nemo_automodel.components.models.minimax_m3_vl.config

Typed configuration classes for the MiniMax M3 VL family.

The released checkpoint ships configuration_minimax_m3_vl.py which coerces the vision_config/text_config sub-dicts into generic PretrainedConfig instances (the text backbone’s model_type="minimax_m2" is not in HF’s CONFIG_MAPPING). For the native AutoModel implementation we declare typed sub-configs so that fields such as sparse_attention_config, moe_layer_freq and the SwiGLU-OAI parameters are real, defaulted attributes.

Mirrors the canonical sglang reference sglang.srt.configs.minimax_vl and the field set in the checkpoint’s config.json; keep them in sync.

Module Contents

Classes

Name	Description
`MiniMaxM3DSparkConfig`	Validated settings of the MiniMax M3 DSpark draft.
`MiniMaxM3DSparkReleasedConfig`	Draft contract read from a target checkpoint’s `dspark/config.json`.
`MiniMaxM3QuantizationConfig`	`quantization_config` block of the released MiniMax M3 checkpoints.
`MiniMaxM3VLConfig`	Top-level configuration for MiniMax M3 vision-language checkpoints.
`MiniMaxM3VLTextConfig`	Configuration for the MiniMax M3 (mixed sparse/dense MoE) text backbone.
`MiniMaxM3VLVisionConfig`	Configuration for the MiniMax M3 VL CLIP-style vision tower (Conv3d + 3D RoPE).
`_DSparkDraftOptions`	Declarative DSpark settings supplied by the generic training recipe.

Functions

Name	Description
`_json_safe_value`	Convert config values that are valid in-memory but not JSON serializable.

Data

_DSPARK_DRAFT_ATTN_IMPLEMENTATION

_DSPARK_RELEASED_ARCHITECTURE

_DSPARK_RELEASED_BACKBONE_FIELDS

_DSPARK_RELEASED_REQUIRED_FIELDS

logger

API

class nemo_automodel.components.models.minimax_m3_vl.config.MiniMaxM3DSparkConfig(
    text_config: nemo_automodel.components.models.minimax_m3_vl.config.MiniMaxM3VLTextConfig,
    num_draft_layers: int,
    target_layer_ids: tuple[int, ...],
    block_size: int,
    num_anchors: int,
    mask_token_id: int,
    markov_rank: int,
    markov_head_type: str,
    confidence_head_alpha: float,
    confidence_head_with_markov: bool,
    released: nemo_automodel.components.models.minimax_m3_vl.config.MiniMaxM3DSparkReleasedConfig | None = None
)

Dataclass

Validated settings of the MiniMax M3 DSpark draft.

block_size

int

Number of tokens drafted per anchor.

confidence_head_alpha

float

Confidence loss weight, zero to disable the head.

confidence_head_with_markov

bool

Whether the confidence head reads Markov embeddings.

markov_head_type

str

Markov head variant.

markov_rank

int

Rank of the Markov head, zero to disable it.

mask_token_id

int

Vocabulary id of the noise token.

num_anchors

int

Number of anchors sampled per training row.

num_draft_layers

int

Number of draft decoder layers.

released

MiniMaxM3DSparkReleasedConfig | None = None

Contract shipped by the target checkpoint, when it has one.

target_layer_ids

tuple[int, ...]

Target layers whose features the draft consumes.

text_config

MiniMaxM3VLTextConfig

Text configuration of the target model.

nemo_automodel.components.models.minimax_m3_vl.config.MiniMaxM3DSparkConfig._validate_released(
    released: nemo_automodel.components.models.minimax_m3_vl.config.MiniMaxM3DSparkReleasedConfig
) -> None

Reject recipe settings that disagree with the target’s released contract.

nemo_automodel.components.models.minimax_m3_vl.config.MiniMaxM3DSparkConfig.build() -> nemo_automodel.components.models.minimax_m3_vl.dspark.MiniMaxM3DSparkModel

Validate the settings and build the draft.

Returns: MiniMaxM3DSparkModel

Newly initialized draft ready for device placement.

