vllm.transformers_utils.configs.qwen3_asr ¶

class Qwen3ASRAudioEncoderConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`Qwen3ASRAudioEncoder`]. It is used to instantiate a
    Qwen3-ASR 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/Qwen3-ASR-1.7B](https://huggingface.co/Qwen/Qwen3-ASR-1.7B)

    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
            `Qwen3ASRProcessor` 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 Qwen3ASRAudioEncoderConfig, Qwen3ASRAudioEncoder

    >>> # Initializing a Qwen3ASRAudioEncoderConfig
    >>> configuration = Qwen3ASRAudioEncoderConfig()

    >>> # Initializing a Qwen3ASRAudioEncoder (with random weights)
    >>> model = Qwen3ASRAudioEncoder(configuration)

    >>> # Accessing the model configuration
    >>> configuration = model.config
    ```"""

    model_type = "qwen3_asr_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 Qwen3ASRConfig(PretrainedConfig):
    """
    This is the configuration class to store the configuration of a [`Qwen3ASRForConditionalGeneration`]. It is used to instantiate a Qwen3ASR
    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/Qwen3-ASR-1.7B](https://huggingface.co/Qwen/Qwen3-ASR-1.7B) 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.
        support_languages (`List[str]`, *optional*): The languages supported by the model.

    Example:

    ```python
    >>> from transformers import (
    ...     Qwen3ASRThinkerConfig,
    ...     Qwen3ASRForConditionalGeneration,
    ...     Qwen3ASRConfig,
    ... )

    >>> # Initializing a Qwen3ASR style configuration
    >>> configuration = Qwen3ASRConfig()

    >>> # Initializing a model from the configuration
    >>> model = Qwen3ASRForConditionalGeneration(configuration)

    >>> # Accessing the model configuration
    >>> configuration = model.config
    ```"""

    model_type = "qwen3_asr"
    sub_configs = {
        "thinker_config": Qwen3ASRThinkerConfig,
    }

    def __init__(
        self,
        thinker_config=None,
        support_languages=None,
        **kwargs,
    ):
        super().__init__(**kwargs)
        if thinker_config is None:
            thinker_config = {}
            logger.info(
                "thinker_config is None. Initializing thinker model with default values"
            )

        self.thinker_config = Qwen3ASRThinkerConfig(**thinker_config)
        self.support_languages = support_languages

    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 Qwen2.5-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()

class Qwen3ASRTextConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`Qwen3ASRTextModel`]. It is used to instantiate a
    Qwen3-ASR model according to the specified arguments, defining the model architecture. Instantiating a configuration
    with the defaults will yield a similar configuration to that of
    Qwen3-ASR-1.7B [Qwen/Qwen3-ASR-1.7B](https://huggingface.co/Qwen/Qwen3-ASR-1.7B)

    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 Qwen3ASR model. Defines the number of different tokens that can be represented by the
            `inputs_ids` passed when calling [`Qwen3ASRModel`]
        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 dimension of the head. If not specified, will default to `hidden_size // num_attention_heads`.
        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 128000):
            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 5000000.0):
            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.
        attention_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for the attention probabilities.

    ```python
    >>> from transformers import Qwen3ASRTextModel, Qwen3ASRTextConfig

    >>> # Initializing a Qwen3ASR style configuration
    >>> configuration = Qwen3ASRTextConfig()

    >>> # Initializing a model from the Qwen3-VL-7B style configuration
    >>> model = Qwen3ASRTextModel(configuration)

    >>> # Accessing the model configuration
    >>> configuration = model.config
    ```"""

    model_type = "qwen3_asr_text"
    base_config_key = "text_config"

    def __init__(
        self,
        vocab_size=151936,
        hidden_size=4096,
        intermediate_size=22016,
        num_hidden_layers=32,
        num_attention_heads=32,
        num_key_value_heads=32,
        head_dim=128,
        hidden_act="silu",
        max_position_embeddings=128000,
        initializer_range=0.02,
        rms_norm_eps=1e-6,
        use_cache=True,
        tie_word_embeddings=False,
        rope_theta=5000000.0,
        rope_scaling=None,
        attention_bias=False,
        attention_dropout=0.0,
        **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

        # 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)

        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)

class Qwen3ASRThinkerConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`Qwen3ASRThinker`]. It is used to instantiate a
    Qwen3-ASR-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-ASR-1.7B](https://huggingface.co/Qwen/Qwen3-ASR-1.7B)

    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.
        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.
        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 Qwen3ASRThinkerModel, Qwen3ASRThinkerConfig

    >>> # Initializing a default Qwen3ASRThinkerConfig
    >>> configuration = Qwen3ASRThinkerConfig()

    >>> # Initializing a model (with random weights) from the default configuration
    >>> model = Qwen3ASRThinkerModel(configuration)

    >>> # Accessing the model configuration
    >>> configuration = model.config
    ```"""

    model_type = "qwen3_asr_thinker"

    attribute_map = {}
    sub_configs = {
        "audio_config": Qwen3ASRAudioEncoderConfig,
        "text_config": Qwen3ASRTextConfig,
    }

    def __init__(
        self,
        audio_config=None,
        text_config=None,
        audio_token_id=151646,
        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.audio_start_token_id = audio_start_token_id
        self.initializer_range = initializer_range

        if isinstance(audio_config, dict):
            audio_config = Qwen3ASRAudioEncoderConfig(**audio_config)
        elif audio_config is None:
            audio_config = Qwen3ASRAudioEncoderConfig()
        self.audio_config = audio_config

        if isinstance(text_config, dict):
            text_config = Qwen3ASRTextConfig(**text_config)
        elif text_config is None:
            text_config = Qwen3ASRTextConfig()
        self.text_config = text_config
        self.audio_token_id = audio_token_id