# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # This file was automatically generated from src/transformers/models/qwen3_omni_moe/modular_qwen3_omni_moe.py. # Do NOT edit this file manually as any edits will be overwritten by the generation of # the file from the modular. If any change should be done, please apply the change to the # modular_qwen3_omni_moe.py file directly. One of our CI enforces this. # 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # coding=utf-8 # Copyright 2025 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved. # # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from ...configuration_utils import PretrainedConfig, layer_type_validation from ...modeling_rope_utils import rope_config_validation from ...utils import logging logger = logging.get_logger(__name__) class Qwen3OmniMoeAudioEncoderConfig(PretrainedConfig): r""" This is the configuration class to store the configuration of a [`Qwen3OmniMoeAudioEncoder`]. It is used to instantiate a Qwen2.5-Omni-Thinker audio encoder according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of the audio encoder of the Qwen2-Audio architecture. e.g. [Qwen/Qwen2.5-Omni-7B](https://huggingface.co/Qwen/Qwen2.5-Omni-7B) Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the documentation from [`PretrainedConfig`] for more information. Args: num_mel_bins (`int`, *optional*, defaults to 128): Number of mel features used per input features. Should correspond to the value used in the `Qwen3OmniMoeProcessor` class. encoder_layers (`int`, *optional*, defaults to 32): Number of encoder layers. encoder_attention_heads (`int`, *optional*, defaults to 20): Number of attention heads for each attention layer in the Transformer encoder. encoder_ffn_dim (`int`, *optional*, defaults to 5120): Dimensionality of the "intermediate" (often named feed-forward) layer in encoder. d_model (`int`, *optional*, defaults to 1280): Dimensionality of the layers. dropout (`float`, *optional*, defaults to 0.0): The dropout probability for all fully connected layers in the embeddings, encoder, and pooler. attention_dropout (`float`, *optional*, defaults to 0.0): The dropout ratio for the attention probabilities. activation_function (`str`, *optional*, defaults to `"gelu"`): The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`, `"relu"`, `"silu"` and `"gelu_new"` are supported. activation_dropout (`float`, *optional*, defaults to 0.0): The dropout ratio for activations inside the fully connected layer. scale_embedding (`bool`, *optional*, defaults to `False`): Scale embeddings by diving by sqrt(d_model). initializer_range (`float`, *optional*, defaults to 0.02): The standard deviation of the truncated_normal_initializer for initializing all weight matrices. max_source_positions (`int`, *optional*, defaults to 1500): The maximum sequence length of log-mel filter-bank features that this model might ever be used with. n_window (`int`, *optional*, defaults to 100): The chunk for conv and flash attn in AudioEncoder. output_dim (`int`, *optional*, defaults to 3584): The output dimension of AudioEncoder. Example: ```python >>> from transformers import Qwen3OmniMoeAudioEncoderConfig, Qwen3OmniMoeAudioEncoder >>> # Initializing a Qwen3OmniMoeAudioEncoderConfig >>> configuration = Qwen3OmniMoeAudioEncoderConfig() >>> # Initializing a Qwen3OmniMoeAudioEncoder (with random weights) >>> model = Qwen3OmniMoeAudioEncoder(configuration) >>> # Accessing the model configuration >>> configuration = model.config ```""" model_type = "qwen3_omni_moe_audio_encoder" def __init__( self, num_mel_bins=128, encoder_layers=32, encoder_attention_heads=20, encoder_ffn_dim=5120, d_model=1280, dropout=0, attention_dropout=0, activation_function="gelu", activation_dropout=0, scale_embedding=False, initializer_range=0.02, max_source_positions=1500, n_window=100, output_dim=3584, n_window_infer=400, conv_chunksize=500, downsample_hidden_size=480, **kwargs, ): super().__init__(**kwargs) self.num_mel_bins = num_mel_bins self.d_model = d_model self.encoder_layers = encoder_layers self.encoder_attention_heads = encoder_attention_heads self.encoder_ffn_dim = encoder_ffn_dim self.dropout = dropout self.attention_dropout = attention_dropout self.activation_function = activation_function self.activation_dropout = activation_dropout self.num_hidden_layers = encoder_layers self.initializer_range = initializer_range self.scale_embedding = scale_embedding # scale factor will be sqrt(d_model) if True self.max_source_positions = max_source_positions self.n_window = n_window self.output_dim = output_dim self.n_window_infer = n_window_infer self.conv_chunksize = conv_chunksize self.downsample_hidden_size = downsample_hidden_size class Qwen3OmniMoeVisionEncoderConfig(PretrainedConfig): model_type = "qwen3_omni_moe_vision_encoder" base_config_key = "vision_config" def __init__( self, depth=27, hidden_size=1152, hidden_act="gelu_pytorch_tanh", intermediate_size=4304, num_heads=16, in_channels=3, patch_size=16, spatial_merge_size=2, temporal_patch_size=2, out_hidden_size=3584, num_position_embeddings=2304, deepstack_visual_indexes=[8, 16, 24], initializer_range=0.02, **kwargs, ): super().__init__(**kwargs) self.depth = depth self.hidden_size = hidden_size self.hidden_act = hidden_act self.intermediate_size = intermediate_size self.num_heads = num_heads self.in_channels = in_channels self.patch_size = patch_size self.spatial_merge_size = spatial_merge_size self.temporal_patch_size = temporal_patch_size self.out_hidden_size = out_hidden_size self.num_position_embeddings = num_position_embeddings self.initializer_range = initializer_range self.deepstack_visual_indexes = deepstack_visual_indexes class Qwen3OmniMoeTextConfig(PretrainedConfig): r""" This is the configuration class to store the configuration of a [`Qwen3OmniMoeTextModel`]. It is used to instantiate a Qwen3OmniMoeText model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of [Qwen/Qwen3-15B-A2B](https://huggingface.co/Qwen/Qwen3-15B-A2B). Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the documentation from [`PretrainedConfig`] for more information. Args: vocab_size (`int`, *optional*, defaults to 151936): Vocabulary size of the Qwen3OmniMoeText model. Defines the number of different tokens that can be represented by the `inputs_ids` passed when calling [`Qwen3OmniMoeTextModel`] hidden_size (`int`, *optional*, defaults to 2048): Dimension of the hidden representations. intermediate_size (`int`, *optional*, defaults to 6144): Dimension of the MLP representations. num_hidden_layers (`int`, *optional*, defaults to 24): Number of hidden layers in the Transformer encoder. num_attention_heads (`int`, *optional*, defaults to 32): Number of attention heads for each attention layer in the Transformer encoder. num_key_value_heads (`int`, *optional*, defaults to 4): This is the number of key_value heads that should be used to implement Grouped Query Attention. If `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed by meanpooling all the original heads within that group. For more details, check out [this paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to `32`. hidden_act (`str` or `function`, *optional*, defaults to `"silu"`): The non-linear activation function (function or string) in the decoder. max_position_embeddings (`int`, *optional*, defaults to 32768): The maximum sequence length that this model might ever be used with. initializer_range (`float`, *optional*, defaults to 0.02): The standard deviation of the truncated_normal_initializer for initializing all weight matrices. rms_norm_eps (`float`, *optional*, defaults to 1e-06): The epsilon used by the rms normalization layers. use_cache (`bool`, *optional*, defaults to `True`): Whether or not the model should return the last key/values attentions (not used by all models). Only relevant if `config.is_decoder=True`. tie_word_embeddings (`bool`, *optional*, defaults to `False`): Whether the model's input and output word embeddings should be tied. rope_theta (`float`, *optional*, defaults to 10000.0): The base period of the RoPE embeddings. rope_scaling (`Dict`, *optional*): Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value accordingly. Expected contents: `rope_type` (`str`): The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope', 'llama3'], with 'default' being the original RoPE implementation. `factor` (`float`, *optional*): Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In most scaling types, a `factor` of x will enable the model to handle sequences of length x * original maximum pre-trained length. `original_max_position_embeddings` (`int`, *optional*): Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during pretraining. `attention_factor` (`float`, *optional*): Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention computation. If unspecified, it defaults to value recommended by the implementation, using the `factor` field to infer the suggested value. `beta_fast` (`float`, *optional*): Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear ramp function. If unspecified, it defaults to 32. `beta_slow` (`float`, *optional*): Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear ramp function. If unspecified, it defaults to 1. `short_factor` (`list[float]`, *optional*): Only used with 'longrope'. The scaling factor to be applied to short contexts (< `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden size divided by the number of attention heads divided by 2 `long_factor` (`list[float]`, *optional*): Only used with 'longrope'. The scaling factor to be applied to long contexts (< `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden size divided by the number of attention heads divided by 2 `low_freq_factor` (`float`, *optional*): Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE `high_freq_factor` (`float`, *optional*): Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`): Whether to use a bias in the query, key, value and output projection layers during self-attention. use_sliding_window (`bool`, *optional*, defaults to `False`): Whether to use sliding window attention. sliding_window (`int`, *optional*, defaults to 4096): Sliding window attention (SWA) window size. If not specified, will default to `4096`. attention_dropout (`float`, *optional*, defaults to 0.0): The dropout ratio for the attention probabilities. decoder_sparse_step (`int`, *optional*, defaults to 1): The frequency of the MoE layer. moe_intermediate_size (`int`, *optional*, defaults to 768): Intermediate size of the routed expert. num_experts_per_tok (`int`, *optional*, defaults to 8): Number of selected experts. num_experts (`int`, *optional*, defaults to 128): Number of routed experts. norm_topk_prob (`bool`, *optional*, defaults to `False`): Whether to normalize the topk probabilities. output_router_logits (`bool`, *optional*, defaults to `False`): Whether or not the router logits should be returned by the model. Enabling this will also allow the model to output the auxiliary loss, including load balancing loss and router z-loss. router_aux_loss_coef (`float`, *optional*, defaults to 0.001): The aux loss factor for the total loss. mlp_only_layers (`list[int]`, *optional*, defaults to `[]`): Indicate which layers use Qwen3OmniMoeTextMLP rather than Qwen3OmniMoeTextSparseMoeBlock The list contains layer index, from 0 to num_layers-1 if we have num_layers layers If `mlp_only_layers` is empty, `decoder_sparse_step` is used to determine the sparsity. ```python >>> from transformers import Qwen3OmniMoeTextModel, Qwen3OmniMoeTextConfig >>> # Initializing a Qwen3OmniMoeText style configuration >>> configuration = Qwen3OmniMoeTextConfig() >>> # Initializing a model from the Qwen3-15B-A2B" style configuration >>> model = Qwen3OmniMoeTextModel(configuration) >>> # Accessing the model configuration >>> configuration = model.config ```""" model_type = "qwen3_omni_moe_text" keys_to_ignore_at_inference = ["past_key_values"] # Default tensor parallel plan for base model `Qwen3OmniMoeText` base_model_tp_plan = { "layers.*.self_attn.q_proj": "colwise", "layers.*.self_attn.k_proj": "colwise", "layers.*.self_attn.v_proj": "colwise", "layers.*.self_attn.o_proj": "rowwise", "layers.*.mlp.experts.*.gate_proj": "colwise", "layers.*.mlp.experts.*.up_proj": "colwise", "layers.*.mlp.experts.*.down_proj": "rowwise", "layers.*.mlp.gate_proj": "colwise", "layers.*.mlp.up_proj": "colwise", "layers.