Raises:

ValueError: If a setting is out of range or disagrees with released.

class nemo_automodel.components.models.minimax_m3_vl.config.MiniMaxM3DSparkReleasedConfig(
    num_draft_layers: int,
    target_layer_ids: tuple[int, ...],
    block_size: int,
    mask_token_id: int,
    markov_rank: int,
    markov_head_type: str,
    ffn_hidden_size: int,
    sliding_window: int,
    exported_config: dict[str, typing.Any]
)

Dataclass

Draft contract read from a target checkpoint’s dspark/config.json.

block_size

int

Number of tokens drafted per anchor.

exported_config

dict[str, Any]

Released config without quantization metadata, written next to the draft weights because the export is not quantized.

ffn_hidden_size

int

Hidden width of the draft’s dense FFN.

markov_head_type

str

Markov head variant.

markov_rank

int

Rank of the Markov head.

mask_token_id

int

Vocabulary id of the noise token.

num_draft_layers

int

Number of draft decoder layers.

sliding_window

int

Context window of every draft layer, in tokens.

target_layer_ids

tuple[int, ...]

Target layers whose features the draft consumes.

nemo_automodel.components.models.minimax_m3_vl.config.MiniMaxM3DSparkReleasedConfig.from_target(
    target_name_or_path: str
) -> nemo_automodel.components.models.minimax_m3_vl.config.MiniMaxM3DSparkReleasedConfig | None

classmethod

Read the contract a target checkpoint ships for its DSpark draft.

Parameters:

target_name_or_path

str

Hub id or local directory of the target checkpoint.

Returns: MiniMaxM3DSparkReleasedConfig | None

The parsed contract, or None when the target has no dspark/config.json.

Raises:

ValueError: If the file does not describe a dense, sliding-window DSparkMiniMaxDraftModel.

class nemo_automodel.components.models.minimax_m3_vl.config.MiniMaxM3QuantizationConfig

Bases: typing.TypedDict

quantization_config block of the released MiniMax M3 checkpoints.

M3 stores MXFP8 projections described by quant_method, activation_scheme, weight_block_size and ignored_layers. M3.1 adds NVFP4 packed routed experts described by moe_quant_algo, moe_quant_format and group_size.

activation_scheme

str

group_size

int

ignored_layers

list[str]

moe_quant_algo

str

moe_quant_format

str

quant_method

Required[str]

weight_block_size

list[int]

class nemo_automodel.components.models.minimax_m3_vl.config.MiniMaxM3VLConfig(
    vision_config: typing.Union[dict, nemo_automodel.components.models.minimax_m3_vl.config.MiniMaxM3VLVisionConfig] | None = None,
    text_config: typing.Union[dict, nemo_automodel.components.models.minimax_m3_vl.config.MiniMaxM3VLTextConfig] | None = None,
    image_token_index: int = 200025,
    video_token_index: int = 200026,
    image_seq_length: int = 576,
    process_image_mode: str = 'dynamic_res',
    projector_hidden_act: str = 'gelu',
    projector_hidden_size: int = 6144,
    multimodal_projector_bias: bool = True,
    patch_merge_bias: bool = True,
    vision_feature_layer: int = -1,
    vision_feature_select_strategy: str = 'full',
    img_token_compression_config: dict | None = None,
    image_grid_pinpoints: str | None = None,
    quantization_config: nemo_automodel.components.models.minimax_m3_vl.config.MiniMaxM3QuantizationConfig | None = None,
    kwargs = {}
)

Bases: PretrainedConfig

Top-level configuration for MiniMax M3 vision-language checkpoints.

hidden_size

= text_config.hidden_size

img_token_compression_config

= img_token_compression_config or {}

max_position_embeddings

= text_config.max_position_embeddings

model_type

= 'minimax_m3_vl'

sub_configs

nemo_automodel.components.models.minimax_m3_vl.config.MiniMaxM3VLConfig.build_dspark_draft(
    options: nemo_automodel.components.models.minimax_m3_vl.config._DSparkDraftOptions
) -> nemo_automodel.components.models.minimax_m3_vl.dspark.MiniMaxM3DSparkModel

Build the model-owned DSpark draft.

A target that ships dspark/config.json fixes the draft contract, and options must agree with it.

Parameters:

options

_DSparkDraftOptions

Declarative draft settings supplied by the DSpark recipe.

