(Ministral3BidirectionalConfig) and checkpoint reload paths that use Auto resolution work consistently.

nemo_automodel.components.models.ministral_bidirectional.model.ModelClass = [Ministral3BidirectionalModel]

nemo_automodel.components.models.ministral_bidirectional.model.__all__ = ['Ministral3BidirectionalModel', 'Ministral3BidirectionalConfig', 'ModelClass']

nemo_automodel.components.models.ministral_bidirectional.model.logger = logging.get_logger(__name__)