transformers 5.0.0rc0__py3-none-any.whl → 5.0.0rc2__py3-none-any.whl

transformers/models/pe_audio_video/modular_pe_audio_video.py

@@ -0,0 +1,764 @@
+# coding=utf-8
+# Copyright 2025 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 collections.abc import Callable
+from dataclasses import dataclass
+from typing import Any, Optional
+
+import torch
+import torch.nn as nn
+
+from ... import initialization as init
+from ...modeling_attn_mask_utils import _prepare_4d_attention_mask
+from ...modeling_outputs import BaseModelOutputWithPooling, MaskedLMOutput
+from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel, eager_attention_forward
+from ...processing_utils import Unpack
+from ...utils import ModelOutput, TransformersKwargs, auto_docstring, can_return_tuple
+from ...utils.generic import check_model_inputs
+from ..auto import AutoModel
+from ..qwen3.modeling_qwen3 import Qwen3Attention, Qwen3DecoderLayer, Qwen3RMSNorm, Qwen3RotaryEmbedding
+from .configuration_pe_audio_video import PeAudioVideoConfig, PeAudioVideoEncoderConfig
+
+
+class PeAudioVideoMaskedGroupNorm(nn.GroupNorm):
+    def forward(self, x, padding_mask=None):
+        if padding_mask is None:
+            return super().forward(x)
+
+        batch_size, hidden_size, seq_len = x.shape
+        group_size = hidden_size // self.num_groups
+        grouped_shape = (batch_size, -1, group_size, seq_len)
+
+        x_grouped = x.view(grouped_shape)
+        padding_mask_grouped = padding_mask.reshape(grouped_shape).bool()
+
+        mean = torch.masked.mean(x_grouped, mask=padding_mask_grouped, dim=(2, 3), keepdim=True)
+        var = torch.masked.var(x_grouped, mask=padding_mask_grouped, dim=(2, 3), keepdim=True, unbiased=False)
+
+        x_norm = (x_grouped - mean) / torch.sqrt(var + self.eps)
+        x_norm = x_norm.view(x.shape)
+
+        if self.affine:
+            x_norm = x_norm * self.weight.view(1, -1, 1) + self.bias.view(1, -1, 1)
+
+        return x_norm * padding_mask
+
+
+class PeAudioVideoConvBlock1d(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.groupnorm = PeAudioVideoMaskedGroupNorm(num_groups=1, num_channels=config.hidden_size)
+        self.activation = nn.SiLU()
+        self.project = nn.Conv1d(
+            in_channels=config.hidden_size,
+            out_channels=config.hidden_size,
+            kernel_size=3,
+            padding="same",
+        )
+
+    def forward(self, x, padding_mask=None):
+        x = self.groupnorm(x, padding_mask=padding_mask)
+        x = self.activation(x)
+        return self.project(x)
+
+
+class PeAudioVideoResnetBlock1d(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.block1 = PeAudioVideoConvBlock1d(config)
+        self.block2 = PeAudioVideoConvBlock1d(config)
+
+    def forward(self, hidden_states, padding_mask=None):
+        """
+        Args:
+            hidden_states: (batch_size, seq_len, hidden_size)
+            padding_mask: (batch_size, seq_len)
+        Returns:
+            hidden_states: (batch_size, seq_len, hidden_size)
+        """
+        # transpose for convolutions
+        # (batch_size, seq_len, hidden_size) -> (batch_size, hidden_size, seq_len)
+        hidden_states = hidden_states.transpose(1, 2)
+
+        if padding_mask is not None:
+            padding_mask = padding_mask.unsqueeze(1).expand_as(hidden_states)
+
+        residual = hidden_states
+        hidden_states = self.block1(hidden_states, padding_mask=padding_mask)
+        hidden_states = self.block2(hidden_states, padding_mask=padding_mask)
+        hidden_states = residual + hidden_states
+
+        return hidden_states.transpose(1, 2)
+
+
+class PeAudioVideoEncoderPatchEmbedder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.resnet_block = PeAudioVideoResnetBlock1d(config)
+        self.class_embedding = nn.Parameter(torch.randn(1, 1, config.hidden_size))
+
+    def forward(self, inputs_embeds, padding_mask=None):
+        # Embedding step: prepend class token and run the ResNet block.
