nemo_automodel.components.models.hy_v3.model

HYV3ForCausalLM — Tencent Hy3-preview (295B MoE) SFT support.

Architecture (from tencent/Hy3-preview config.json):

80 transformer layers; layer 0 is dense, layers 1-79 are MoE
MoE: 192 routed experts + 1 shared expert, top-8 activated
Sigmoid routing with expert-bias correction (e_score_correction_bias)
GQA: 64 Q heads, 8 KV heads, head_dim=128
Per-head QK RMSNorm before RoPE
256K context, rope_theta=11158840

Module Contents

Classes

Name	Description
`Block`	-
`HYV3ForCausalLM`	-
`HYV3Model`	-

Data

ModelClass

API

class nemo_automodel.components.models.hy_v3.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

input_layernorm

mlp

post_attention_layernorm

self_attn

= HYV3Attention(config, backend)

nemo_automodel.components.models.hy_v3.model.Block._mlp(
    x: torch.Tensor,
    padding_mask: torch.Tensor | None
) -> torch.Tensor

nemo_automodel.components.models.hy_v3.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_v3.model.Block.init_weights(
    buffer_device: torch.device
)

class nemo_automodel.components.models.hy_v3.model.HYV3ForCausalLM(
    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

backend

= backend or BackendConfig()

lm_head

model

state_dict_adapter

nemo_automodel.components.models.hy_v3.model.HYV3ForCausalLM.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:~transformers.modeling_outputs.CausalLMOutputWithPast.

Parameters:

input_ids

torch.Tensor

Input 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 [B, S].

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; otherwise only compute logits for the last logits_to_keep token positions (avoids materialising the full logit matrix during generation).

output_hidden_states

Optional[bool]Defaults to None

Whether to carry the final hidden states on the output.

**attn_kwargs

AnyDefaults to {}

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

Returns: CausalLMOutputWithPast

class:~transformers.modeling_outputs.CausalLMOutputWithPast with logits

nemo_automodel.components.models.hy_v3.model.HYV3ForCausalLM.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_v3.model.HYV3ForCausalLM.from_pretrained(
    pretrained_model_name_or_path: str,
    model_args = (),
    kwargs = {}
)

classmethod

nemo_automodel.components.models.hy_v3.model.HYV3ForCausalLM.get_input_embeddings()

nemo_automodel.components.models.hy_v3.model.HYV3ForCausalLM.get_output_embeddings()

nemo_automodel.components.models.hy_v3.model.HYV3ForCausalLM.initialize_weights(
    buffer_device: torch.device | None = None,
    dtype: torch.dtype = torch.bfloat16
) -> None

nemo_automodel.components.models.hy_v3.model.HYV3ForCausalLM.set_input_embeddings(
    value
)

nemo_automodel.components.models.hy_v3.model.HYV3ForCausalLM.set_output_embeddings(
    new_embeddings
)

nemo_automodel.components.models.hy_v3.model.HYV3ForCausalLM.update_moe_gate_bias() -> None

class nemo_automodel.components.models.hy_v3.model.HYV3Model(
    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

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_v3.model.HYV3Model.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_v3.model.HYV3Model.init_weights(
    buffer_device: torch.device | None = None
) -> None

nemo_automodel.components.models.hy_v3.model.ModelClass = HYV3ForCausalLM