nemo_automodel.components.models.hy_mt2.model

HyMT2ForCausalLM — Tencent Hy-MT2-30B-A3B (translation MoE) SFT support.

Architecture (from tencent/Hy-MT2-30B-A3B config.json):

48 transformer layers; layer 0 is dense, layers 1-47 are MoE
MoE: 128 routed experts + 1 shared expert, top-8 activated
Sigmoid routing with expert-bias correction, router_scaling_factor=2.826
route_norm = True (normalize top-k weights to sum to 1)
GQA: 32 Q heads, 4 KV heads, head_dim=128, hidden_size=2048
Per-head Q/K RMSNorm (qk_norm=True) before RoPE
256K context, rope_theta=11158840
dense intermediate_size=6912, moe_intermediate_size=expert_hidden_dim=768
vocab_size=120832
enable_lm_head_fp32 = True (HF reference upcasts lm_head to fp32)

Notes vs. components/models/hy_v3 (Hy3-preview 295B):

Smaller everywhere (48L / 128 experts / 32+4 heads / hidden=2048).
Adds an in-model enable_lm_head_fp32 fallback (applies when the YAML’s lm_head_precision is not set). The preferred path is to set distributed.moe.lm_head_precision: float32 in the YAML, which the MoE parallelizer handles via MixedPrecisionPolicy.
score_func is driven by config.moe_router_use_sigmoid instead of being hard-coded.

Module Contents

Classes

Name	Description
`Block`	Single Hy-MT2 transformer block: attention + (dense MLP \| MoE) + residual norms.
`HyMT2ForCausalLM`	Hy-MT2-30B-A3B causal-LM wrapper.
`HyMT2Model`	Hy-MT2 backbone: token embeddings + transformer blocks + final RMSNorm.

Functions

Name	Description
`_resolve_score_func`	Map `config.moe_router_use_sigmoid` to a gate `score_func` name.

Data

ModelClass

API

class nemo_automodel.components.models.hy_mt2.model.Block(
    layer_idx: int,
    config: typing.Any,
    moe_config: nemo_automodel.components.moe.config.MoEConfig,
    backend: nemo_automodel.components.models.common.BackendConfig
)

Bases: Module

Single Hy-MT2 transformer block: attention + (dense MLP | MoE) + residual norms.

input_layernorm

mlp

post_attention_layernorm

self_attn

= HyMT2Attention(config, backend)

nemo_automodel.components.models.hy_mt2.model.Block._mlp(
    x: torch.Tensor,
    padding_mask: torch.Tensor | None
) -> torch.Tensor

nemo_automodel.components.models.hy_mt2.model.Block.forward(
    x: torch.Tensor,
    freqs_cis: torch.Tensor,
    attention_mask: torch.Tensor | None = None,
    padding_mask: torch.Tensor | None = None,
    attn_kwargs: typing.Any = {}
) -> torch.Tensor

nemo_automodel.components.models.hy_mt2.model.Block.init_weights(
    buffer_device: torch.device
)

class nemo_automodel.components.models.hy_mt2.model.HyMT2ForCausalLM(
    config: typing.Any,
    moe_config: nemo_automodel.components.moe.config.MoEConfig | None = None,
    backend: nemo_automodel.components.models.common.BackendConfig | None = None,
    kwargs = {}
)

Bases: HFCheckpointingMixin, Module, MoEFSDPSyncMixin

Hy-MT2-30B-A3B causal-LM wrapper.

Mixes in MoEFSDPSyncMixin so EP / FSDP2 expert-gradient sync works out of the box (set distributed.ep_size in the YAML; must divide num_experts=128). The HFCheckpointingMixin provides from_pretrained / save_pretrained over the HF safetensors layout.

_enable_lm_head_fp32

backend

= backend or BackendConfig()

lm_head

model

state_dict_adapter

nemo_automodel.components.models.hy_mt2.model.HyMT2ForCausalLM.forward(
    input_ids: torch.Tensor,
    position_ids: torch.Tensor | None = None,
    attention_mask: torch.Tensor | None = None,
    padding_mask: torch.Tensor | None = None,
    logits_to_keep: typing.Union[int, torch.Tensor] = 0,
    output_hidden_states: typing.Optional[bool] = None,
    attn_kwargs: typing.Any = {}
) -> transformers.modeling_outputs.CausalLMOutputWithPast

Forward pass returning :class:CausalLMOutputWithPast.