*.mlp.down_proj": "rowwise", } base_model_pp_plan = { "embed_tokens": (["input_ids"], ["inputs_embeds"]), "layers": (["hidden_states", "attention_mask"], ["hidden_states"]), "norm": (["hidden_states"], ["hidden_states"]), } def __init__( self, vocab_size=3584, hidden_size=2048, intermediate_size=18944, num_hidden_layers=28, num_attention_heads=28, num_key_value_heads=4, hidden_act="silu", max_position_embeddings=32768, initializer_range=0.02, rms_norm_eps=1e-6, use_cache=True, tie_word_embeddings=False, rope_theta=1000000.0, rope_scaling=None, attention_bias=False, sliding_window=None, attention_dropout=0, decoder_sparse_step=1, moe_intermediate_size=768, num_experts_per_tok=8, num_experts=128, norm_topk_prob=True, output_router_logits=False, router_aux_loss_coef=0.001, mlp_only_layers=None, **kwargs, ): super().__init__( tie_word_embeddings=tie_word_embeddings, **kwargs, ) self.vocab_size = vocab_size self.max_position_embeddings = max_position_embeddings self.hidden_size = hidden_size self.intermediate_size = intermediate_size self.num_hidden_layers = num_hidden_layers self.num_attention_heads = num_attention_heads self.sliding_window = sliding_window self.num_key_value_heads = num_key_value_heads self.hidden_act = hidden_act self.initializer_range = initializer_range self.rms_norm_eps = rms_norm_eps self.use_cache = use_cache self.rope_theta = rope_theta self.rope_scaling = rope_scaling self.attention_bias = attention_bias self.attention_dropout = attention_dropout # Validate the correctness of rotary position embeddings parameters # BC: if there is a 'type' field, move it to 'rope_type'. if self.rope_scaling is not None and "type" in self.rope_scaling: self.rope_scaling["rope_type"] = self.rope_scaling["type"] rope_config_validation(self) # MoE arguments self.decoder_sparse_step = decoder_sparse_step self.moe_intermediate_size = moe_intermediate_size self.num_experts_per_tok = num_experts_per_tok self.num_experts = num_experts self.norm_topk_prob = norm_topk_prob self.output_router_logits = output_router_logits self.router_aux_loss_coef = router_aux_loss_coef self.mlp_only_layers = [] if mlp_only_layers is None else mlp_only_layers class Qwen3OmniMoeThinkerConfig(PretrainedConfig): r""" This is the configuration class to store the configuration of a [`Qwen3OmniMoeThinker`]. It is used to instantiate a Qwen3-Omni-Thinker model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of the thinker component of the Qwen3-Omni architecture. e.g. [Qwen/Qwen3-Omni-7B](https://huggingface.co/Qwen/Qwen3-Omni-7B) Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the documentation from [`PretrainedConfig`] for more information. Args: audio_config (`dict`, *optional*): The config dictionary of the audio backbone. vision_config (`dict`, *optional*): The config dictionary of the vision backbone. text_config (`dict`, *optional*): The config dictionary of the text backbone. audio_token_id (`int`, *optional*, defaults to 151646): The audio token id to encode the audio prompt. image_token_id (`int`, *optional*, defaults to 151655): The image token id to encode the image prompt. video_token_id (`int`, *optional*, defaults to 151656): The video token id to encode the video prompt. position_id_per_seconds (`int`, *optional*, defaults to 25): The increment of position id per second. audio_start_token_id (`int`, *optional*, defaults to 151647): The audio start token id to encode the audio prompt. user_token_id (`int`, *optional*, defaults to 872): The user token id to encode the user token. initializer_range (`float`, *optional*, defaults to 0.02): The standard deviation of the truncated_normal_initializer for initializing all weight matrices. Example: ```python >>> from transformers import Qwen3OmniMoeThinkerModel, Qwen3OmniMoeThinkerConfig >>> # Initializing a default Qwen3OmniMoeThinkerConfig >>> configuration = Qwen3OmniMoeThinkerConfig() >>> # Initializing a model (with random weights) from the default configuration >>> model = Qwen3OmniMoeThinkerModel(configuration) >>> # Accessing the model configuration >>> configuration = model.config ```""" model_type = "qwen3_omni_moe_thinker" # Override parent's attribute_map as we use audio_token_id directly, not audio_token_index attribute_map = {} sub_configs = { "audio_config": Qwen3OmniMoeAudioEncoderConfig, "vision_config": Qwen3OmniMoeVisionEncoderConfig, "text_config": Qwen3OmniMoeTextConfig, } def __init__( self, audio_config=None, vision_config=None, text_config=None, audio_token_id=151646, image_token_id=151655, video_token_id=151656, position_id_per_seconds=25, audio_start_token_id=151647, user_token_id=872, initializer_range=0.02, **kwargs, ): super().__init__(**kwargs) self.user_token_id = user_token_id self.position_id_per_seconds = position_id_per_seconds self.audio_start_token_id = audio_start_token_id self.initializer_range = initializer_range if isinstance(vision_config, dict): vision_config = Qwen3OmniMoeVisionEncoderConfig(**vision_config) elif vision_config is None: vision_config = Qwen3OmniMoeVisionEncoderConfig() self.vision_config = vision_config if isinstance(audio_config, dict): audio_config = Qwen3OmniMoeAudioEncoderConfig(**audio_config) elif audio_config is None: audio_config = Qwen3OmniMoeAudioEncoderConfig() self.audio_config = audio_config if isinstance(text_config, dict): text_config = Qwen3OmniMoeTextConfig(**text_config) elif text_config is None: text_config = Qwen3OmniMoeTextConfig() self.text_config = text_config self.audio_token_id = audio_token_id self.image_token_id = image_token_id self.video_token_id = video_token_id class Qwen3OmniMoeTalkerCodePredictorConfig(PretrainedConfig): r""" This is the configuration class to store the configuration of a [`Qwen3OmniMoeTalkerCodePredictorModel`]. It is used to instantiate a Qwen3OmniMoeTalkerCodePredictor model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of Qwen3OmniMoeTalkerCodePredictor-8B [Qwen/Qwen3OmniMoeTalkerCodePredictor-8B](https://huggingface.co/Qwen/Qwen3OmniMoeTalkerCodePredictor-8B). Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the documentation from [`PretrainedConfig`] for more information. Args: vocab_size (`int`, *optional*, defaults to 151936): Vocabulary size of the Qwen3OmniMoeTalkerCodePredictor model. Defines the number of different tokens that can be represented by the `inputs_ids` passed when calling [`Qwen3OmniMoeTalkerCodePredictorModel`] hidden_size (`int`, *optional*, defaults to 4096): Dimension of the hidden representations. intermediate_size (`int`, *optional*, defaults to 22016): Dimension of the MLP representations. num_hidden_layers (`int`, *optional*, defaults to 32): Number of hidden layers in the Transformer encoder. num_attention_heads (`int`, *optional*, defaults to 32): Number of attention heads for each attention layer in the Transformer encoder. num_key_value_heads (`int`, *optional*, defaults to 32): This is the number of key_value heads that should be used to implement Grouped Query Attention. If `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed by meanpooling all the original heads within that group. For more details, check out [this paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to `32`. head_dim (`int`, *optional*, defaults to 128): The attention head dimension. hidden_act (`str` or `function`, *optional*, defaults to `"silu"`): The non-linear activation function (function or string) in the decoder. max_position_embeddings (`int`, *optional*, defaults to 32768): The maximum sequence length that this model might ever be used with. initializer_range (`float`, *optional*, defaults to 0.02): The standard deviation of the truncated_normal_initializer for initializing all weight matrices. rms_norm_eps (`float`, *optional*, defaults to 1e-06): The epsilon used by the rms normalization layers. use_cache (`bool`, *optional*, defaults to `True`): Whether or not the model should return the last key/values attentions (not used by all models). Only relevant if `config.is_decoder=True`. tie_word_embeddings (`bool`, *optional*, defaults to `False`): Whether the model's input and output word embeddings should be tied. rope_theta (`float`, *optional*, defaults to 10000.0): The base period of the RoPE embeddings. rope_scaling (`Dict`, *optional*): Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value accordingly. Expected contents: `rope_type` (`str`): The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope', 'llama3'], with 'default' being the original RoPE implementation. `factor` (`float`, *optional*): Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In most scaling types, a `factor` of x will enable the model to handle sequences of length x * original maximum pre-trained length. `original_max_position_embeddings` (`int`, *optional*): Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during pretraining. `attention_factor` (`float`, *optional*): Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention computation. If unspecified, it defaults to value recommended by the implementation, using the `factor` field to infer the suggested value. `beta_fast` (`float`, *optional*): Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear ramp function. If unspecified, it defaults to 32. `beta_slow` (`float`, *optional*): Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear ramp function. If unspecified, it defaults to 1. `short_factor` (`list[float]`, *optional*): Only used with 'longrope'. The scaling factor to be applied to short contexts (< `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden size divided by the number of attention heads divided by 2 `long_factor` (`list[float]`, *optional*): Only used with 'longrope'. The scaling factor to be applied to long contexts (< `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden size divided by the number of attention heads divided by 2 `low_freq_factor` (`float`, *optional*): Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE `high_freq_factor` (`float`, *optional*): Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`): Whether to use a bias in the query, key, value and output projection layers during self-attention. use_sliding_window (`bool`, *optional*, defaults to `False`): Whether to use sliding window attention. sliding_window (`int`, *optional*, defaults to 4096): Sliding window attention (SWA) window size. If not specified, will default to `4096`. max_window_layers (`int`, *optional*, defaults to 28): The number of layers using full attention. The first `max_window_layers` layers will use full attention, while any additional layer afterwards will use SWA (Sliding Window Attention). layer_types (`list`, *optional*): Attention pattern for each layer. attention_dropout (`float`, *optional*, defaults to 0.0): The dropout ratio for the attention probabilities. ```python >>> from transformers import Qwen3OmniMoeTalkerCodePredictorModel, Qwen3OmniMoeTalkerCodePredictorConfig >>> # Initializing a Qwen3OmniMoeTalkerCodePredictor style configuration >>> configuration = Qwen3OmniMoeTalkerCodePredictorConfig() >>> # Initializing a model from the Qwen3OmniMoeTalkerCodePredictor-8B style configuration >>> model = Qwen3OmniMoeTalkerCodePredictorModel(configuration) >>> # Accessing the model configuration >>> configuration = model.config ```""" model_type = "qwen3_omni_moe_talker_code_predictor" keys_to_ignore_at_inference = ["past_key_values"] # Default tensor parallel plan for base model `Qwen3OmniMoeTalkerCodePredictor` base_model_tp_plan = { "layers.*.self_attn.q_proj": "colwise", "layers.*.self_attn.k_proj": "colwise", "layers.*.self_attn.v_proj": "colwise", "layers.*.self_attn.o_proj": "rowwise", "layers.*.mlp.gate_proj": "colwise", "layers.*.mlp.up_proj": "colwise", "layers.*.mlp.down_proj": "rowwise", } base_model_pp_plan = { "embed_tokens": (["input_ids"], ["inputs_embeds"]), "layers": (["hidden_states", "attention_mask"], ["hidden_states"]), "norm": (["hidden_states"], ["hidden_states"]), } def __init__( self, vocab_size=2048, hidden_size=1024, intermediate_size=3072, num_hidden_layers=5, num_attention_heads=16, num_key_value_heads=8, head_dim=128, hidden_act="silu", max_position_embeddings=32768, initializer_range=0.02, rms_norm_eps=0.000001, use_cache=True, tie_word_embeddings=False, rope_theta=10000, rope_scaling=None, attention_bias=False, sliding_window=None, layer_types=None, attention_dropout=0, num_code_groups=32, **kwargs, ): super().