image-match/python/py/Lib/site-packages/transformers/models/audioflamingo3/modeling_audioflamingo3.py

#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/audioflamingo3/modular_audioflamingo3.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_audioflamingo3.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2025 NVIDIA CORPORATION 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.

import math
from collections.abc import Callable

import torch
from torch import nn

from ...activations import ACT2FN
from ...cache_utils import Cache, EncoderDecoderCache
from ...generation import GenerationMixin
from ...masking_utils import create_bidirectional_mask
from ...modeling_flash_attention_utils import FlashAttentionKwargs
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import BaseModelOutputWithPooling, CausalLMOutputWithPast
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from ...processing_utils import Unpack
from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, logging
from ...utils.generic import merge_with_config_defaults
from ...utils.output_capturing import capture_outputs
from ..auto import AutoModel, AutoModelForCausalLM
from .configuration_audioflamingo3 import AudioFlamingo3Config, AudioFlamingo3EncoderConfig


logger = logging.get_logger(__name__)


def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: torch.Tensor | None,
    scaling: float | None = None,
    dropout: float = 0.0,
    **kwargs,
):
    if scaling is None:
        scaling = query.size(-1) ** -0.5

    attn_weights = torch.matmul(query, key.transpose(2, 3)) * scaling
    if attention_mask is not None:
        attn_weights = attn_weights + attention_mask

    attn_weights = nn.functional.softmax(attn_weights, dim=-1)

    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
    attn_output = torch.matmul(attn_weights, value)
    attn_output = attn_output.transpose(1, 2).contiguous()

    return attn_output, attn_weights


class AudioFlamingo3Attention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(
        self,
        embed_dim: int,
        num_heads: int,
        dropout: float = 0.0,
        is_decoder: bool = False,
        bias: bool = True,
        is_causal: bool = False,
        layer_idx: int | None = None,
        config: AudioFlamingo3Config | None = None,
    ):
        super().__init__()
        self.embed_dim = embed_dim
        self.num_heads = num_heads
        self.dropout = dropout
        self.head_dim = embed_dim // num_heads
        self.config = config

        if (self.head_dim * num_heads) != self.embed_dim:
            raise ValueError(
                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
                f" and `num_heads`: {num_heads})."
            )
        self.scaling = self.head_dim**-0.5
        self.is_decoder = is_decoder
        self.is_causal = is_causal

        if layer_idx is None and is_decoder:
            logger.warning_once(
                f"Instantiating a decoder {self.__class__.__name__} without passing `layer_idx` is not recommended and "
                "will to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
                "when creating this class."
            )
        self.layer_idx = layer_idx

        self.k_proj = nn.Linear(embed_dim, embed_dim, bias=False)
        self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)

    def forward(
        self,
        hidden_states: torch.Tensor,
        key_value_states: torch.Tensor | None = None,
        past_key_values: Cache | None = None,
        attention_mask: torch.Tensor | None = None,
        output_attentions: bool = False,
        # TODO: we need a refactor so that the different attention modules can get their specific kwargs
        # ATM, we have mixed things encoder, decoder, and encoder-decoder attn
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]:
        """Input shape: Batch x Time x Channel"""

        # if key_value_states are provided this layer is used as a cross-attention layer
        # for the decoder
        is_cross_attention = key_value_states is not None

        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)

        # Scaling is susceptible to floating point arithmetics' inprecisions
        # which can lead to different results (this is dependent from model
        # to model, e.g. audioflamingo3 is one such case). We therefore keep the
        # original order of scaling to follow the original implementation
        # and enforce no scaling (1.0) in the attention call below.
        query_states = (self.q_proj(hidden_states) * self.scaling).view(hidden_shape).transpose(1, 2).contiguous()

        # Check is encoder-decoder model is being used. Otherwise we'll get `DynamicCache`
        if past_key_values is not None and isinstance(past_key_values, EncoderDecoderCache):
            is_updated = past_key_values.is_updated.get(self.layer_idx)
            if is_cross_attention:
                # after the first generated id, we can subsequently re-use all key/value_states from cache
                past_key_values.is_updated[self.layer_idx] = True
                past_key_values = past_key_values.cross_attention_cache
            else:
                past_key_values = past_key_values.self_attention_cache

        # use key_value_states if cross attention
        current_states = key_value_states if key_value_states is not None else hidden_states
        if is_cross_attention and past_key_values and is_updated:
            # reuse k,v, cross_attentions
            key_states = past_key_values.layers[self.layer_idx].keys
            value_states = past_key_values.layers[self.layer_idx].values
        else:
            # Use the query's batch dimension for kv view so that a different-batch
            # encoder output (e.g. in tests) gets absorbed into the sequence axis,
            # preserving backward-compatible behaviour.
            kv_shape = (input_shape[0], -1, self.num_heads, self.head_dim)
            key_states = self.k_proj(current_states).view(kv_shape).transpose(1, 2).contiguous()
            value_states = self.v_proj(current_states).view(kv_shape).transpose(1, 2).contiguous()
            if past_key_values is not None:
                key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )

