nemo_automodel.components.models.ling_v2.layers

Attention layer for BailingMoeV2 (Ling 2.0).

GQA + per-head QK-RMSNorm + partial RoPE. Equivalent to Qwen3-MoE attention with an additional partial_rotary_factor knob that rotates only the first head_dim * partial_rotary_factor channels and passes the rest through (GPT-J / GPT-NeoX half-RoPE).

Module Contents

Classes

Name	Description
`BailingMoeV2Attention`	Bailing MoE V2 attention block.

API

class nemo_automodel.components.models.ling_v2.layers.BailingMoeV2Attention(
    config,
    backend: nemo_automodel.components.models.common.BackendConfig
)

Bases: Module

Bailing MoE V2 attention block.

Shapes:

Input x: [B, S, H] (or [T, H] in THD format).
Projections: q: [B, S, n_heads, head_dim] k, v: [B, S, n_kv_heads, head_dim]
Output: [B, S, H].

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

= bool(getattr(config, 'use_qk_norm', False))

v_proj

nemo_automodel.components.models.ling_v2.layers.BailingMoeV2Attention.forward(
    x: torch.Tensor,
    freqs_cis: torch.Tensor,
    attention_mask: torch.Tensor | None = None,
    attn_kwargs: typing.Any = {}
) -> torch.Tensor

nemo_automodel.components.models.ling_v2.layers.BailingMoeV2Attention.init_weights(
    buffer_device: torch.device,
    init_std: float = 0.02
) -> None