__init__( tie_word_embeddings=tie_word_embeddings, **kwargs, ) self.vocab_size = vocab_size self.max_position_embeddings = max_position_embeddings self.hidden_size = hidden_size self.intermediate_size = intermediate_size self.num_hidden_layers = num_hidden_layers self.num_attention_heads = num_attention_heads self.sliding_window = sliding_window # for backward compatibility if num_key_value_heads is None: num_key_value_heads = num_attention_heads self.num_key_value_heads = num_key_value_heads self.head_dim = head_dim self.hidden_act = hidden_act self.initializer_range = initializer_range self.rms_norm_eps = rms_norm_eps self.use_cache = use_cache self.rope_theta = rope_theta self.rope_scaling = rope_scaling self.attention_bias = attention_bias self.attention_dropout = attention_dropout # Validate the correctness of rotary position embeddings parameters # BC: if there is a 'type' field, move it to 'rope_type'. if self.rope_scaling is not None and "type" in self.rope_scaling: self.rope_scaling["rope_type"] = self.rope_scaling["type"] rope_config_validation(self) self.layer_types = layer_types if self.layer_types is None: self.layer_types = [ "sliding_attention" if self.sliding_window is not None and i >= self.max_window_layers else "full_attention" for i in range(self.num_hidden_layers) ] layer_type_validation(self.layer_types, self.num_hidden_layers) self.num_code_groups = num_code_groups class Qwen3OmniMoeTalkerTextConfig(PretrainedConfig): r""" This is the configuration class to store the configuration of a [`Qwen3OmniMoeTalkerTextModel`]. It is used to instantiate a Qwen3OmniMoeTalkerText model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of [Qwen/Qwen3-15B-A2B](https://huggingface.co/Qwen/Qwen3-15B-A2B). Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the documentation from [`PretrainedConfig`] for more information. Args: vocab_size (`int`, *optional*, defaults to 151936): Vocabulary size of the Qwen3OmniMoeTalkerText model. Defines the number of different tokens that can be represented by the `inputs_ids` passed when calling [`Qwen3OmniMoeTalkerTextModel`] hidden_size (`int`, *optional*, defaults to 2048): Dimension of the hidden representations. intermediate_size (`int`, *optional*, defaults to 6144): Dimension of the MLP representations. num_hidden_layers (`int`, *optional*, defaults to 24): Number of hidden layers in the Transformer encoder. num_attention_heads (`int`, *optional*, defaults to 32): Number of attention heads for each attention layer in the Transformer encoder. num_key_value_heads (`int`, *optional*, defaults to 4): This is the number of key_value heads that should be used to implement Grouped Query Attention. If `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed by meanpooling all the original heads within that group. For more details, check out [this paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to `32`. hidden_act (`str` or `function`, *optional*, defaults to `"silu"`): The non-linear activation function (function or string) in the decoder. max_position_embeddings (`int`, *optional*, defaults to 32768): The maximum sequence length that this model might ever be used with. initializer_range (`float`, *optional*, defaults to 0.02): The standard deviation of the truncated_normal_initializer for initializing all weight matrices. rms_norm_eps (`float`, *optional*, defaults to 1e-06): The epsilon used by the rms normalization layers. use_cache (`bool`, *optional*, defaults to `True`): Whether or not the model should return the last key/values attentions (not used by all models). Only relevant if `config.is_decoder=True`. tie_word_embeddings (`bool`, *optional*, defaults to `False`): Whether the model's input and output word embeddings should be tied. rope_theta (`float`, *optional*, defaults to 10000.0): The base period of the RoPE embeddings. rope_scaling (`Dict`, *optional*): Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value accordingly. Expected contents: `rope_type` (`str`): The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope', 'llama3'], with 'default' being the original RoPE implementation. `factor` (`float`, *optional*): Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In most scaling types, a `factor` of x will enable the model to handle sequences of length x * original maximum pre-trained length. `original_max_position_embeddings` (`int`, *optional*): Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during pretraining. `attention_factor` (`float`, *optional*): Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention computation. If unspecified, it defaults to value recommended by the implementation, using the `factor` field to infer the suggested value. `beta_fast` (`float`, *optional*): Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear ramp function. If unspecified, it defaults to 32. `beta_slow` (`float`, *optional*): Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear ramp function. If unspecified, it defaults to 1. `short_factor` (`list[float]`, *optional*): Only used with 'longrope'. The scaling factor to be applied to short contexts (< `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden size divided by the number of attention heads divided by 2 `long_factor` (`list[float]`, *optional*): Only used with 'longrope'. The scaling factor to be applied to long contexts (< `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden size divided by the number of attention heads divided by 2 `low_freq_factor` (`float`, *optional*): Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE `high_freq_factor` (`float`, *optional*): Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`): Whether to use a bias in the query, key, value and output projection layers during self-attention. use_sliding_window (`bool`, *optional*, defaults to `False`): Whether to use sliding window attention. sliding_window (`int`, *optional*, defaults to 4096): Sliding window attention (SWA) window size. If not specified, will default to `4096`. attention_dropout (`float`, *optional*, defaults to 0.0): The dropout ratio for the attention probabilities. decoder_sparse_step (`int`, *optional*, defaults to 1): The frequency of the MoE