nemo_automodel.components.models.hy_v3.layers

Module Contents

Classes

Name	Description
`HYV3Attention`	HYV3 attention with GQA, per-head Q/K RMSNorm, and RoPE.

API

class nemo_automodel.components.models.hy_v3.layers.HYV3Attention(
    config: typing.Any,
    backend: nemo_automodel.components.models.common.BackendConfig
)

Bases: Module

HYV3 attention with GQA, per-head Q/K RMSNorm, and RoPE.

Architecture:

q_proj: [hidden, n_heads * head_dim]
k_proj / v_proj: [hidden, n_kv_heads * head_dim]
q_norm / k_norm: RMSNorm applied per-head before RoPE
RoPE applied after per-head norm

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

v_proj

nemo_automodel.components.models.hy_v3.layers.HYV3Attention.forward(
    x: torch.Tensor,
    freqs_cis: torch.Tensor,
    attention_mask: torch.Tensor | None = None,
    attn_kwargs: typing.Any = {}
) -> torch.Tensor

nemo_automodel.components.models.hy_v3.layers.HYV3Attention.init_weights(
    buffer_device: torch.device,
    init_std: float = 0.02
)