nemo_automodel.components.models.glm4_moe.layers

Module Contents

Classes

Name	Description
`Glm4MoeAttention`	GLM4 MoE attention with optional query/key per-head RMSNorm + partial RoPE.

API

class nemo_automodel.components.models.glm4_moe.layers.Glm4MoeAttention(
    config: transformers.models.glm4_moe.configuration_glm4_moe.Glm4MoeConfig,
    backend: nemo_automodel.components.models.common.BackendConfig
)

Bases: Module

GLM4 MoE attention with optional query/key per-head RMSNorm + partial RoPE.

Key differences from Qwen3 MoE:

Optional QK normalization (controlled by use_qk_norm)
Partial rotary embeddings (controlled by partial_rotary_factor)
o_proj bias is False (unlike Qwen3 which has configurable attention_bias)

Shapes:

Input: x -> [B, S, H]
Projections: q: [B, S, n_heads, head_dim] k/v: [B, S, n_kv_heads, head_dim] -> repeated to n_heads via groups
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

partial_rotary_factor

= getattr(config, 'partial_rotary_factor', 0.5)

q_norm

q_proj

use_qk_norm

= config.use_qk_norm

v_proj

nemo_automodel.components.models.glm4_moe.layers.Glm4MoeAttention.forward(
    x: torch.Tensor,
    freqs_cis: torch.Tensor,
    attention_mask: torch.Tensor | None = None,
    attn_kwargs: typing.Any = {}
) -> torch.Tensor

nemo_automodel.components.models.glm4_moe.layers.Glm4MoeAttention.init_weights(
    buffer_device: torch.device,
    init_std: float = 0.02
)