mstar.model.command_a_plus.config

Classes

CommandAPlusConfig(text_config, vision_config)	The outer Cohere2 Vision checkpoint configuration.
CommandAPlusTextConfig(vocab_size, ...)
CommandAPlusVisionConfig(hidden_size, ...)	The SigLIP tower, pixel shuffle and SwiGLU projector that feed image tiles to the LLM.

class mstar.model.command_a_plus.config.CommandAPlusConfig(text_config, vision_config)

Bases: object

The outer Cohere2 Vision checkpoint configuration.

Parameters:

• text_config (CommandAPlusTextConfig)

• vision_config (CommandAPlusVisionConfig)

classmethod from_dict(data)

Parameters:

data (dict)

Return type:

CommandAPlusConfig

classmethod from_json(path)

Load a local config.json without resolving or downloading weights.

Parameters:

path (str | Path)

Return type:

CommandAPlusConfig

text_config: CommandAPlusTextConfig

vision_config: CommandAPlusVisionConfig

class mstar.model.command_a_plus.config.CommandAPlusTextConfig(vocab_size: int, hidden_size: int, num_hidden_layers: int, num_attention_heads: int, num_key_value_heads: int, head_dim: int, intermediate_size: int, num_experts: int, num_experts_per_tok: int, num_shared_experts: int, layer_types: list[str], sliding_window: int, max_position_embeddings: int, rope_theta: float, layer_norm_eps: float, logit_scale: float, bos_token_id: int, eos_token_id: int, pad_token_id: int)

Bases: object

Parameters:

• vocab_size (int)

• hidden_size (int)

• num_hidden_layers (int)

• num_attention_heads (int)

• num_key_value_heads (int)

• head_dim (int)

• intermediate_size (int)

• num_experts (int)

• num_experts_per_tok (int)

• num_shared_experts (int)

• layer_types (list[str])

• sliding_window (int)

• max_position_embeddings (int)

• rope_theta (float)

• layer_norm_eps (float)

• logit_scale (float)

• bos_token_id (int)

• eos_token_id (int)

• pad_token_id (int)

bos_token_id: int

eos_token_id: int

classmethod from_dict(data)

Parse the text config; shapes are checked when weights load, so only the math is pinned here.

Parameters:

data (dict)

Return type:

CommandAPlusTextConfig

head_dim: int

hidden_size: int

intermediate_size: int

layer_norm_eps: float

layer_types: list[str]

logit_scale: float

max_position_embeddings: int

num_attention_heads: int

num_experts: int

num_experts_per_tok: int

num_hidden_layers: int

num_key_value_heads: int

num_shared_experts: int

pad_token_id: int

rope_theta: float

property shared_intermediate_size: int

Width of the shared MLP, derived from the expert dimensions.

sliding_window: int

vocab_size: int

class mstar.model.command_a_plus.config.CommandAPlusVisionConfig(hidden_size, intermediate_size, num_hidden_layers, num_attention_heads, image_size, patch_size, layer_norm_eps, downsample_factor, alignment_intermediate_size, image_token_id)

Bases: object

The SigLIP tower, pixel shuffle and SwiGLU projector that feed image tiles to the LLM.

Parameters:

• hidden_size (int)

• intermediate_size (int)

• num_hidden_layers (int)

• num_attention_heads (int)

• image_size (int)

• patch_size (int)

• layer_norm_eps (float)

• downsample_factor (int)

• alignment_intermediate_size (int)

• image_token_id (int)

alignment_intermediate_size: int

downsample_factor: int

classmethod from_dict(data)

data is the whole checkpoint config; the tower’s own fields sit in vision_config.

Parameters:

data (dict)

Return type:

CommandAPlusVisionConfig

hidden_size: int

image_size: int

image_token_id: int

intermediate_size: int

layer_norm_eps: float

num_attention_heads: int

num_hidden_layers: int

patch_size: int

property tokens_per_tile: int