configuration_meralion.py

"""MERaLiON AudioLLM model configuration"""

from collections import OrderedDict
from typing import TYPE_CHECKING, Any, Mapping, Optional, Union

from transformers.configuration_utils import PretrainedConfig
from transformers.onnx import OnnxConfig
from transformers.utils import logging


if TYPE_CHECKING:
    from transformers.feature_extraction_utils import FeatureExtractionMixin
    from transformers.tokenization_utils_base import PreTrainedTokenizerBase
    from transformers.utils import TensorType


logger = logging.get_logger(__name__)


# fmt: off
NON_SPEECH_TOKENS = [
    1, 2, 7, 8, 9, 10, 14, 25,
    26, 27, 28, 29, 31, 58, 59, 60, 61, 62,
    63, 90, 91, 92, 93, 357, 366, 438, 532, 685,
    705, 796, 930, 1058, 1220, 1267, 1279, 1303, 1343, 1377,
    1391, 1635, 1782, 1875, 2162, 2361, 2488, 3467, 4008, 4211,
    4600, 4808, 5299, 5855, 6329, 7203, 9609, 9959, 10563, 10786,
    11420, 11709, 11907, 13163, 13697, 13700, 14808, 15306, 16410, 16791,
    17992, 19203, 19510, 20724, 22305, 22935, 27007, 30109, 30420, 33409,
    34949, 40283, 40493, 40549, 47282, 49146, 50257, 50359, 50360, 50361
]
NON_SPEECH_TOKENS_MULTI = [
    1, 2, 7, 8, 9, 10, 14, 25,
    26, 27, 28, 29, 31, 58, 59, 60, 61, 62,
    63, 90, 91, 92, 93, 359, 503, 522, 542, 873,
    893, 902, 918, 922, 931, 1350, 1853, 1982, 2460, 2627,
    3246, 3253, 3268, 3536, 3846, 3961, 4183, 4667, 6585, 6647,
    7273, 9061, 9383, 10428, 10929, 11938, 12033, 12331, 12562, 13793,
    14157, 14635, 15265, 15618, 16553, 16604, 18362, 18956, 20075, 21675,
    22520, 26130, 26161, 26435, 28279, 29464, 31650, 32302, 32470, 36865,
    42863, 47425, 49870, 50254, 50258, 50360, 50361, 50362
]
# fmt: on

# Copied from transformers.models.whisper.configuration_whisper.WhisperConfig
class MERaLiONSpeechConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`MERaLiONSpeechModel`]. It is used to instantiate a
    MERaLiONSpeech 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 MERaLiONSpeech
    [openai/whisper-tiny](https://huggingface.co/openai/whisper-tiny) architecture.