layer. moe_intermediate_size (`int`, *optional*, defaults to 768): Intermediate size of the routed expert. num_experts_per_tok (`int`, *optional*, defaults to 8): Number of selected experts. num_experts (`int`, *optional*, defaults to 128): Number of routed experts. norm_topk_prob (`bool`, *optional*, defaults to `False`): Whether to normalize the topk probabilities. output_router_logits (`bool`, *optional*, defaults to `False`): Whether or not the router logits should be returned by the model. Enabling this will also allow the model to output the auxiliary loss, including load balancing loss and router z-loss. router_aux_loss_coef (`float`, *optional*, defaults to 0.001): The aux loss factor for the total loss. mlp_only_layers (`list[int]`, *optional*, defaults to `[]`): Indicate which layers use Qwen3OmniMoeTalkerTextMLP rather than Qwen3OmniMoeTalkerTextSparseMoeBlock The list contains layer index, from 0 to num_layers-1 if we have num_layers layers If `mlp_only_layers` is empty, `decoder_sparse_step` is used to determine the sparsity. ```python >>> from transformers import Qwen3OmniMoeTalkerTextModel, Qwen3OmniMoeTalkerTextConfig >>> # Initializing a Qwen3OmniMoeTalkerText style configuration >>> configuration = Qwen3OmniMoeTalkerTextConfig() >>> # Initializing a model from the Qwen3-15B-A2B" style configuration >>> model = Qwen3OmniMoeTalkerTextModel(configuration) >>> # Accessing the model configuration >>> configuration = model.config ```""" model_type = "qwen3_omni_moe_talker_text" keys_to_ignore_at_inference = ["past_key_values"] # Default tensor parallel plan for base model `Qwen3OmniMoeTalkerText` base_model_tp_plan = { "layers.*.self_attn.q_proj": "colwise", "layers.*.self_attn.k_proj": "colwise", "layers.*.self_attn.v_proj": "colwise", "layers.*.self_attn.o_proj": "rowwise", "layers.*.mlp.experts.*.gate_proj": "colwise", "layers.*.mlp.experts.*.up_proj": "colwise", "layers.*.mlp.experts.*.down_proj": "rowwise", "layers.*.mlp.gate_proj": "colwise", "layers.*.mlp.up_proj": "colwise", "layers.*.mlp.down_proj": "rowwise", } base_model_pp_plan = { "embed_tokens": (["input_ids"], ["inputs_embeds"]), "layers": (["hidden_states", "attention_mask"], ["hidden_states"]), "norm": (["hidden_states"], ["hidden_states"]), } def __init__( self, vocab_size=3072, hidden_size=1024, intermediate_size=2048, num_hidden_layers=20, num_attention_heads=16, num_key_value_heads=2, hidden_act="silu", max_position_embeddings=32768, initializer_range=0.02, rms_norm_eps=0.000001, use_cache=True, tie_word_embeddings=False, rope_theta=10000, rope_scaling=None, attention_bias=False, sliding_window=None, attention_dropout=0, decoder_sparse_step=1, moe_intermediate_size=384, num_experts_per_tok=8, num_experts=128, norm_topk_prob=False, output_router_logits=False, router_aux_loss_coef=0.001, mlp_only_layers=None, **kwargs, ): super().__init__( tie_word_embeddings=tie_word_embeddings, **kwargs, ) self.vocab_size = vocab_size self.max_position_embeddings = max_position_embeddings self.hidden_size = hidden_size self.intermediate_size = intermediate_size self.num_hidden_layers = num_hidden_layers self.num_attention_heads = num_attention_heads self.sliding_window = sliding_window self.num_key_value_heads = num_key_value_heads self.hidden_act = hidden_act self.initializer_range = initializer_range self.rms_norm_eps = rms_norm_eps self.use_cache = use_cache self.rope_theta = rope_theta self.rope_scaling = rope_scaling self.attention_bias = attention_bias self.attention_dropout = attention_dropout # Validate the correctness of rotary position embeddings parameters # BC: if there is a 'type' field, move it to 'rope_type'. if self.rope_scaling is not None and "type" in self.rope_scaling: self.rope_scaling["rope_type"] = self.rope_scaling["type"] rope_config_validation(self) # MoE arguments self.decoder_sparse_step = decoder_sparse_step self.moe_intermediate_size = moe_intermediate_size self.num_experts_per_tok = num_experts_per_tok self.num_experts = num_experts self.norm_topk_prob = norm_topk_prob self.output_router_logits = output_router_logits self.router_aux_loss_coef = router_aux_loss_coef self.mlp_only_layers = [] if mlp_only_layers is None else mlp_only_layers class Qwen3OmniMoeTalkerConfig(PretrainedConfig): r""" This is the configuration class to store the configuration of a [`Qwen3OmniMoeTalker`]. It is used to instantiate a Qwen3-Omni multi-modal talker model capable of handling text, audio, and vision modalities in a unified architecture. The model integrates a text decoder with a code predictor for autoregressive generation of both semantic and acoustic tokens, enabling speech and multimodal content generation. This configuration wraps sub-configurations for the text and code predictor components, allowing modular setup and initialization. e.g. [Qwen/Qwen3-Omni-7B](https://huggingface.co/Qwen/Qwen3-Omni-7B) Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the documentation from [`PretrainedConfig`] for more information. Args: code_predictor_config (`dict`, *optional*): A dictionary of configuration parameters used to initialize a [`Qwen3OmniMoeTalkerCodePredictorConfig`]. If not provided, defaults will be used. text_config (`dict`, *optional*): A dictionary of configuration parameters used to initialize a [`Qwen3OmniMoeTalkerTextConfig`]. If not provided, defaults will be used. num_code_groups (`int`, *optional*, defaults to 32): Number of codebook groups used in the predicted acoustic token sequence, corresponding to multi-codebook VQ representation. thinker_hidden_size (`int`, *optional*, defaults to 2048): Hidden dimension size of the thinker module used for intermediate reasoning or latent planning before audio generation. codec_eos_token_id (`int`, *optional*, defaults to 4198): Token ID representing the end-of-speech token in the codec-generated sequence. accept_hidden_layer (`int`, *optional*, defaults to 18): Index of the hidden layer whose output is used for accepting or refining generated tokens during think-and-speak process. codec_nothink_id (`int`, *optional*, defaults to 4203): Token ID indicating no thinking step is required during generation. codec_think_bos_id (`int`, *optional*, defaults to 4204): Token ID marking