nemo_automodel.components.models.mistral3.model

Module Contents

Classes

Name	Description
`GradientCheckpointingLayer`	-
`Ministral3Attention`	-
`Ministral3CausalLMOutputWithPast`	-
`Ministral3Config`	Configuration for Ministral3 text decoder.
`Ministral3DecoderLayer`	-
`Ministral3ForCausalLM`	-
`Ministral3MLP`	-
`Ministral3Model`	-
`Ministral3ModelOutputWithPast`	-
`Ministral3PreTrainedModel`	-
`Ministral3RMSNorm`	-
`Ministral3RotaryEmbedding`	-

Functions

Name	Description
`_force_register`	Map a native transformers config to our model class in `auto_cls`.
`_get_llama_4_attn_scale`	-
`_register_ministral3_with_transformers`	Register Ministral3Config and models with transformers Auto classes.
`apply_rotary_pos_emb`	-
`eager_attention_forward`	-
`repeat_kv`	-
`rotate_half`	-

Data

ModelClass

logger

API

class nemo_automodel.components.models.mistral3.model.GradientCheckpointingLayer()

Bases: Module

nemo_automodel.components.models.mistral3.model.GradientCheckpointingLayer.forward(
    args = (),
    kwargs = {}
)

class nemo_automodel.components.models.mistral3.model.Ministral3Attention(
    config: nemo_automodel.components.models.mistral3.model.Ministral3Config,
    layer_idx: int
)

Bases: Module

attention_dropout

= config.attention_dropout

head_dim

k_proj

num_key_value_groups

o_proj

q_proj

scaling

= self.head_dim ** -0.5

v_proj

nemo_automodel.components.models.mistral3.model.Ministral3Attention.forward(
    hidden_states: torch.Tensor,
    position_embeddings: tuple[torch.Tensor, torch.Tensor],
    attention_mask: torch.Tensor | None,
    past_key_values: transformers.cache_utils.Cache | None = None,
    cache_position: torch.LongTensor | None = None,
    kwargs: transformers.processing_utils.Unpack[transformers.modeling_flash_attention_utils.FlashAttentionKwargs] = {}
) -> tuple[torch.Tensor, torch.Tensor | None]

class nemo_automodel.components.models.mistral3.model.Ministral3CausalLMOutputWithPast()

Dataclass

Bases: CausalLMOutputWithPast

class nemo_automodel.components.models.mistral3.model.Ministral3Config(
    vocab_size: int | None = 131072,
    hidden_size: int | None = 4096,
    intermediate_size: int | None = 14336,
    num_hidden_layers: int | None = 34,
    num_attention_heads: int | None = 32,
    num_key_value_heads: int | None = 8,
    head_dim: int | None = 128,
    hidden_act: str | None = 'silu',
    max_position_embeddings: int | None = 262144,
    initializer_range: float | None = 0.02,
    rms_norm_eps: float | None = 1e-05,
    use_cache: bool | None = True,
    pad_token_id: int | None = 11,
    bos_token_id: int | None = 1,
    eos_token_id: int | None = 2,
    tie_word_embeddings: bool | None = False,
    rope_parameters: dict | None = None,
    sliding_window: int | None = None,
    attention_dropout: float | None = 0.0,
    kwargs = {}
)

Bases: PretrainedConfig

Configuration for Ministral3 text decoder.

base_model_pp_plan

base_model_tp_plan

head_dim

keys_to_ignore_at_inference

= ['past_key_values']

model_type

= 'ministral3'

rope_scaling

rope_theta

= self.rope_parameters.get('rope_theta', 1000000.0)

class nemo_automodel.components.models.mistral3.model.Ministral3DecoderLayer(
    config: nemo_automodel.components.models.mistral3.model.Ministral3Config,
    layer_idx: int
)

Bases: GradientCheckpointingLayer

hidden_size

= config.hidden_size

input_layernorm

mlp

= Ministral3MLP(config)

post_attention_layernorm

self_attn

nemo_automodel.components.models.mistral3.model.Ministral3DecoderLayer.forward(
    hidden_states: torch.Tensor,
    attention_mask: torch.Tensor | None = None,
    position_ids: torch.LongTensor | None = None,
    past_key_values: transformers.cache_utils.Cache | None = None,
    use_cache: bool | None = False,
    cache_position: torch.LongTensor | None = None,
    position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
    kwargs: transformers.processing_utils.Unpack[transformers.utils.TransformersKwargs] = {}
) -> torch.Tensor

class nemo_automodel.components.models.mistral3.model.Ministral3ForCausalLM(
    config: nemo_automodel.components.models.mistral3.model.Ministral3Config
)

Bases: HFCheckpointingMixin, Ministral3PreTrainedModel, GenerationMixin

_pp_plan

= {'lm_head': (['hidden_states'], ['logits'])}

_tied_weights_keys

= {'lm_head.weight': 'model.embed_tokens.weight'}

_tp_plan

= {'lm_head': 'colwise_rep'}

lm_head

model

= Ministral3Model(config)

state_dict_adapter

= Mistral3FP8StateDictAdapter.for_causal_lm()

tie_word_embeddings_support

TieSupport = TieSupport.UNTIED_ONLY

vocab_size

= config.vocab_size

nemo_automodel.components.models.mistral3.model.Ministral3ForCausalLM.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,
    labels: torch.LongTensor | None = None,
    use_cache: bool | None = None,
    cache_position: torch.LongTensor | None = None,
    logits_to_keep: typing.Union[int, torch.Tensor] = 0,
    output_hidden_states: bool | None = None,
    kwargs: transformers.processing_utils.Unpack[transformers.utils.TransformersKwargs] = {}
) -> transformers.modeling_outputs.CausalLMOutputWithPast