    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 51865):
            Vocabulary size of the MERaLiONSpeech model. Defines the number of different tokens that can be represented by the
            `decoder_input_ids` passed when calling [`MERaLiONSpeechModel`]
        num_mel_bins (`int`, *optional*, defaults to 80):
            Number of mel features used per input features. Should correspond to the value used in the
            `MERaLiONSpeechProcessor` class.
        encoder_layers (`int`, *optional*, defaults to 4):
            Number of encoder layers.
        decoder_layers (`int`, *optional*, defaults to 4):
            Number of decoder layers.
        encoder_attention_heads (`int`, *optional*, defaults to 6):
            Number of attention heads for each attention layer in the Transformer encoder.
        decoder_attention_heads (`int`, *optional*, defaults to 6):
            Number of attention heads for each attention layer in the Transformer decoder.
        encoder_ffn_dim (`int`, *optional*, defaults to 1536):
            Dimensionality of the "intermediate" (often named feed-forward) layer in encoder.
        decoder_ffn_dim (`int`, *optional*, defaults to 1536):
            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
        encoder_layerdrop (`float`, *optional*, defaults to 0.0):
            The LayerDrop probability for the encoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
            for more details.
        decoder_layerdrop (`float`, *optional*, defaults to 0.0):
            The LayerDrop probability for the decoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
            for more details.
        decoder_start_token_id (`int`, *optional*, defaults to 50257):
            Corresponds to the "<|startoftranscript|>" token, which is automatically used when no `decoder_input_ids`
            are provided to the `generate` function. It is used to guide the model`s generation process depending on
            the task.
        use_cache (`bool`, *optional*, defaults to `True`):
            Whether or not the model should return the last key/values attentions (not used by all models).
        is_encoder_decoder (`bool`, *optional*, defaults to `True`):
            Whether the model is used as an encoder/decoder or not.
        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.
        d_model (`int`, *optional*, defaults to 384):
            Dimensionality of the layers.
        dropout (`float`, *optional*, defaults to 0.1):
            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_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for activations inside the fully connected layer.
        init_std (`float`, *optional*, defaults to 0.02):
            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
        scale_embedding (`bool`, *optional*, defaults to False):
            Scale embeddings by diving by sqrt(d_model).
        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.
        max_target_positions (`int`, *optional*, defaults to 448):
            The maximum sequence length that this model might ever be used with. Typically set this to something large
            just in case (e.g., 512 or 1024 or 2048).
        pad_token_id (`int`, *optional*, defaults to 50256):
            Padding token id.
        bos_token_id (`int`, *optional*, defaults to 50256):
            Begin of stream token id.
        eos_token_id (`int`, *optional*, defaults to 50256):
            End of stream token id.
        suppress_tokens (`List[int]`, *optional*):
            A list containing the non-speech tokens that will be used by the logit processor in the `generate`
            function. NON_SPEECH_TOKENS and NON_SPEECH_TOKENS_MULTI each correspond to the `english-only` and the
            `multilingual` model.
        begin_suppress_tokens (`List[int]`, *optional*, defaults to `[220,50256]`):
            A list containing tokens that will be supressed at the beginning of the sampling process. Initialized as
            the token for `" "` (`blank_token_id`) and the `eos_token_id`
        use_weighted_layer_sum (`bool`, *optional*, defaults to `False`):
            Whether to use a weighted average of layer outputs with learned weights. Only relevant when using an
            instance of [`MERaLiONSpeechForAudioClassification`].
        classifier_proj_size (`int`, *optional*, defaults to 256):
            Dimensionality of the projection before token mean-pooling for classification. Only relevant when using an
            instance of [`MERaLiONSpeechForAudioClassification`].
        apply_spec_augment (`bool`, *optional*, defaults to `False`):
            Whether to apply *SpecAugment* data augmentation to the outputs of the feature encoder. For reference see
            [SpecAugment: A Simple Data Augmentation Method for Automatic Speech
            Recognition](https://arxiv.org/abs/1904.08779).
        mask_time_prob (`float`, *optional*, defaults to 0.05):
            Percentage (between 0 and 1) of all feature vectors along the time axis which will be masked. The masking
            procecure generates `mask_time_prob*len(time_axis)/mask_time_length` independent masks over the axis. If
            reasoning from the propability of each feature vector to be chosen as the start of the vector span to be
            masked, *mask_time_prob* should be `prob_vector_start*mask_time_length`. Note that overlap may decrease the
            actual percentage of masked vectors. This is only relevant if `apply_spec_augment == True`.
        mask_time_length (`int`, *optional*, defaults to 10):
            Length of vector span along the time axis.
        mask_time_min_masks (`int`, *optional*, defaults to 2),:
            The minimum number of masks of length `mask_feature_length` generated along the time axis, each time step,
            irrespectively of `mask_feature_prob`. Only relevant if ''mask_time_prob*len(time_axis)/mask_time_length <
            mask_time_min_masks''
        mask_feature_prob (`float`, *optional*, defaults to 0.0):
            Percentage (between 0 and 1) of all feature vectors along the feature axis which will be masked. The
            masking procecure generates `mask_feature_prob*len(feature_axis)/mask_time_length` independent masks over
            the axis. If reasoning from the propability of each feature vector to be chosen as the start of the vector
            span to be masked, *mask_feature_prob* should be `prob_vector_start*mask_feature_length`. Note that overlap
            may decrease the actual percentage of masked vectors. This is only relevant if `apply_spec_augment is
            True`.
        mask_feature_length (`int`, *optional*, defaults to 10):
            Length of vector span along the feature axis.
        mask_feature_min_masks (`int`, *optional*, defaults to 0),:
            The minimum number of masks of length `mask_feature_length` generated along the feature axis, each time
            step, irrespectively of `mask_feature_prob`. Only relevant if
            `mask_feature_prob*len(feature_axis)/mask_feature_length < mask_feature_min_masks`.
        median_filter_width (`int`, *optional*, defaults to 7):
            Width of the median filter used to smoothen to cross-attention outputs when computing token timestamps.
            Should be an odd number.
    """