        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            dropout=0.0 if not self.training else self.dropout,
            scaling=1.0,
            output_attentions=output_attentions,
            **kwargs,
        )

        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        attn_output = self.out_proj(attn_output)

        return attn_output, attn_weights


class AudioFlamingo3EncoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: AudioFlamingo3Config):
        super().__init__()
        self.embed_dim = config.d_model

        self.self_attn = AudioFlamingo3Attention(
            embed_dim=self.embed_dim,
            num_heads=config.encoder_attention_heads,
            dropout=config.attention_dropout,
            config=config,
        )
        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.dropout = config.dropout
        self.activation_fn = ACT2FN[config.activation_function]
        self.activation_dropout = config.activation_dropout
        self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
        self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
        self.final_layer_norm = nn.LayerNorm(self.embed_dim)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor,
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`): attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
        """
        residual = hidden_states
        hidden_states = self.self_attn_layer_norm(hidden_states)
        hidden_states, _ = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            **kwargs,
        )
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.final_layer_norm(hidden_states)
        hidden_states = self.activation_fn(self.fc1(hidden_states))
        hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
        hidden_states = self.fc2(hidden_states)
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states

        if hidden_states.dtype == torch.float16:
            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)

        return hidden_states


@auto_docstring
class AudioFlamingo3PreTrainedModel(PreTrainedModel):
    config: AudioFlamingo3Config
    base_model_prefix = "model"
    input_modalities = ("audio", "text")
    supports_gradient_checkpointing = True
    _no_split_modules = ["AudioFlamingo3Attention"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn = True
    _supports_sdpa = True


@auto_docstring(
    custom_intro="""
    The audio model from AudioFlamingo3 without any head or projection on top.
    """
)
class AudioFlamingo3Encoder(AudioFlamingo3PreTrainedModel):
    """
    AudioFlamingo3 encoder: Whisper encoder, average pool (time/2), then LayerNorm.
    """

    # Ignore copy
    config: AudioFlamingo3EncoderConfig
    main_input_name = "input_features"
    input_modalities = "audio"
    _no_split_modules = ["AudioFlamingo3EncoderLayer"]

    _can_record_outputs = {
        "hidden_states": AudioFlamingo3EncoderLayer,
        "attentions": AudioFlamingo3Attention,
    }

    def __init__(self, config: AudioFlamingo3EncoderConfig):
        super().__init__(config)
        self.dropout = config.dropout
        self.layerdrop = config.encoder_layerdrop

        embed_dim = config.d_model
        self.num_mel_bins = config.num_mel_bins
        self.max_source_positions = config.max_source_positions
        self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0

        self.conv1 = nn.Conv1d(self.num_mel_bins, embed_dim, kernel_size=3, padding=1)
        self.conv2 = nn.Conv1d(embed_dim, embed_dim, kernel_size=3, stride=2, padding=1)

        self.embed_positions = nn.Embedding(self.max_source_positions, embed_dim)
        self.embed_positions.requires_grad_(False)

        self.layers = nn.ModuleList([AudioFlamingo3EncoderLayer(config) for _ in range(config.encoder_layers)])
        self.layer_norm = nn.LayerNorm(config.d_model)
        # Ignore copy
        self.avg_pooler = nn.AvgPool1d(2, stride=2)

        self.gradient_checkpointing = False
        # Initialize weights and apply final processing
        self.post_init()

    def _freeze_parameters(self):
        for param in self.parameters():
            param.requires_grad = False
        self._requires_grad = False

    def get_input_embeddings(self) -> nn.Module:
        return self.conv1

    def set_input_embeddings(self, value: nn.Module):
        self.conv1 = value

    @merge_with_config_defaults
    @capture_outputs
    def forward(
        self,
        input_features: torch.Tensor,
        input_features_mask: torch.Tensor | None = None,
        **kwargs,
    ) -> tuple | BaseModelOutputWithPooling:
        r"""
        Args:
            input_features (`torch.FloatTensor` of shape `(batch_size, feature_size, sequence_length)`):
                Log-Mel features extracted from raw audio. Use the processor/feature extractor to compute and pad
                these features from waveform input.
            input_features_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding feature indices. Mask values selected in `[0, 1]`:

