Qwen3MoeBackbone model

[source]

Qwen3MoeBackbone class

keras_hub.models.Qwen3MoeBackbone(
    vocabulary_size,
    num_layers,
    num_query_heads,
    num_key_value_heads,
    hidden_dim,
    intermediate_dim,
    moe_intermediate_dim,
    num_experts,
    head_dim=None,
    top_k=4,
    norm_top_k_prob=False,
    decoder_sparse_step=1,
    rope_max_wavelength=10000,
    rope_scaling_factor=1.0,
    layer_norm_epsilon=1e-06,
    dropout=0,
    dtype=None,
    tie_word_embeddings=False,
    sliding_window_size=32768,
    router_aux_loss_coefficient=0.001,
    mlp_only_layers=None,
    training=None,
    **kwargs
)

Qwen3 MoE core network with hyperparameters.

This backbone implements the base Transformer network for the Qwen MoE model. It includes embedding lookups and transformer layers with a Mixture of Experts (MoE) architecture, where each layer uses a sparse set of experts for efficient computation. This backbone outputs the final hidden states for each token, not generative predictions over the vocabulary space. For higher -level object for text generation, see keras_hub.models.Qwen3MoeCausalLM.

The default constructor gives a fully customizable, randomly initialized Qwen MoE model with any number of layers, heads, and embedding dimensions. To load preset architectures and weights, use the from_preset constructor.

Arguments

• vocabulary_size: int. The size of the token vocabulary.
• num_layers: int. The number of transformer layers.
• num_query_heads: int. The number of heads for the query projections in the attention layer.
• num_key_value_heads: int. The number of heads for the key and value projections in the attention layer.
• hidden_dim: int. The size of the transformer hidden state at the end of each transformer layer.
• intermediate_dim: int. The output dimension of the first Dense layer in the feedforward network for each transformer.
• moe_intermediate_dim: int. The intermediate dimension for each expert in the MoE feedforward network.
• num_experts: int. The number of experts in each MoE layer.
• top_k: int. The number of top experts to select for each token in the MoE layer.
• head_dim: int. The size of each attention head.
• layer_norm_epsilon: float. The epsilon value used for every layer norm in the transformer model.
• dropout: float. Dropout probability for the transformer encoder.
• sliding_window_size: int. Size of the sliding local window. Defaults to 4096.
• max_sequence_length: int. The maximum sequence length supported by the model. Defaults to 4096.
• dtype: str or keras.mixed_precision.DTypePolicy. The dtype to use for the model's computations and weights. Note that some computations, such as softmax and layer normalization, will always be done at float32 precision regardless of dtype.

Example

input_data = {
    "token_ids": np.ones(shape=(1, 12), dtype="int32"),
    "padding_mask": np.array([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]]),
}

__Pretrained Qwen MoE decoder__

.

model = keras_hub.models.Qwen3MoeBackbone.from_preset("qwen3_moe_a2_7b")
model(input_data)

__Randomly initialized Qwen MoE decoder with custom config__

.

model = keras_hub.models.Qwen3MoeBackbone(
    vocabulary_size=151936,
    num_layers=28,
    num_query_heads=16,
    num_key_value_heads=8,
    hidden_dim=2048,
    intermediate_dim=4096,
    moe_intermediate_dim=128,
    num_experts=60,
    top_k=4,
    head_dim=128,
    max_sequence_length=4096,
)
model(input_data)


----

<span style="float:right;">[[source]](https://github.com/keras-team/keras-hub/tree/v0.23.0/keras_hub/src/models/backbone.py#L117)</span>

### `from_preset` method


```python
Qwen3MoeBackbone.from_preset(preset, load_weights=True, **kwargs)

Instantiate a keras_hub.models.Backbone from a model preset.

A preset is a directory of configs, weights and other file assets used to save and load a pre-trained model. The preset can be passed as a one of:

1. a built-in preset identifier like 'bert_base_en'
2. a Kaggle Models handle like 'kaggle://user/bert/keras/bert_base_en'
3. a Hugging Face handle like 'hf://user/bert_base_en'
4. a ModelScope handle like 'modelscope://user/bert_base_en'
5. a path to a local preset directory like './bert_base_en'

This constructor can be called in one of two ways. Either from the base class like keras_hub.models.Backbone.from_preset(), or from a model class like keras_hub.models.GemmaBackbone.from_preset(). If calling from the base class, the subclass of the returning object will be inferred from the config in the preset directory.

For any Backbone subclass, you can run cls.presets.keys() to list all built-in presets available on the class.

Arguments

• preset: string. A built-in preset identifier, a Kaggle Models handle, a Hugging Face handle, or a path to a local directory.
• load_weights: bool. If True, the weights will be loaded into the model architecture. If False, the weights will be randomly initialized.

Examples

# Load a Gemma backbone with pre-trained weights.
model = keras_hub.models.Backbone.from_preset(
    "gemma_2b_en",
)

# Load a Bert backbone with a pre-trained config and random weights.
model = keras_hub.models.Backbone.from_preset(
    "bert_base_en",
    load_weights=False,
)

token_embedding property

keras_hub.models.Qwen3MoeBackbone.token_embedding

A keras.layers.Embedding instance for embedding token ids.

This layer embeds integer token ids to the hidden dim of the model.

[source]

enable_lora method

Qwen3MoeBackbone.enable_lora(rank, target_layer_names=None)

Enable Lora on the backbone.

Calling this method will freeze all weights on the backbone, while enabling Lora on the query & value EinsumDense layers of the attention layers.

Arguments

• rank: The rank of the LoRA factorization.
• target_layer_names: A list of strings, the names of the layers to apply LoRA to. If None, this will be populated with the default LoRA layer names as returned by backbone.default_lora_layer_names().