    model_type = "meralion_speech_encoder"
    keys_to_ignore_at_inference = ["past_key_values"]
    attribute_map = {
        "num_key_value_heads": "encoder_attention_heads",
        "num_attention_heads": "encoder_attention_heads",
        "hidden_size": "d_model",
    }

    def __init__(
        self,
        vocab_size=51865,
        num_mel_bins=80,
        encoder_layers=4,
        encoder_attention_heads=6,
        decoder_layers=4,
        decoder_attention_heads=6,
        decoder_ffn_dim=1536,
        encoder_ffn_dim=1536,
        encoder_layerdrop=0.0,
        decoder_layerdrop=0.0,
        decoder_start_token_id=50257,
        use_cache=True,
        is_encoder_decoder=True,
        activation_function="gelu",
        d_model=384,
        dropout=0.0,
        attention_dropout=0.0,
        activation_dropout=0.0,
        init_std=0.02,
        scale_embedding=False,
        max_source_positions=1500,
        max_target_positions=448,
        pad_token_id=50256,
        bos_token_id=50256,
        eos_token_id=50256,
        suppress_tokens=None,
        begin_suppress_tokens=[220, 50256],
        use_weighted_layer_sum=False,
        classifier_proj_size=256,
        apply_spec_augment=False,
        mask_time_prob=0.05,
        mask_time_length=10,
        mask_time_min_masks=2,
        mask_feature_prob=0.0,
        mask_feature_length=10,
        mask_feature_min_masks=0,
        median_filter_width=7,
        **kwargs,
    ):
        self.vocab_size = vocab_size
        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.decoder_layers = decoder_layers
        self.decoder_attention_heads = decoder_attention_heads
        self.decoder_ffn_dim = decoder_ffn_dim
        self.encoder_ffn_dim = encoder_ffn_dim
        self.dropout = dropout
        self.attention_dropout = attention_dropout
        self.activation_dropout = activation_dropout
        self.activation_function = activation_function
        self.init_std = init_std
        self.encoder_layerdrop = encoder_layerdrop
        self.decoder_layerdrop = decoder_layerdrop
        self.use_cache = use_cache
        self.num_hidden_layers = encoder_layers
        self.scale_embedding = scale_embedding  # scale factor will be sqrt(d_model) if True
        self.max_source_positions = max_source_positions
        self.max_target_positions = max_target_positions

        # Audio Classification-specific parameters. Feel free to ignore for other classes.
        self.classifier_proj_size = classifier_proj_size
        self.use_weighted_layer_sum = use_weighted_layer_sum

        # fine-tuning config parameters for SpecAugment: https://arxiv.org/abs/1904.08779
        self.apply_spec_augment = apply_spec_augment
        self.mask_time_prob = mask_time_prob
        self.mask_time_length = mask_time_length
        self.mask_time_min_masks = mask_time_min_masks
        self.mask_feature_prob = mask_feature_prob
        self.mask_feature_length = mask_feature_length
        self.mask_feature_min_masks = mask_feature_min_masks

        self.median_filter_width = median_filter_width

        super().__init__(
            pad_token_id=pad_token_id,
            bos_token_id=bos_token_id,
            eos_token_id=eos_token_id,
            is_encoder_decoder=is_encoder_decoder,
            decoder_start_token_id=decoder_start_token_id,
            suppress_tokens=suppress_tokens,
            begin_suppress_tokens=begin_suppress_tokens,
            **kwargs,
        )
    @property
    def inputs(self) -> Mapping[str, Mapping[int, str]]:
        common_inputs = OrderedDict(
            [
                ("input_features", {0: "batch", 1: "feature_size", 2: "encoder_sequence"}),
            ]
        )
        if self.use_past:
            common_inputs["decoder_input_ids"] = {0: "batch"}
        else:
            common_inputs["decoder_input_ids"] = {0: "batch", 1: "decoder_sequence"}

