nemo_automodel.components.models.hy_mt2.layers

Module Contents

Classes

Name	Description
`HyMT2Attention`	Hy-MT2-30B-A3B attention: GQA, per-head Q/K RMSNorm, and RoPE.

API

class nemo_automodel.components.models.hy_mt2.layers.HyMT2Attention(
    config: typing.Any,
    backend: nemo_automodel.components.models.common.BackendConfig
)

Bases: Module

Hy-MT2-30B-A3B attention: GQA, per-head Q/K RMSNorm, and RoPE.

Differences vs. the existing Hy3-preview HYV3Attention:

qk_norm is gated by config.qk_norm (defaults to True). For Hy-MT2-30B-A3B this is always True; the flag is here so the same module can also be reused for non-qk-norm variants without code edits.
Dimensions follow Hy-MT2-30B-A3B: 32 Q heads / 4 KV heads, head_dim=128, hidden_size=2048.

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

qk_norm_enabled

= bool(getattr(config, 'qk_norm', True))

v_proj

nemo_automodel.components.models.hy_mt2.layers.HyMT2Attention.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_mt2.layers.HyMT2Attention.init_weights(
    buffer_device: torch.device,
    init_std: float = 0.02
)