nemo_automodel.components.models.minimax_m2.layers

Module Contents

Classes

Name	Description
`MiniMaxM2Attention`	MiniMax-M2 attention with optional Q/K RMSNorm and partial RoPE.

API

class nemo_automodel.components.models.minimax_m2.layers.MiniMaxM2Attention(
    config: typing.Any,
    backend: nemo_automodel.components.models.common.BackendConfig
)

Bases: Module

MiniMax-M2 attention with optional Q/K RMSNorm and partial RoPE.

head_dim

k_norm

k_proj

num_heads

= config.num_attention_heads

num_kv_heads

= config.num_key_value_heads

o_proj

q_norm

q_proj

use_qk_norm

= getattr(config, 'use_qk_norm', False)

v_proj

nemo_automodel.components.models.minimax_m2.layers.MiniMaxM2Attention.forward(
    x: torch.Tensor,
    freqs_cis: torch.Tensor,
    attention_mask: torch.Tensor | None = None,
    attn_kwargs: typing.Any = {}
) -> torch.Tensor

nemo_automodel.components.models.minimax_m2.layers.MiniMaxM2Attention.init_weights(
    buffer_device: torch.device,
    init_std: float = 0.02
)