+        hidden_states = torch.cat(
+            [self.class_embedding.expand(inputs_embeds.size(0), -1, -1), inputs_embeds],
+            dim=1,
+        )
+
+        if padding_mask is not None:
+            # TODO: any reason why we take padding_mask[0] and not just 1?
+            padding_mask = torch.cat([padding_mask[:, [0]], padding_mask], dim=1)
+
+        hidden_states = self.resnet_block(hidden_states, padding_mask=padding_mask)
+        return hidden_states, padding_mask
+
+
+class PeAudioVideoContrastiveHead(nn.Module):
+    def __init__(
+        self,
+        in_dim: int,
+        out_dim: int,
+    ) -> None:
+        super().__init__()
+        self.layer_norm = nn.LayerNorm(normalized_shape=in_dim, eps=1e-6)
+        self.proj = nn.Linear(in_dim, out_dim, bias=False)
+
+    def forward(self, x: torch.Tensor) -> torch.FloatTensor:
+        return self.proj(self.layer_norm(x))
+
+
+class PeAudioVideoEncoderEmbedder(nn.Module):
+    def __init__(self, config: PeAudioVideoEncoderConfig):
+        super().__init__()
+        self.audio_encoder = AutoModel.from_config(config.audio_config)
+        self.video_encoder = AutoModel.from_config(config.video_config)
+
+        self.video_proj = nn.Conv1d(config.video_config.hidden_size, config.audio_config.hidden_size, 1)
+        self.video_norm = nn.LayerNorm(config.audio_config.hidden_size)
+
+        self.concat_modality_proj = nn.Linear(
+            config.audio_config.hidden_size + config.video_config.hidden_size,
+            config.hidden_size,
+        )
+        self.data_proj = nn.Linear(config.hidden_size, config.hidden_size)
+
+    def _align_video_hidden_state(
+        self,
+        video_hidden_state: torch.Tensor,
+        audio_hidden_state: torch.Tensor,
+        padding_mask_videos: Optional[torch.Tensor] = None,
+        padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Align video_hidden_state to audio_hidden_state by nearest neighbor interpolation.
+        """
+        if video_hidden_state.shape[1] == audio_hidden_state.shape[1]:
+            return video_hidden_state
+
+        if padding_mask_videos is not None:
+            video_lengths = padding_mask_videos.sum(dim=-1)
+        else:
+            video_lengths = video_hidden_state.shape[1] * video_hidden_state.new_ones(
+                video_hidden_state.shape[0], dtype=torch.long
+            )
+
+        if padding_mask is not None:
+            audio_lengths = padding_mask.sum(dim=-1)
+        else:
+            audio_lengths = audio_hidden_state.shape[1] * audio_hidden_state.new_ones(
+                audio_hidden_state.shape[0], dtype=torch.long
+            )
+
+        if (audio_lengths == video_hidden_state.shape[1]).all() or (
+            video_lengths == audio_hidden_state.shape[1]
+        ).all():
+            # no need to align taking into account the padding masks
+            # note: when one of the above is true, we can expect the other to be true as there is no reason
+            # to have masked audio without masked video and vice versa
+
+            return nn.functional.interpolate(video_hidden_state, size=audio_hidden_state.shape[1], mode="nearest")
+
+        aligned_shape = (*audio_hidden_state.shape[:2], video_hidden_state.shape[-1])
+        aligned_hidden_state = audio_hidden_state.new_zeros(aligned_shape)
+
+        for i, (hidden_state, video_length, audio_length) in enumerate(
+            zip(video_hidden_state, video_lengths, audio_lengths)
+        ):
+            hidden_state = hidden_state[:video_length]
+            if hidden_state.numel() > 0 and audio_length > 0:
+                interpolated_hidden_state = nn.functional.interpolate(
+                    hidden_state[None].transpose(1, 2), size=audio_length, mode="nearest"
+                ).transpose(1, 2)[0]
+                aligned_hidden_state[i, :audio_length, :] = interpolated_hidden_state