nemo_automodel.components.models.mistral3.model.Ministral3ForCausalLM.get_input_embeddings()

nemo_automodel.components.models.mistral3.model.Ministral3ForCausalLM.get_output_embeddings()

nemo_automodel.components.models.mistral3.model.Ministral3ForCausalLM.set_input_embeddings(
    value
)

nemo_automodel.components.models.mistral3.model.Ministral3ForCausalLM.set_output_embeddings(
    new_embeddings
)

class nemo_automodel.components.models.mistral3.model.Ministral3MLP(
    config: nemo_automodel.components.models.mistral3.model.Ministral3Config
)

Bases: Module

act_fn

= ACT2FN[config.hidden_act]

down_proj

gate_proj

hidden_size

= config.hidden_size

intermediate_size

= config.intermediate_size

up_proj

nemo_automodel.components.models.mistral3.model.Ministral3MLP.forward(
    x
)

class nemo_automodel.components.models.mistral3.model.Ministral3Model(
    config: nemo_automodel.components.models.mistral3.model.Ministral3Config
)

Bases: Ministral3PreTrainedModel

embed_tokens

layers

norm

padding_idx

= config.pad_token_id

rotary_emb

= Ministral3RotaryEmbedding(config=config)

vocab_size

= config.vocab_size

nemo_automodel.components.models.mistral3.model.Ministral3Model.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.processing_utils.Unpack[transformers.utils.TransformersKwargs] = {}
) -> transformers.modeling_outputs.BaseModelOutputWithPast

class nemo_automodel.components.models.mistral3.model.Ministral3ModelOutputWithPast(
    image_hidden_states: torch.FloatTensor | None = None
)

Dataclass

Bases: BaseModelOutputWithPast

image_hidden_states

FloatTensor | None = None

class nemo_automodel.components.models.mistral3.model.Ministral3PreTrainedModel()

Bases: PreTrainedModel

_can_record_outputs

= {}

_no_split_modules

= ['Ministral3DecoderLayer']

_skip_keys_device_placement

= ['past_key_values']

base_model_prefix

= 'model'

config

Ministral3Config

class nemo_automodel.components.models.mistral3.model.Ministral3RMSNorm(
    hidden_size,
    eps = 1e-06
)

Bases: Module

weight

= nn.Parameter(torch.ones(hidden_size))

nemo_automodel.components.models.mistral3.model.Ministral3RMSNorm.forward(
    hidden_states
)

class nemo_automodel.components.models.mistral3.model.Ministral3RotaryEmbedding(
    config: nemo_automodel.components.models.mistral3.model.Ministral3Config,
    device = None
)

Bases: Module

inv_freq

Tensor

max_seq_len_cached

= config.max_position_embeddings

original_max_seq_len

= config.max_position_embeddings

rope_type

nemo_automodel.components.models.mistral3.model.Ministral3RotaryEmbedding.compute_default_rope_parameters(
    config: nemo_automodel.components.models.mistral3.model.Ministral3Config | None = None,
    device: typing.Optional[torch.device] = None,
    seq_len: int | None = None
) -> tuple[torch.Tensor, float]

staticmethod

nemo_automodel.components.models.mistral3.model.Ministral3RotaryEmbedding.forward(
    x,
    position_ids
)

nemo_automodel.components.models.mistral3.model._force_register(
    auto_cls,
    config_cls,
    model_cls
) -> None

Map a native transformers config to our model class in auto_cls.

transformers >=5.15 makes _LazyAutoMapping.register return without doing anything when the config class lives under transformers.*, so that hub remote code cannot hijack a native config (models/auto/auto_factory.py: if getattr(key, "__module__", "").startswith("transformers."): return). Ministral-3 checkpoints declare model_type: mistral3, so AutoConfig hands back HF’s Mistral3Config and the public register silently no-ops — AutoModelForCausalLM then rejects the config outright. Writing the entry the way register did before that guard keeps Ministral-3 loadable as a causal LM; it is a no-op once the public call has already taken effect (transformers <=5.12).

nemo_automodel.components.models.mistral3.model._get_llama_4_attn_scale(
    positions_ids: torch.Tensor,
    beta: float,
    max_position_embeddings: int
) -> torch.Tensor

nemo_automodel.components.models.mistral3.model._register_ministral3_with_transformers()

This uses the official transformers registration API. Registration is idempotent (re-registering the same config/model is a no-op in recent transformers versions).

nemo_automodel.components.models.mistral3.model.apply_rotary_pos_emb(
    q,
    k,
    cos,
    sin,
    position_ids = None,
    unsqueeze_dim = 1
)

nemo_automodel.components.models.mistral3.model.eager_attention_forward(
    module: torch.nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: torch.Tensor | None,
    scaling: float,
    dropout: float = 0.0,
    kwargs: transformers.processing_utils.Unpack[transformers.utils.TransformersKwargs] = {}
)

nemo_automodel.components.models.mistral3.model.repeat_kv(
    hidden_states: torch.Tensor,
    n_rep: int
) -> torch.Tensor

nemo_automodel.components.models.mistral3.model.rotate_half(
    x
)

nemo_automodel.components.models.mistral3.model.ModelClass = Ministral3ForCausalLM

nemo_automodel.components.models.mistral3.model.logger = logging.get_logger(__name__)