Returns: MiniMaxM3DSparkModel

Newly initialized draft ready for device placement.

nemo_automodel.components.models.minimax_m3_vl.config.MiniMaxM3VLConfig.to_dict()

class nemo_automodel.components.models.minimax_m3_vl.config.MiniMaxM3VLTextConfig(
    hidden_size: int = 6144,
    intermediate_size: int = 3072,
    dense_intermediate_size: int = 12288,
    shared_intermediate_size: int = 3072,
    num_hidden_layers: int = 60,
    num_attention_heads: int = 64,
    num_key_value_heads: int = 4,
    head_dim: int = 128,
    vocab_size: int = 200064,
    max_position_embeddings: int = 524288,
    rms_norm_eps: float = 1e-06,
    use_gemma_norm: bool = True,
    attention_output_gate: bool = False,
    rope_theta: float = 5000000.0,
    rotary_dim: int = 64,
    partial_rotary_factor: float = 0.5,
    hidden_act: str = 'swigluoai',
    use_qk_norm: bool = True,
    qk_norm_type: str = 'per_head',
    tie_word_embeddings: bool = False,
    num_local_experts: int = 128,
    num_experts_per_tok: int = 4,
    n_shared_experts: int = 1,
    scoring_func: str = 'sigmoid',
    use_routing_bias: bool = True,
    routed_scaling_factor: float = 2.0,
    moe_layer_freq: list[int] | None = None,
    swiglu_alpha: float = 1.702,
    swiglu_limit: float = 7.0,
    sparse_attention_config: dict | None = None,
    num_mtp_modules: int = 1,
    quantization_config: nemo_automodel.components.models.minimax_m3_vl.config.MiniMaxM3QuantizationConfig | None = None,
    pad_token_id: int | None = None,
    kwargs = {}
)

Bases: PretrainedConfig

Configuration for the MiniMax M3 (mixed sparse/dense MoE) text backbone.

architectures

= ['MiniMaxM3SparseForCausalLM']

model_type

= 'minimax_m3'

nemo_automodel.components.models.minimax_m3_vl.config.MiniMaxM3VLTextConfig.to_dict()

class nemo_automodel.components.models.minimax_m3_vl.config.MiniMaxM3VLVisionConfig(
    hidden_size: int = 1280,
    num_attention_heads: int = 16,
    num_hidden_layers: int = 32,
    intermediate_size: int = 5120,
    patch_size: int = 14,
    image_size: int = 672,
    projection_dim: int = 6144,
    num_channels: int = 3,
    position_embedding_type: str = 'rope',
    rope_mode: str = '3d',
    rope_theta: float = 10000.0,
    attention_dropout: float = 0.0,
    hidden_act: str = 'gelu',
    layer_norm_eps: float = 1e-05,
    img_token_compression_config: dict | None = None,
    vision_segment_max_frames: int = 4,
    kwargs = {}
)

Bases: PretrainedConfig

Configuration for the MiniMax M3 VL CLIP-style vision tower (Conv3d + 3D RoPE).

img_token_compression_config

model_type

= 'minimax_m3_vision'

nemo_automodel.components.models.minimax_m3_vl.config.MiniMaxM3VLVisionConfig.to_dict()

class nemo_automodel.components.models.minimax_m3_vl.config._DSparkDraftOptions()

Protocol

Declarative DSpark settings supplied by the generic training recipe.

block_size

int

confidence_head_alpha

float

confidence_head_with_markov

bool

markov_head_type

str

markov_rank

int

mask_token_id

int

num_anchors

int

num_draft_layers

int

target_layer_ids

list[int]

nemo_automodel.components.models.minimax_m3_vl.config._json_safe_value(
    value: typing.Any
) -> typing.Any

Convert config values that are valid in-memory but not JSON serializable.

nemo_automodel.components.models.minimax_m3_vl.config._DSPARK_DRAFT_ATTN_IMPLEMENTATION = 'flex_attention'

nemo_automodel.components.models.minimax_m3_vl.config._DSPARK_RELEASED_ARCHITECTURE = 'DSparkMiniMaxDraftModel'

nemo_automodel.components.models.minimax_m3_vl.config._DSPARK_RELEASED_BACKBONE_FIELDS = ('hidden_size', 'num_attention_heads', 'num_key_value_heads', 'head_dim', 'rms_n...

nemo_automodel.components.models.minimax_m3_vl.config._DSPARK_RELEASED_REQUIRED_FIELDS = (*_DSPARK_RELEASED_BACKBONE_FIELDS, 'dspark_block_size', 'dspark_ffn_hidden_size...

nemo_automodel.components.models.minimax_m3_vl.config.logger = logging.getLogger(__name__)