+
+        return aligned_hidden_state
+
+    def forward(
+        self,
+        input_values: torch.Tensor,
+        pixel_values_videos: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+        padding_mask_videos: Optional[torch.Tensor] = None,
+    ):
+        audio_output = self.audio_encoder(input_values, padding_mask=padding_mask)
+        video_output = self.video_encoder(pixel_values_videos, padding_mask_videos=padding_mask_videos)
+
+        audio_hidden_state = audio_output.last_hidden_state
+        video_hidden_state = video_output.last_hidden_state
+        padding_mask = audio_output.output_mask
+
+        video_hidden_state = self.video_proj(video_hidden_state.transpose(1, 2)).transpose(1, 2)
+        video_hidden_state = self._align_video_hidden_state(
+            video_hidden_state=video_hidden_state,
+            audio_hidden_state=audio_hidden_state,
+            padding_mask_videos=padding_mask_videos,
+            padding_mask=padding_mask,
+        )
+        video_hidden_state = self.video_norm(video_hidden_state)
+        inputs_embeds = torch.cat([audio_hidden_state, video_hidden_state], dim=-1)
+        inputs_embeds = self.concat_modality_proj(inputs_embeds)
+        inputs_embeds = self.data_proj(inputs_embeds)
+
+        return inputs_embeds, padding_mask, audio_output, video_output
+
+
+class PeAudioVideoEncoderAttention(Qwen3Attention):
+    def __init__(self, config, layer_idx):
+        super().__init__(config, layer_idx)
+        self.is_causal = False
+        del self.sliding_window
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = self.q_norm(self.q_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
+        key_states = self.k_norm(self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            **kwargs,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+class PeAudioVideoEncoderLayer(Qwen3DecoderLayer):
+    def __init__(self, config, layer_idx):
+        super().__init__(config, layer_idx)
+        del self.attention_type
+
+
+class PeAudioVideoEncoderRMSNorm(Qwen3RMSNorm): ...
+
+
+def stack_freqs(cos: torch.Tensor, sin: torch.Tensor):
+    dim = cos.size(-1)
+    cos = cos.narrow(-1, 0, dim // 2)
+    sin = sin.narrow(-1, 0, dim // 2)
+    freqs_cis = torch.stack((cos, -sin, sin, cos), dim=-1).view(*cos.size(), 2, 2)
+    return freqs_cis
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
+    freqs_cis = stack_freqs(cos, sin)
+    freqs_cis = freqs_cis.unsqueeze(unsqueeze_dim)
+    q_ = q.reshape(*q.shape[:-1], -1, 1, 2)
+    k_ = k.reshape(*k.shape[:-1], -1, 1, 2)
+    return (q_ * freqs_cis).sum(5).flatten(3), (k_ * freqs_cis).sum(5).flatten(3)
+
+
+class PeAudioVideoEncoderRotaryEmbedding(Qwen3RotaryEmbedding): ...
+
+
+@auto_docstring
+class PeAudioVideoPreTrainedModel(PreTrainedModel):
+    config: PeAudioVideoConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["PeAudioVideoEncoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+
+    _can_compile_fullgraph = True
+    _supports_attention_backend = True
+    _can_record_outputs = {
+        "hidden_states": PeAudioVideoEncoderLayer,
+        "attentions": PeAudioVideoEncoderAttention,
+    }
+
+    def _init_weights(self, module):
+        super()._init_weights(module)
+
+        if hasattr(self.config, "initializer_range"):
+            std = self.config.initializer_range
+        else:
+            # 0.02 is the standard default value across the library
+            std = getattr(self.config.get_text_config(), "initializer_range", 0.02)
+
+        if isinstance(module, PeAudioVideoEncoderPatchEmbedder):
+            embed_dim = module.class_embedding.shape[-1]
+            init.normal_(module.class_embedding, mean=0.0, std=embed_dim**-0.5 * std)
+
+
+@dataclass
+@auto_docstring(
+    custom_intro="""
+    Class for outputs of [`PeAudioVideoEncoder`].