Parameters:

input_ids

torch.Tensor

Token IDs. BSHD: [B, S]; THD: [1, T] (squeezed internally).

position_ids

torch.Tensor | NoneDefaults to None

Optional position indices.

attention_mask

torch.Tensor | NoneDefaults to None

Optional 2D padding mask.

padding_mask

torch.Tensor | NoneDefaults to None

Optional padding mask used by the THD squeeze helper.

logits_to_keep

Union[int, torch.Tensor]Defaults to 0

If 0 (default) compute logits for all positions; if > 0 only compute logits for the last logits_to_keep positions (avoids materialising the full logit matrix during generation / enables fused-linear cross-entropy in training).

output_hidden_states

Optional[bool]Defaults to None

Whether to return the final hidden states (the input to lm_head) on the output. Defaults to the config flag.

**attn_kwargs

AnyDefaults to {}

Additional attention kwargs forwarded to the base model (e.g. qkv_format, cu_seqlens, CP kwargs).

Returns: CausalLMOutputWithPast

class:~transformers.modeling_outputs.CausalLMOutputWithPast with

nemo_automodel.components.models.hy_mt2.model.HyMT2ForCausalLM.from_config(
    config: typing.Any,
    moe_config: nemo_automodel.components.moe.config.MoEConfig | None = None,
    backend: nemo_automodel.components.models.common.BackendConfig | None = None,
    kwargs = {}
)

classmethod

nemo_automodel.components.models.hy_mt2.model.HyMT2ForCausalLM.from_pretrained(
    pretrained_model_name_or_path: str,
    model_args = (),
    kwargs = {}
)

classmethod

nemo_automodel.components.models.hy_mt2.model.HyMT2ForCausalLM.get_input_embeddings()

nemo_automodel.components.models.hy_mt2.model.HyMT2ForCausalLM.get_output_embeddings()

nemo_automodel.components.models.hy_mt2.model.HyMT2ForCausalLM.initialize_weights(
    buffer_device: torch.device | None = None,
    dtype: torch.dtype = torch.bfloat16
) -> None

nemo_automodel.components.models.hy_mt2.model.HyMT2ForCausalLM.set_input_embeddings(
    value
)

nemo_automodel.components.models.hy_mt2.model.HyMT2ForCausalLM.set_output_embeddings(
    new_embeddings
)

nemo_automodel.components.models.hy_mt2.model.HyMT2ForCausalLM.update_moe_gate_bias() -> None

class nemo_automodel.components.models.hy_mt2.model.HyMT2Model(
    config: typing.Any,
    backend: nemo_automodel.components.models.common.BackendConfig,
    moe_config: nemo_automodel.components.moe.config.MoEConfig | None = None,
    moe_overrides: dict | None = None
)

Bases: Module

Hy-MT2 backbone: token embeddings + transformer blocks + final RMSNorm.

The MoE / dense split is governed by config.first_k_dense_replace (layer 0 dense, the rest MoE for the published Hy-MT2-30B-A3B). The MoE configuration is assembled from the HF config fields and forwarded to every MoE-bearing Block.

embed_tokens

head_dim

layers

= torch.nn.ModuleDict()

max_seq_len

= config.max_position_embeddings

moe_config

= moe_config or MoEConfig(**moe_defaults)

norm

rotary_emb

nemo_automodel.components.models.hy_mt2.model.HyMT2Model.forward(
    input_ids: torch.Tensor,
    position_ids: torch.Tensor | None = None,
    attention_mask: torch.Tensor | None = None,
    padding_mask: torch.Tensor | None = None,
    attn_kwargs: typing.Any = {}
) -> torch.Tensor

nemo_automodel.components.models.hy_mt2.model.HyMT2Model.init_weights(
    buffer_device: torch.device | None = None
) -> None

nemo_automodel.components.models.hy_mt2.model._resolve_score_func(
    config: typing.Any
) -> str

Map config.moe_router_use_sigmoid to a gate score_func name.

Returns “sigmoid” when the flag is True (Hy-MT2 default) and “softmax” otherwise. The bias-aware variants (“sigmoid_with_bias” / “softmax_with_bias”) are selected at the gate level by the presence of e_score_correction_bias plus expert-group routing, which Hy-MT2 does not use (n_expert_groups=0).

nemo_automodel.components.models.hy_mt2.model.ModelClass = HyMT2ForCausalLM