the beginning of a thinking sequence. codec_think_eos_id (`int`, *optional*, defaults to 4205): Token ID marking the end of a thinking sequence. codec_pad_id (`int`, *optional*, defaults to 4196): Padding token ID used in codec input sequences. codec_bos_id (`int`, *optional*, defaults to 4197): Beginning-of-speech token ID in codec sequences. audio_token_id (`int`, *optional*, defaults to 151646): Special token ID used to indicate the position of audio tokens in the input sequence. image_token_id (`int`, *optional*, defaults to 151655): Special token ID used to represent image inputs in the multimodal context. video_token_id (`int`, *optional*, defaults to 151656): Special token ID used to represent video inputs. vision_start_token_id (`int`, *optional*, defaults to 151652): Token ID indicating the start of a visual input sequence (e.g., image or video embeddings). position_id_per_seconds (`int`, *optional*, defaults to 25): Number of position IDs allocated per second of audio content, used for temporal alignment in generation. audio_start_token_id (`int`, *optional*, defaults to 151669): Token ID that indicates the start of an audio generation segment in the output. speaker_id (`dict`, *optional*): Speaker name to speaker id dict. Example: ```python >>> from transformers import Qwen3OmniMoeTalkerConfig, Qwen3OmniMoeTalker >>> # Initialize a Qwen3OmniMoeTalkerConfig with default sub-configurations >>> config = Qwen3OmniMoeTalkerConfig( ... num_code_groups=32, ... thinker_hidden_size=2048, ... ) >>> # Initialize the full Qwen3-Omni Talker model >>> model = Qwen3OmniMoeTalker(config) >>> # Access the model configuration >>> config = model.config >>> print(config.text_config) # Access text decoder configuration >>> print(config.code_predictor_config) # Access code predictor configuration ```""" sub_configs = { "code_predictor_config": Qwen3OmniMoeTalkerCodePredictorConfig, "text_config": Qwen3OmniMoeTalkerTextConfig, } def __init__( self, code_predictor_config=None, text_config=None, num_code_groups=32, thinker_hidden_size=2048, codec_eos_token_id=4198, accept_hidden_layer=18, codec_nothink_id=4203, codec_think_bos_id=4204, codec_think_eos_id=4205, codec_pad_id=4196, codec_bos_id=4197, audio_token_id=151646, image_token_id=151655, video_token_id=151656, vision_start_token_id=151652, position_id_per_seconds=25, audio_start_token_id=151669, speaker_id=None, **kwargs, ): super().__init__(**kwargs) if code_predictor_config is None: code_predictor_config = {} self.code_predictor_config = Qwen3OmniMoeTalkerCodePredictorConfig() logger.info("code_predictor_config is None. Initializing code_predictor_config model with default values") elif isinstance(code_predictor_config, Qwen3OmniMoeTalkerCodePredictorConfig): self.code_predictor_config = code_predictor_config else: self.code_predictor_config = Qwen3OmniMoeTalkerCodePredictorConfig(**code_predictor_config) if text_config is None: text_config = {} self.text_config = Qwen3OmniMoeTalkerTextConfig() logger.info("talker text_config is None. Initializing talker text model with default values") elif isinstance(text_config, Qwen3OmniMoeTalkerTextConfig): self.text_config = text_config else: self.text_config = Qwen3OmniMoeTalkerTextConfig(**text_config) self.num_code_groups = num_code_groups self.thinker_hidden_size = thinker_hidden_size self.codec_eos_token_id = codec_eos_token_id self.accept_hidden_layer = accept_hidden_layer self.codec_nothink_id = codec_nothink_id self.codec_think_bos_id = codec_think_bos_id self.codec_think_eos_id = codec_think_eos_id self.codec_pad_id = codec_pad_id self.codec_bos_id = codec_bos_id self.audio_token_id = audio_token_id self.image_token_id = image_token_id self.video_token_id = video_token_id self.position_id_per_seconds = position_id_per_seconds self.audio_start_token_id = audio_start_token_id self.vision_start_token_id = vision_start_token_id self.speaker_id = speaker_id class Qwen3OmniMoeCode2WavConfig(PretrainedConfig): r""" This is the configuration class to store the configuration of a [`Qwen3OmniMoeCode2WavConfig`]. It is used to instantiate a Qwen3-Omni code-to-waveform decoder, responsible for converting discrete audio codes into high-fidelity waveforms. The configuration defines the architecture of the decoder, including parameters for vector quantization, autoregressive modeling, and upsampling layers. e.g. [Qwen/Qwen3-Omni-7B](https://huggingface.co/Qwen/Qwen3-Omni-7B) Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the documentation from [`PretrainedConfig`] for more information. Args: codebook_size (`int`, *optional*, defaults to 2048): Number of entries in each residual codebook used for acoustic token quantization. hidden_size (`int`, *optional*, defaults to 1024): Dimensionality of the hidden states and embeddings in the autoregressive transformer decoder. max_position_embeddings (`int`, *optional*, defaults to 8000): Maximum sequence length that the autoregressive decoder can handle. Determines positional embedding size. rope_theta (`float`, *optional*, defaults to 10000.0): The base period for rotary position embeddings (RoPE) applied to attention layers. num_attention_heads (`int`, *optional*, defaults to 16): Number of attention heads for each attention layer in the decoder. num_key_value_heads (`int`, *optional*, defaults to 16): Number of key and value attention heads used in grouped-query attention (if applicable). attention_bias (`bool`, *optional*, defaults to `False`): Whether to use bias in the attention projection layers. sliding_window (`int`, *optional*, defaults to 72): Window size for local attention mechanism, limiting attention context to improve efficiency. intermediate_size (`int`, *optional*, defaults to 3072): Dimensionality of the feed-forward (intermediate) layer in each transformer block. hidden_act (`str` or `function`, *optional*, defaults to `"silu"`): The non-linear activation function used in the feed-forward layers. Supports `"silu"`, `"relu"`, `"gelu"`, etc. layer_scale_initial_scale (`float`, *optional*, defaults to 0.01): Initial value for LayerScale applied in transformer blocks, helping stabilize training. rms_norm_eps (`float`, *optional*, defaults to 1e-5): Epsilon