+    """
+)
+class PeAudioVideoEncoderOutput(BaseModelOutputWithPooling):
+    audio_model_output: Optional[BaseModelOutputWithPooling] = None
+    video_model_output: Optional[BaseModelOutputWithPooling] = None
+
+
+@auto_docstring(
+    custom_intro="""
+    The PeAudioVideo Encoder model.
+    """
+)
+class PeAudioVideoEncoder(PeAudioVideoPreTrainedModel):
+    config: PeAudioVideoEncoderConfig
+    main_input_name = "input_values"
+    base_model_prefix = "audio_video_encoder"
+
+    def __init__(self, config: PeAudioVideoEncoderConfig):
+        super().__init__(config)
+        self.embedder = PeAudioVideoEncoderEmbedder(config)
+        self.patch_embedder = PeAudioVideoEncoderPatchEmbedder(config)
+        self.layers = nn.ModuleList(
+            [PeAudioVideoEncoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = PeAudioVideoEncoderRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = PeAudioVideoEncoderRotaryEmbedding(config=config)
+        self.output = nn.Linear(config.hidden_size, config.hidden_size, bias=False)
+        self.gradient_checkpointing = False
+
+        self.post_init()
+
+    @can_return_tuple
+    @check_model_inputs
+    def forward(
+        self,
+        input_values: torch.Tensor | None = None,
+        pixel_values_videos: torch.Tensor | None = None,
+        padding_mask: Optional[torch.Tensor] = None,
+        padding_mask_videos: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> PeAudioVideoEncoderOutput:
+        inputs_embeds, padding_mask, audio_output, video_output = self.embedder(
+            input_values,
+            pixel_values_videos,
+            padding_mask=padding_mask,
+            padding_mask_videos=padding_mask_videos,
+        )
+        inputs_embeds, attention_mask = self.patch_embedder(inputs_embeds, padding_mask=padding_mask)
+
+        if attention_mask is not None:
+            attention_mask = _prepare_4d_attention_mask(attention_mask, inputs_embeds.dtype)
+
+        position_ids = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device).unsqueeze(0)
+        position_embeddings = self.rotary_emb(inputs_embeds, position_ids)
+
+        hidden_states = inputs_embeds
+        for encoder_layer in self.layers[: self.config.num_hidden_layers]:
+            hidden_states = encoder_layer(
+                hidden_states,
+                attention_mask=attention_mask,
+                position_embeddings=position_embeddings,
+                **kwargs,
+            )
+
+        hidden_states = self.norm(hidden_states)
+        hidden_states = self.output(hidden_states)
+
+        return PeAudioVideoEncoderOutput(
+            last_hidden_state=hidden_states[:, 1:],
+            pooler_output=hidden_states[:, 0],
+            audio_model_output=audio_output,
+            video_model_output=video_output,
+        )
+
+
+@dataclass
+@auto_docstring(
+    custom_intro="""
+    Class for outputs of [`PeAudioVideoModel`] when using text, audio, and/or video.