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.
        """

        seq_len = (input_features.shape[-1] - 1) // 2 + 1  # After conv2 downsampling
        input_features_lengths = input_features_mask.sum(-1)
        input_features_lengths = (input_features_lengths - 1) // 2 + 1  # conv2 downsampling
        input_features_mask = torch.arange(seq_len, device=input_features.device) < input_features_lengths[:, None]

        # Conv front-end
        inputs_embeds = nn.functional.gelu(self.conv1(input_features))
        inputs_embeds = nn.functional.gelu(self.conv2(inputs_embeds))
        inputs_embeds = inputs_embeds.permute(0, 2, 1)

        # Add positions, dropout
        hidden_states = inputs_embeds + self.embed_positions.weight
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)

        attention_mask = create_bidirectional_mask(
            config=self.config,
            inputs_embeds=hidden_states,
            attention_mask=input_features_mask,
        )

        # Transformer stack
        for layer in self.layers:
            drop = self.training and torch.rand([]) < self.layerdrop
            if not drop:
                hidden_states = layer(hidden_states, attention_mask)

        # AvgPool (time/2) + LayerNorm
        hidden_states = hidden_states.permute(0, 2, 1)
        hidden_states = self.avg_pooler(hidden_states).permute(0, 2, 1)
        hidden_states = self.layer_norm(hidden_states)

        return BaseModelOutputWithPooling(
            last_hidden_state=hidden_states,
        )

    # Ignore copy
    def _get_feat_extract_output_lengths(self, input_lengths: torch.LongTensor):
        """
        Computes the output length of the convolutional layers and the output length of the audio encoder
        """
        input_lengths = (input_lengths - 1) // 2 + 1
        output_lengths = (input_lengths - 2) // 2 + 1
        return input_lengths, output_lengths


class AudioFlamingo3MultiModalProjector(nn.Module):
    """
    Audio adaptor (small MLP) that projects AudioFlamingo3Encoder features
    to the LLM embedding space so they can replace `<sound>` tokens.
    """

    def __init__(self, config: AudioFlamingo3Config):
        super().__init__()
        self.linear_1 = nn.Linear(
            config.audio_config.hidden_size, config.text_config.hidden_size, bias=config.projector_bias
        )
        self.act = ACT2FN[config.projector_hidden_act]
        self.linear_2 = nn.Linear(
            config.text_config.hidden_size, config.text_config.hidden_size, bias=config.projector_bias
        )

    def forward(self, audio_features):
        hidden_states = self.linear_1(audio_features)
        hidden_states = self.act(hidden_states)
        hidden_states = self.linear_2(hidden_states)
        return hidden_states


@auto_docstring(
    custom_intro="""
    The AudioFlamingo3 model which consists of a fine-tuned Whisper encoder, a multi-modal projector and a Qwen2 language model.
    """
)
class AudioFlamingo3ForConditionalGeneration(AudioFlamingo3PreTrainedModel, GenerationMixin):
    _keep_in_fp32_modules_strict = None
    _tp_plan = None
    _pp_plan = None

    def __init__(self, config):
        super().__init__(config)
        self.vocab_size = config.text_config.vocab_size
        self.audio_tower = AutoModel.from_config(config.audio_config)
        self.language_model = AutoModelForCausalLM.from_config(config.text_config)
        self.multi_modal_projector = AudioFlamingo3MultiModalProjector(config)

        # Initialize weights and apply final processing
        self.post_init()

    def get_input_embeddings(self):
        return self.language_model.get_input_embeddings()

    def set_input_embeddings(self, value):
        self.language_model.set_input_embeddings(value)

    def get_output_embeddings(self):
        return self.language_model.get_output_embeddings()

    def set_output_embeddings(self, new_embeddings):
        self.language_model.set_output_embeddings(new_embeddings)

    def set_decoder(self, decoder):
        self.language_model.set_decoder(decoder)

    def get_decoder(self):
        return self.language_model.get_decoder()

    @can_return_tuple
    @auto_docstring(
        custom_intro="This method is used to get the audio embeddings from input features (a log mel spectrogram), meaning inferring the audio encoder and the multi-modal projector."
    )
    def get_audio_features(
        self,
        input_features: torch.FloatTensor,
        input_features_mask: torch.Tensor,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPooling:
        r"""
        input_features (`torch.FloatTensor`):
            Float values of mel features extracted from the raw speech waveform. Raw speech waveform can be
            obtained by loading a `.flac` or `.wav` audio file into an array of type `list[float]` or a
            `numpy.ndarray`, *e.g.* via the soundfile library (`pip install soundfile`). To prepare the array into
            `input_features`, the [`AutoFeatureExtractor`] should be used for extracting the mel features, padding
            and conversion into a tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`]
        input_features_mask (`torch.Tensor` of shape `(batch_size, feature_sequence_length)`):
            Mask to avoid performing attention on padded feature indices.
        """

        audio_output = self.audio_tower(
            input_features, input_features_mask=input_features_mask, return_dict=True, **kwargs
        )
        audio_embeds = self.multi_modal_projector(audio_output.last_hidden_state)