        if self.use_past:
            self.fill_with_past_key_values_(common_inputs, direction="inputs")

        return common_inputs

    def generate_dummy_inputs(
        self,
        preprocessor: Union["PreTrainedTokenizerBase", "FeatureExtractionMixin"],
        batch_size: int = -1,
        seq_length: int = -1,
        is_pair: bool = False,
        framework: Optional["TensorType"] = None,
        sampling_rate: int = 22050,
        time_duration: float = 5.0,
        frequency: int = 220,
    ) -> Mapping[str, Any]:
        dummy_inputs = OrderedDict()
        encoder_inputs = OnnxConfig.generate_dummy_inputs(
            self,
            preprocessor=preprocessor.feature_extractor,
            batch_size=batch_size,
            framework=framework,
            sampling_rate=sampling_rate,
            time_duration=time_duration,
            frequency=frequency,
        )
        encoder_sequence_length = encoder_inputs["input_features"].shape[2]
        seq_length = encoder_sequence_length // 2 if self.use_past else seq_length

        decoder_inputs = super().generate_dummy_inputs(
            preprocessor.tokenizer, batch_size, seq_length, is_pair, framework
        )

        dummy_inputs["input_features"] = encoder_inputs.pop("input_features")
        dummy_inputs["decoder_input_ids"] = decoder_inputs.pop("decoder_input_ids")

        if "past_key_values" in decoder_inputs:
            dummy_inputs["past_key_values"] = decoder_inputs.pop("past_key_values")

        return dummy_inputs

    @property
    def atol_for_validation(self) -> float:
        return 1e-3
    

# Copied from transformers.models.gemma2.configuration_gemma2.Gemma2Config
class MERaLiONTextConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`MERaLiONTextModel`]. It is used to instantiate an MERaLiONText
    model according to the specified arguments, defining the model architecture. 
    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 256000):
            Vocabulary size of the MERaLiONText model. Defines the number of different tokens that can be represented by the
            `inputs_ids` passed when calling [`MERaLiONTextModel`]
        hidden_size (`int`, *optional*, defaults to 3072):
            Dimension of the hidden representations.
        intermediate_size (`int`, *optional*, defaults to 24576):
            Dimension of the MLP representations.
        num_hidden_layers (`int`, *optional*, defaults to 28):
            Number of hidden layers in the Transformer decoder.
        num_attention_heads (`int`, *optional*, defaults to 16):
            Number of attention heads for each attention layer in the Transformer decoder.
        num_key_value_heads (`int`, *optional*, defaults to 16):
            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 checkout [this
            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
            `num_attention_heads`.
        head_dim (`int`, *optional*, defaults to 256):
            The attention head dimension.
        hidden_activation (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`):
            The non-linear activation function (function or string) in the decoder. Will default to `"gelu_pytorch_tanh"`
            if not specified. `"gelu_pytorch_tanh"` uses an approximation of the `"gelu"` activation function.
        max_position_embeddings (`int`, *optional*, defaults to 8192):
            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`.
        pad_token_id (`int`, *optional*, defaults to 0):
            Padding token id.
        eos_token_id (`int`, *optional*, defaults to 1):
            End of stream token id.
        bos_token_id (`int`, *optional*, defaults to 2):
            Beginning of stream token id.
        tie_word_embeddings (`bool`, *optional*, defaults to `True`):
            Whether to tie weight embeddings
        rope_theta (`float`, *optional*, defaults to 10000.0):
            The base period of the RoPE embeddings.
        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.
        query_pre_attn_scalar (`float`, *optional*, defaults to 224): scaling factor used on the attention scores
        sliding_window (`int`, *optional*, defaults to 4096): in MERaLiONText, every other layer uses sliding window attention. This is the
            size of the sliding window.
        final_logit_softcapping (`float`, *optional*, defaults to 30.0): scaling factor when applying tanh softcapping on the logits.
        attn_logit_softcapping (`float`, *optional*, defaults to 50.0): scaling factor when applying tanh softcapping on the attention scores.
        cache_implementation (`str`, *optional*, defaults to `"hybrid"`): the cache type to be used with `generate`.
    """