+    """
+)
+class PeAudioVideoOutput(ModelOutput):
+    # embeddings
+    audio_embeds: Optional[torch.FloatTensor] = None
+    video_embeds: Optional[torch.FloatTensor] = None
+    audio_video_embeds: Optional[torch.FloatTensor] = None
+    text_audio_embeds: Optional[torch.FloatTensor] = None
+    text_video_embeds: Optional[torch.FloatTensor] = None
+    text_audio_video_embeds: Optional[torch.FloatTensor] = None
+    audio_plus_text_embeds: Optional[torch.FloatTensor] = None
+    video_plus_text_embeds: Optional[torch.FloatTensor] = None
+
+    # model outputs
+    # TODO: update types to the correct ones
+    text_outputs: Optional[MaskedLMOutput] = None
+    audio_outputs: Optional[BaseModelOutputWithPooling] = None
+    video_outputs: Optional[BaseModelOutputWithPooling] = None
+    audio_video_outputs: Optional[BaseModelOutputWithPooling] = None
+
+    # logits
+    logits_audio_text: Optional[torch.FloatTensor] = None
+    logits_video_text: Optional[torch.FloatTensor] = None
+    logits_audio_video: Optional[torch.FloatTensor] = None
+    logits_audio_video_text: Optional[torch.FloatTensor] = None
+    logits_audio_plus_text_video: Optional[torch.FloatTensor] = None
+    logits_video_plus_text_audio: Optional[torch.FloatTensor] = None
+
+    audio_text_loss: Optional[torch.FloatTensor] = None
+    video_text_loss: Optional[torch.FloatTensor] = None
+    audio_video_loss: Optional[torch.FloatTensor] = None
+    audio_video_text_loss: Optional[torch.FloatTensor] = None
+    audio_plus_text_video_loss: Optional[torch.FloatTensor] = None
+    video_plus_text_audio_loss: Optional[torch.FloatTensor] = None
+    loss: Optional[torch.FloatTensor] = None
+
+    def to_tuple(self) -> tuple[Any]:
+        return tuple(self[k] if not k.endswith("model_output") else getattr(self, k).to_tuple() for k in self.keys())
+
+
+@dataclass
+class AudioVideoEmbeddings(ModelOutput):
+    audio_embeds: Optional[torch.FloatTensor] = None
+    video_embeds: Optional[torch.FloatTensor] = None
+    audio_video_embeds: Optional[torch.FloatTensor] = None
+
+
+class PeAudioVideoModel(PeAudioVideoPreTrainedModel):
+    _tied_weights_keys = {
+        r"audio_model\.text_model\.(?!rotary_emb)": r"^text_model\.(?!rotary_emb)",
+        r"video_model\.text_model\.(?!rotary_emb)": r"^text_model\.(?!rotary_emb)",
+        r"audio_video_encoder\.embedder\.audio_encoder\.(?!rotary_emb)": r"audio_model\.audio_encoder\.(?!rotary_emb)",
+        r"audio_video_encoder\.embedder\.video_encoder\.(?!rotary_emb|.*\.rope\.pos_embed)": r"video_model\.video_encoder\.(?!rotary_emb|.*\.rope\.pos_embed)",
+    }
+
+    def __init__(self, config: PeAudioVideoConfig):
+        super().__init__(config)
+        self.text_model = AutoModel.from_config(config.text_config)
+        self.audio_model = AutoModel.from_config(config.audio_config)
+        self.video_model = AutoModel.from_config(config.video_config)
+        self.audio_video_encoder = PeAudioVideoEncoder(config.audio_video_config)
+
+        text_hidden_size = config.text_config.hidden_size
+        audio_hidden_size = config.audio_video_config.audio_config.hidden_size
+        video_hidden_size = config.audio_video_config.video_config.hidden_size
+
+        # audio-video
+        self.audio_video_head = PeAudioVideoContrastiveHead(config.audio_video_config.hidden_size, text_hidden_size)
+        self.text_audio_video_head = PeAudioVideoContrastiveHead(text_hidden_size, text_hidden_size)
+        self.audio_video_logit_scale = nn.Parameter(torch.zeros(1))
+        self.audio_video_logit_bias = nn.Parameter(torch.zeros(1))
+        self.text_audio_video_logit_scale = nn.Parameter(torch.zeros(1))
+        self.text_audio_video_logit_bias = nn.Parameter(torch.zeros(1))
+
+        # text-audio
+        self.audio_plus_text_head = PeAudioVideoContrastiveHead(text_hidden_size + audio_hidden_size, text_hidden_size)
+        self.audio_plus_text_logit_scale = nn.Parameter(torch.zeros(1))
+        self.audio_plus_text_logit_bias = nn.Parameter(torch.zeros(1))