        # Mask according to the audio tower output lengths, accounting for both conv downsampling and final avg pooling
        input_lengths = input_features_mask.sum(-1).to(torch.long)
        _, post_lengths = self.audio_tower._get_feat_extract_output_lengths(input_lengths)
        valid_mask = torch.arange(audio_embeds.shape[1], device=post_lengths.device)[None, :] < post_lengths[:, None]
        audio_output.pooler_output = audio_embeds[valid_mask.to(audio_embeds.device)]

        return audio_output

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        input_features: torch.FloatTensor | None = None,
        input_features_mask: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> CausalLMOutputWithPast:
        r"""
        input_features_mask (`torch.Tensor` of shape `(batch_size, feature_sequence_length)`):
            Mask to avoid performing attention on padding feature indices. Mask values selected in `[0, 1]`:

            - 1 for tokens that are **not masked**,
            - 0 for tokens that are **masked**.
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.

        Example:

        ```python
        >>> from transformers import AudioFlamingo3ForConditionalGeneration, AutoProcessor

        >>> model_id = "nvidia/audio-flamingo-3-hf"
        >>> processor = AutoProcessor.from_pretrained(model_id)
        >>> model = AudioFlamingo3ForConditionalGeneration.from_pretrained(model_id, device_map="auto")

        >>> conversations = [
        >>>     [
        >>>         {
        >>>             "role": "user",
        >>>             "content": [
        >>>                 {"type": "text", "text": "Transcribe the input speech."},
        >>>                 {
        >>>                     "type": "audio",
        >>>                     "path": "https://huggingface.co/datasets/nvidia/AudioSkills/resolve/main/assets/t_837b89f2-26aa-4ee2-bdf6-f73f0dd59b26.wav",
        >>>                 },
        >>>             ],
        >>>         }
        >>>     ],
        >>>     [
        >>>         {
        >>>             "role": "user",
        >>>             "content": [
        >>>                 {
        >>>                     "type": "text",
        >>>                     "text": "This track feels really peaceful and introspective. What elements make it feel so calming and meditative?",
        >>>                 },
        >>>                 {"type": "audio", "path": "https://huggingface.co/datasets/nvidia/AudioSkills/resolve/main/assets/FPSbCAANfbJLVSwD.mp3"},
        >>>             ],
        >>>         }
        >>>     ],
        >>> ]

        >>> inputs = processor.apply_chat_template(
        >>>     conversations,
        >>>     tokenize=True,
        >>>     add_generation_prompt=True,
        >>>     return_dict=True,
        >>> ).to(model.device)

        >>> outputs = model.generate(**inputs, max_new_tokens=500)

        >>> decoded_outputs = processor.batch_decode(
        >>>     outputs[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True
        >>> )
        >>> print(decoded_outputs)
        ["The spoken content of the audio is...", "The track's calming and meditative feel can be attributed to..."]
        ```"""

        if inputs_embeds is None:
            inputs_embeds = self.get_input_embeddings()(input_ids)

        if input_features is not None and input_ids is not None:
            audio_embeds = self.get_audio_features(input_features, input_features_mask, return_dict=True).pooler_output

            # replace text-audio token placeholders with audio embeddings
            audio_token_mask = (input_ids == self.config.audio_token_id).unsqueeze(-1)
            inputs_embeds = inputs_embeds.masked_scatter(
                audio_token_mask.to(inputs_embeds.device), audio_embeds.to(inputs_embeds.device)
            )

        outputs: CausalLMOutputWithPast = self.language_model(
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            labels=labels,
            use_cache=use_cache,
            logits_to_keep=logits_to_keep,
            **kwargs,
        )
        return outputs

    def prepare_inputs_for_generation(self, *args, is_first_iteration: bool = False, **kwargs):
        input_features = kwargs.pop("input_features", None)
        input_features_mask = kwargs.pop("input_features_mask", None)

        model_inputs = super().prepare_inputs_for_generation(*args, **kwargs)

        if is_first_iteration or not model_inputs.get("use_cache", False):
            if input_features is not None:
                model_inputs["input_features"] = input_features
            if input_features_mask is not None:
                model_inputs["input_features_mask"] = input_features_mask

        return model_inputs


__all__ = ["AudioFlamingo3ForConditionalGeneration", "AudioFlamingo3PreTrainedModel", "AudioFlamingo3Encoder"]