nemo_automodel.components.models.qwen3_next.model

Module Contents

Classes

Name	Description
`Block`	-
`Qwen3NextForCausalLM`	-
`Qwen3NextModel`	-

Data

ModelClass

API

class nemo_automodel.components.models.qwen3_next.model.Block(
    layer_idx: int,
    config: transformers.models.qwen3_next.configuration_qwen3_next.Qwen3NextConfig,
    moe_config: nemo_automodel.components.moe.config.MoEConfig,
    backend: nemo_automodel.components.models.common.BackendConfig
)

Bases: Module

input_layernorm

layer_type

= config.layer_types[layer_idx]

linear_attn

= Qwen3NextFp32GatedDeltaNet(config, layer_idx)

mlp

= MoE(moe_config, backend)

post_attention_layernorm

self_attn

= Qwen3NextAttention(config, layer_idx, backend)

nemo_automodel.components.models.qwen3_next.model.Block._mlp(
    x: torch.Tensor,
    padding_mask: torch.Tensor | None
) -> torch.Tensor

nemo_automodel.components.models.qwen3_next.model.Block.forward(
    x: torch.Tensor,
    freqs_cis: torch.Tensor,
    attention_mask: torch.Tensor | None = None,
    padding_mask: torch.Tensor | None = None,
    position_ids: torch.Tensor | None = None,
    attn_kwargs: typing.Any = {}
) -> torch.Tensor

nemo_automodel.components.models.qwen3_next.model.Block.init_weights(
    buffer_device: torch.device
)

class nemo_automodel.components.models.qwen3_next.model.Qwen3NextForCausalLM(
    config: transformers.models.qwen3_next.configuration_qwen3_next.Qwen3NextConfig,
    moe_config: nemo_automodel.components.moe.config.MoEConfig | None = None,
    backend: nemo_automodel.components.models.common.BackendConfig | None = None,
    kwargs = {}
)

Bases: HFCheckpointingMixin, Module, MoEFSDPSyncMixin

_keep_in_fp32_modules_strict

= ['_fp32_params']

backend

= backend or BackendConfig()

lm_head

model

state_dict_adapter

tie_word_embeddings_support

TieSupport = TieSupport.UNTIED_ONLY

nemo_automodel.components.models.qwen3_next.model.Qwen3NextForCausalLM.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: bool | None = None,
    attn_kwargs: typing.Any = {}
) -> transformers.modeling_outputs.CausalLMOutputWithPast

Run Qwen3Next without unsupported packed FA4 recurrent attention.

Parameters:

input_ids

torch.Tensor

Token IDs of shape [batch, sequence].

position_ids

torch.Tensor | NoneDefaults to None

Optional positions of shape [batch, sequence].

attention_mask

torch.Tensor | NoneDefaults to None

Padding/document IDs of shape [batch, sequence] or an explicit mask of shape [batch, 1, sequence, sequence].

padding_mask

torch.Tensor | NoneDefaults to None

Optional padding mask of shape [batch, sequence].

logits_to_keep

Union[int, torch.Tensor]Defaults to 0

Number of trailing logits, or indices of shape [kept_tokens].

output_hidden_states

bool | NoneDefaults to None

Whether to return the final hidden states.

**attn_kwargs

AnyDefaults to {}

Attention metadata, including cumulative document lengths of shape [documents + 1] and packed indices of shape [tokens].

Returns: CausalLMOutputWithPast

Model output with logits of shape [batch, kept_sequence, vocab] and

Raises:

ValueError: If packed FA4 is requested, since recurrent layers do not reset state at document boundaries.

nemo_automodel.components.models.qwen3_next.model.Qwen3NextForCausalLM.from_config(
    config: transformers.models.qwen3_next.configuration_qwen3_next.Qwen3NextConfig,
    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.qwen3_next.model.Qwen3NextForCausalLM.from_pretrained(
    pretrained_model_name_or_path: str,
    model_args = (),
    kwargs = {}
)

classmethod

nemo_automodel.components.models.qwen3_next.model.Qwen3NextForCausalLM.get_input_embeddings()

nemo_automodel.components.models.qwen3_next.model.Qwen3NextForCausalLM.get_output_embeddings()

nemo_automodel.components.models.qwen3_next.model.Qwen3NextForCausalLM.initialize_weights(
    buffer_device: torch.device | None = None,
    dtype: torch.dtype = torch.bfloat16
) -> None

nemo_automodel.components.models.qwen3_next.model.Qwen3NextForCausalLM.set_input_embeddings(
    value
)

nemo_automodel.components.models.qwen3_next.model.Qwen3NextForCausalLM.set_output_embeddings(
    new_embeddings
)

class nemo_automodel.components.models.qwen3_next.model.Qwen3NextModel(
    config: transformers.models.qwen3_next.configuration_qwen3_next.Qwen3NextConfig,
    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.qwen3_next.model.Qwen3NextModel.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.qwen3_next.model.Qwen3NextModel.init_weights(
    buffer_device: torch.device | None = None
) -> None

nemo_automodel.components.models.qwen3_next.model.ModelClass = Qwen3NextForCausalLM