+
+        # text-video
+        self.video_plus_text_head = PeAudioVideoContrastiveHead(text_hidden_size + video_hidden_size, text_hidden_size)
+        self.video_plus_text_logit_scale = nn.Parameter(torch.zeros(1))
+        self.video_plus_text_logit_bias = nn.Parameter(torch.zeros(1))
+
+        self.post_init()
+
+    def _contrastive_loss(self, logits: torch.Tensor) -> torch.Tensor:
+        labels = torch.eye(logits.shape[0], device=logits.device)
+        loss = -nn.functional.logsigmoid(labels * logits).sum() / logits.shape[0]
+        return loss
+
+    def get_text_audio_embeds(self, input_ids, attention_mask=None):
+        return self.audio_model.get_text_embeds(input_ids, attention_mask)
+
+    def get_text_video_embeds(self, input_ids, attention_mask=None):
+        return self.video_model.get_text_embeds(input_ids, attention_mask)
+
+    def get_text_audio_video_embeds(self, input_ids, attention_mask=None):
+        text_outputs: MaskedLMOutput = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            return_dict=True,
+        )
+        text_embeds = text_outputs.hidden_states[-1][:, 0]
+        return self.text_audio_video_head(text_embeds)
+
+    def get_audio_embeds(self, input_values, padding_mask=None):
+        return self.audio_model.get_audio_embeds(input_values, padding_mask)
+
+    def get_video_embeds(self, pixel_values_videos, padding_mask_videos=None):
+        return self.video_model.get_video_embeds(pixel_values_videos, padding_mask_videos)
+
+    def get_audio_video_embeds(
+        self,
+        input_values: torch.Tensor,
+        pixel_values_videos: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+        padding_mask_videos: Optional[torch.Tensor] = None,
+        return_audio_embeds: bool = False,
+        return_video_embeds: bool = False,
+        **kwargs,
+    ) -> AudioVideoEmbeddings:
+        audio_video_outputs = self.audio_video_encoder(
+            input_values=input_values,
+            pixel_values_videos=pixel_values_videos,
+            padding_mask=padding_mask,
+            padding_mask_videos=padding_mask_videos,
+            **kwargs,
+        )
+        if return_audio_embeds:
+            audio_embeds = self.audio_model.audio_head(audio_video_outputs.audio_model_output.pooler_output)
+        if return_video_embeds:
+            video_embeds = self.video_model.video_head(audio_video_outputs.video_model_output.pooler_output)
+
+        audio_video_embeds = self.audio_video_head(audio_video_outputs.pooler_output)
+        return AudioVideoEmbeddings(
+            audio_embeds=audio_embeds if return_audio_embeds else None,
+            video_embeds=video_embeds if return_video_embeds else None,
+            audio_video_embeds=audio_video_embeds,
+        )
+
+    def get_audio_plus_text_embeds(
+        self,
+        input_ids: torch.Tensor,
+        input_values: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        audio_embeds = self.audio_model.audio_encoder(
+            input_values=input_values,
+            padding_mask=padding_mask,
+            return_dict=True,
+        )
+        text_outputs: MaskedLMOutput = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            return_dict=True,
+        )
+        text_embeds = text_outputs.hidden_states[-1][:, 0]
+
+        audio_plus_text_embeds = torch.cat([text_embeds, audio_embeds], dim=-1)
+        return self.audio_plus_text_head(audio_plus_text_embeds)
+
+    def get_video_plus_text_embeds(
+        self,
+        input_ids: torch.Tensor,
+        pixel_values_videos: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        padding_mask_videos: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        video_embeds = self.video_model.video_encoder(
+            pixel_values_videos=pixel_values_videos,
+            padding_mask_videos=padding_mask_videos,
+            return_dict=True,
+        )
+        text_outputs: MaskedLMOutput = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            return_dict=True,