value for RMS normalization layers to prevent division by zero. num_hidden_layers (`int`, *optional*, defaults to 8): Number of transformer blocks in the autoregressive decoder. num_quantizers (`int`, *optional*, defaults to 16): Number of residual vector quantizers used in the vocoder for fine-grained audio reconstruction. upsample_rates (`Tuple[int]`, *optional*, defaults to `(8, 5, 4, 3)`): Rate at which features are upsampled in the final waveform synthesis stage. upsampling_ratios (`Tuple[int]`, *optional*, defaults to `(2, 2)`): Ratios used in transposed convolutional layers to progressively upsample feature maps to waveform. decoder_dim (`int`, *optional*, defaults to 1536): Final dimensionality of the decoder's output before waveform generation. attention_dropout (`float`, *optional*, defaults to 0.0): Dropout probability applied to attention weights in the decoder. Example: ```python >>> from transformers import Qwen3OmniMoeCode2WavConfig, Qwen3OmniMoeCode2WavModel >>> # Initializing a default Qwen3OmniMoeCode2WavConfig >>> config = Qwen3OmniMoeCode2WavConfig() >>> # Initializing the Code2Wav model with the configuration >>> model = Qwen3OmniMoeCode2WavModel(config) >>> # Accessing configuration >>> config = model.config ```""" def __init__( self, codebook_size=2048, hidden_size=1024, max_position_embeddings=8000, rope_theta=10000, num_attention_heads=16, num_key_value_heads=16, attention_bias=False, sliding_window=72, intermediate_size=3072, hidden_act="silu", layer_scale_initial_scale=0.01, rms_norm_eps=1e-5, num_hidden_layers=8, num_quantizers=16, upsample_rates=(8, 5, 4, 3), upsampling_ratios=(2, 2), decoder_dim=1536, attention_dropout=0.0, **kwargs, ): super().__init__(**kwargs) self.codebook_size = codebook_size self.hidden_size = hidden_size self.max_position_embeddings = max_position_embeddings self.rope_theta = rope_theta self.num_attention_heads = num_attention_heads self.num_key_value_heads = num_key_value_heads self.attention_bias = attention_bias self.sliding_window = sliding_window self.intermediate_size = intermediate_size self.hidden_act = hidden_act self.layer_scale_initial_scale = layer_scale_initial_scale self.rms_norm_eps = rms_norm_eps self.num_hidden_layers = num_hidden_layers self.num_quantizers = num_quantizers self.upsample_rates = upsample_rates self.upsampling_ratios = upsampling_ratios self.decoder_dim = decoder_dim self.attention_dropout = attention_dropout @property def layer_types(self): """ All layer in code2wav should be sliding attention """ return ["sliding_attention"] * self.num_hidden_layers class Qwen3OmniMoeConfig(PretrainedConfig): """ This is the configuration class to store the configuration of a [`Qwen3OmniMoeForConditionalGeneration`]. It is used to instantiate a Qwen3Omni model according to the specified sub-models configurations, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of the [Qwen/Qwen2.5-Omni-7B](https://huggingface.co/Qwen/Qwen2.5-Omni-7B) architecture. Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the documentation from [`PretrainedConfig`] for more information. Args: thinker_config (`dict`, *optional*): Configuration of the underlying thinker sub-model. talker_config (`dict`, *optional*): Configuration of the underlying talker sub-model. code2wav_config (`dict`, *optional*): Configuration of the underlying code2wav sub-model. enable_audio_output (`bool`, *optional*, defaults to `True`): Whether enable audio output and load talker and code2wav module. Example: ```python >>> from transformers import ( ... Qwen3OmniMoeThinkerConfig, ... Qwen3OmniMoeTalkerConfig, ... Qwen3OmniMoeCode2WavConfig, ... Qwen3OmniMoeForConditionalGeneration, ... Qwen3OmniMoeConfig, ... ) >>> # Initializing a Qwen3OmniMoe style configuration >>> configuration = Qwen3OmniMoeConfig() >>> # Initializing a model from the configuration >>> model = Qwen3OmniMoeForConditionalGeneration(configuration) >>> # Accessing the model configuration >>> configuration = model.config ```""" model_type = "qwen3_omni_moe" sub_configs = { "thinker_config": Qwen3OmniMoeThinkerConfig, "talker_config": Qwen3OmniMoeTalkerConfig, "code2wav_config": Qwen3OmniMoeCode2WavConfig, } def __init__( self, thinker_config=None, talker_config=None, code2wav_config=None, enable_audio_output=True, im_start_token_id=151644, im_end_token_id=151645, tts_pad_token_id=151671, tts_bos_token_id=151672, tts_eos_token_id=151673, system_token_id=8948, user_token_id=872, assistant_token_id=77091, **kwargs, ): super().__init__(**kwargs) if thinker_config is None: thinker_config = {} logger.info("thinker_config is None. Initializing thinker model with default values") if talker_config is None: talker_config = {} logger.info("talker_config is None. Initializing talker model with default values") if code2wav_config is None: code2wav_config = {} logger.info("code2wav_config is None. Initializing code2wav model with default values") self.thinker_config = Qwen3OmniMoeThinkerConfig(**thinker_config) self.talker_config = Qwen3OmniMoeTalkerConfig(**talker_config) self.code2wav_config = Qwen3OmniMoeCode2WavConfig(**code2wav_config) self.enable_audio_output = enable_audio_output self.im_start_token_id = im_start_token_id self.im_end_token_id = im_end_token_id self.tts_pad_token_id = tts_pad_token_id self.tts_bos_token_id = tts_bos_token_id self.tts_eos_token_id = tts_eos_token_id self.system_token_id = system_token_id self.user_token_id = user_token_id self.assistant_token_id = assistant_token_id def get_text_config(self, decoder=False) -> "PretrainedConfig": """ Returns the config that is meant to be used with text IO. On most models, it is the original config instance itself. On specific composite models, it is under a set of valid names. Args: decoder (`Optional[bool]`, *optional*, defaults to `False`): If set to `True`, then only search for decoder config names. """ # Overridden for deeply nested config like Qwen2-Omni. We don't have any omni model # except for Qwen yet. This has to be generalized if more deeply nested configs are # added. NOTE: currently method used only by vLLM return self.thinker_config.get_text_config() __all__ = ["Qwen3OmniMoeConfig", "Qwen3OmniMoeThinkerConfig", "Qwen3OmniMoeTalkerConfig"]