    model_type = "meralion_text_decoder"
    keys_to_ignore_at_inference = ["past_key_values"]

    def __init__(
        self,
        vocab_size=256000,
        hidden_size=3072,
        intermediate_size=24576,
        num_hidden_layers=28,
        num_attention_heads=16,
        num_key_value_heads=16,
        head_dim=256,
        hidden_activation="gelu_pytorch_tanh",
        max_position_embeddings=8192,
        initializer_range=0.02,
        rms_norm_eps=1e-6,
        use_cache=True,
        pad_token_id=0,
        eos_token_id=1,
        bos_token_id=2,
        tie_word_embeddings=True,
        rope_theta=10000.0,
        attention_bias=False,
        attention_dropout=0.0,
        query_pre_attn_scalar=224,
        sliding_window=4096,
        final_logit_softcapping=30.0,
        attn_logit_softcapping=50.0,
        cache_implementation="hybrid",
        **kwargs,
    ):
        super().__init__(
            pad_token_id=pad_token_id,
            bos_token_id=bos_token_id,
            eos_token_id=eos_token_id,
            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.head_dim = head_dim
        self.num_key_value_heads = num_key_value_heads
        self.initializer_range = initializer_range
        self.rms_norm_eps = rms_norm_eps
        self.use_cache = use_cache
        self.rope_theta = rope_theta
        self.attention_bias = attention_bias
        self.attention_dropout = attention_dropout
        self.hidden_activation = hidden_activation
        self.query_pre_attn_scalar = query_pre_attn_scalar
        self.sliding_window = sliding_window
        self.final_logit_softcapping = final_logit_softcapping
        self.attn_logit_softcapping = attn_logit_softcapping
        self.cache_implementation = cache_implementation


class MERaLiONConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`MERaLiONForConditionalGeneration`]. It is used to instantiate an
    MERaLiON 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 MERaLiON.

    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 (`Union[AutoConfig, dict]`,  *optional*, defaults to `CLIPVisionConfig`):
            The config object or dictionary of the audio backbone.
        text_config (`Union[AutoConfig, dict]`, *optional*, defaults to `LlamaConfig`):
            The config object or dictionary of the text backbone.
        audio_token_index (`int`, *optional*, defaults to 151646):
            The image token index to encode the image prompt.
    """

    model_type = "meralion"
    is_composition = False

    def __init__(
        self,
        speech_config=None,
        text_config=None,
        speech_mlp_scale_factor=15,
        speech_token_index=255999,
        **kwargs,
    ):
        
        if isinstance(speech_config, dict):
            speech_config = MERaLiONSpeechConfig(**speech_config)
        elif speech_config is None:
            speech_config = MERaLiONSpeechConfig(
                d_model=1280,
                encoder_attention_heads=20,
                encoder_ffn_dim=5120,
                encoder_layerdrop=0.0,
                encoder_layers=32,
                num_mel_bins=128,
                max_source_positions=1500,
                scale_embedding=False,
                activation_function="gelu",
            )

        self.speech_config = speech_config

        if isinstance(text_config, dict):
            text_config = MERaLiONTextConfig(**text_config)
        elif text_config is None:
            text_config = MERaLiONTextConfig()

        self.text_config = text_config

        self.speech_mlp_scale_factor = speech_mlp_scale_factor
        self.speech_token_index = speech_token_index
        
        self.sliding_window = self.text_config.sliding_window
        self.hidden_size = self.text_config.hidden_size
        self.num_attention_heads = self.text_config.num_attention_heads
        self.num_hidden_layers = self.text_config.num_hidden_layers
        self.num_key_value_heads = self.text_config.num_key_value_heads
        self.head_dim = self.text_config.head_dim
        self.intermediate_size = self.text_config.intermediate_size
        
        super().__init__(**kwargs)