+        )
+        text_embeds = text_outputs.hidden_states[-1][:, 0]
+
+        video_plus_text_embeds = torch.cat([text_embeds, video_embeds], dim=-1)
+        return self.video_plus_text_head(video_plus_text_embeds)
+
+    @can_return_tuple
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        pixel_values_videos: Optional[torch.Tensor] = None,
+        input_values: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        padding_mask_videos: Optional[torch.Tensor] = None,
+        padding_mask: Optional[torch.Tensor] = None,
+        return_loss=False,
+        **kwargs,
+    ) -> PeAudioVideoOutput:
+        if sum([input_ids is not None, pixel_values_videos is not None, input_values is not None]) < 2:
+            raise ValueError("At least two of input_ids, pixel_values_videos, or input_values must be provided")
+
+        if pixel_values_videos is None:
+            outputs = self.audio_model(
+                input_ids=input_ids,
+                input_values=input_values,
+                attention_mask=attention_mask,
+                padding_mask=padding_mask,
+                return_dict=True,
+            )
+            audio_plus_text_embeds = torch.cat(
+                [outputs.audio_outputs.pooler_output, outputs.text_outputs.hidden_states[-1][:, 0]], dim=-1
+            )
+            audio_plus_text_embeds = self.audio_plus_text_head(audio_plus_text_embeds)
+            return PeAudioVideoOutput(audio_plus_text_embeds=audio_plus_text_embeds, **outputs)
+
+        if input_values is None:
+            outputs = self.video_model(
+                input_ids=input_ids,
+                pixel_values_videos=pixel_values_videos,
+                attention_mask=attention_mask,
+                padding_mask_videos=padding_mask_videos,
+                return_dict=True,
+            )
+            video_plus_text_embeds = torch.cat(
+                [outputs.video_outputs.pooler_output, outputs.text_outputs.hidden_states[-1][:, 0]], dim=-1
+            )
+            video_plus_text_embeds = self.video_plus_text_head(video_plus_text_embeds)
+            return PeAudioVideoOutput(video_plus_text_embeds=video_plus_text_embeds, **outputs)
+
+        audio_video_outputs = self.audio_video_encoder(
+            input_values=input_values,
+            pixel_values_videos=pixel_values_videos,
+            padding_mask=padding_mask,
+            padding_mask_videos=padding_mask_videos,
+            **kwargs,
+        )
+        audio_embeds = audio_video_outputs.audio_model_output.pooler_output
+        video_embeds = audio_video_outputs.video_model_output.pooler_output
+        audio_video_embeds = audio_video_outputs.pooler_output
+
+        audio_embeds = self.audio_model.audio_head(audio_embeds)
+        video_embeds = self.video_model.video_head(video_embeds)
+        audio_video_embeds = self.audio_video_head(audio_video_embeds)
+        logits_audio_video = audio_embeds @ video_embeds.T
+        logits_audio_video = logits_audio_video * self.audio_video_logit_scale + self.audio_video_logit_bias
+        audio_video_loss = self._contrastive_loss(logits_audio_video) if return_loss else None
+
+        if input_ids is None:
+            return PeAudioVideoOutput(
+                logits_audio_video=logits_audio_video,
+                audio_embeds=audio_embeds,
+                video_embeds=video_embeds,
+                audio_video_embeds=audio_video_embeds,
+                loss=audio_video_loss,
+                audio_video_loss=audio_video_loss,
+            )
+
+        kwargs["output_hidden_states"] = True
+        text_outputs = self.text_model(input_ids=input_ids, attention_mask=attention_mask, **kwargs)
+        text_embeds = text_outputs.hidden_states[-1][:, 0]
+        audio_plus_text_embeds = torch.cat([audio_video_outputs.audio_model_output.pooler_output, text_embeds], dim=-1)
+        video_plus_text_embeds = torch.cat([audio_video_outputs.video_model_output.pooler_output, text_embeds], dim=-1)
+
+        text_audio_embeds = self.audio_model.text_audio_head(text_embeds)
+        text_video_embeds = self.video_model.text_video_head(text_embeds)
+        text_audio_video_embeds = self.text_audio_video_head(text_embeds)
+        audio_plus_text_embeds = self.audio_plus_text_head(audio_plus_text_embeds)
+        video_plus_text_embeds = self.video_plus_text_head(video_plus_text_embeds)
+
+        logits_audio_text = audio_embeds @ text_audio_embeds.T
+        logits_video_text = video_embeds @ text_video_embeds.T
+        logits_audio_video_text = audio_video_embeds @ text_audio_video_embeds.T
+
+        logits_audio_plus_text_video = audio_plus_text_embeds @ video_embeds.T
+        logits_video_plus_text_audio = video_plus_text_embeds @ audio_embeds.T
+
+        logits_audio_text = (
+            logits_audio_text * self.audio_model.text_audio_logit_scale + self.audio_model.text_audio_logit_bias
+        )
+        logits_video_text = (
+            logits_video_text * self.video_model.text_video_logit_scale + self.video_model.text_video_logit_bias
+        )
+        logits_audio_video_text = (
+            logits_audio_video_text * self.text_audio_video_logit_scale + self.text_audio_video_logit_bias
+        )
+
+        logits_audio_plus_text_video = (
+            logits_audio_plus_text_video * self.audio_plus_text_logit_scale + self.audio_plus_text_logit_bias
+        )
+        logits_video_plus_text_audio = (
+            logits_video_plus_text_audio * self.video_plus_text_logit_scale + self.video_plus_text_logit_bias
+        )
+
+        if return_loss:
+            audio_text_loss = self._contrastive_loss(logits_audio_text)
+            video_text_loss = self._contrastive_loss(logits_video_text)
+            audio_video_text_loss = self._contrastive_loss(logits_audio_video_text)
+            audio_plus_text_video_loss = self._contrastive_loss(logits_audio_plus_text_video)
+            video_plus_text_audio_loss = self._contrastive_loss(logits_video_plus_text_audio)
+            loss = (
+                audio_video_text_loss
+                + audio_text_loss
+                + video_text_loss
+                + audio_video_loss
+                + audio_plus_text_video_loss
+                + video_plus_text_audio_loss
+            )
+
+        return PeAudioVideoOutput(
+            # embeddings
+            audio_embeds=audio_embeds,
+            video_embeds=video_embeds,
+            audio_video_embeds=audio_video_embeds,
+            text_audio_embeds=text_audio_embeds,
+            text_video_embeds=text_video_embeds,
+            text_audio_video_embeds=text_audio_video_embeds,
+            audio_plus_text_embeds=audio_plus_text_embeds,
+            video_plus_text_embeds=video_plus_text_embeds,
+            # model outputs
+            text_outputs=text_outputs,
+            audio_outputs=audio_video_outputs.audio_model_output,
+            video_outputs=audio_video_outputs.video_model_output,
+            audio_video_outputs=audio_video_outputs,
+            # logits
+            logits_audio_text=logits_audio_text,
+            logits_video_text=logits_video_text,
+            logits_audio_video=logits_audio_video,
+            logits_audio_video_text=logits_audio_video_text,
+            logits_audio_plus_text_video=logits_audio_plus_text_video,
+            logits_video_plus_text_audio=logits_video_plus_text_audio,
+            # losses
+            audio_text_loss=audio_text_loss if return_loss else None,
+            video_text_loss=video_text_loss if return_loss else None,
+            audio_video_loss=audio_video_loss if return_loss else None,
+            audio_video_text_loss=audio_video_text_loss if return_loss else None,
+            audio_plus_text_video_loss=audio_plus_text_video_loss if return_loss else None,
+            video_plus_text_audio_loss=video_plus_text_audio_loss if return_loss else None,
+            loss=loss if return_loss else None,
+        )
+
+
+__all__ = [
+    "PeAudioVideoModel",
+    "PeAudioVideoEncoder",
+]