modeling_travisionlm.py

"""PyTorch TraVisionLM"""
import torch
from transformers import PreTrainedModel, AutoModel, AutoModelForCausalLM
from transformers.utils import logging, add_start_docstrings, ModelOutput
from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask_for_sdpa
from dataclasses import dataclass
from typing import List, Optional, Tuple, Union
from torch import nn
from transformers.cache_utils import Cache

logger = logging.get_logger(__name__)

from .configuration_travisionlm import TraVisionLMConfig

_CONFIG_FOR_DOC = "TraVisionLMConfig"

@dataclass
class TraVisionCausalLMOutputWithPast(ModelOutput):
    """
    Base class for TraVision language model (or autoregressive) outputs.

    Args:
        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
            Language modeling loss (for next-token prediction).
        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
            `(batch_size, num_heads, sequence_length, embed_size_per_head)`)

            Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
            `past_key_values` input) to speed up sequential decoding.
        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.

            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
            heads.
        image_hidden_states (`tuple(torch.FloatTensor)`, *optional*):
            Tuple of `torch.FloatTensor` (one for the output of the image embeddings, `(batch_size, num_images,
            sequence_length, hidden_size)`.

            image_hidden_states of the model produced by the vision encoder, and optionally by the perceiver
    """
    loss: Optional[torch.FloatTensor] = None
    logits: torch.FloatTensor = None
    past_key_values: Optional[Union[List[torch.FloatTensor], Cache]] = None
    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[torch.FloatTensor]] = None
    image_hidden_states: Optional[Tuple[torch.FloatTensor]] = None


class TraVisionMultiModalProjector(nn.Module):
    """
    Multimodal projector that cast the image features into the same dimension space as the language model
    """
    def __init__(self, config: TraVisionLMConfig, dropout=0.1):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(config.vision_config.projection_dim, 4*config.vision_config.projection_dim, bias=True),
            nn.GELU(),
            nn.Linear(4*config.vision_config.projection_dim, config.hidden_size, bias=True),
            nn.Dropout(dropout)
        )

    def forward(self, image_features):
        hidden_states = self.net(image_features).to(image_features.dtype)
        return hidden_states
    

TRAVISIONLM_START_DOCSTRING = r"""
    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
    etc.)

    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
    and behavior.

    Parameters:
        config ([`TraVisionLMConfig`]):
            Model configuration class with all the parameters of the model. Initializing with a config file does not
            load the weights associated with the model, only the configuration. Check out the
            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""

@add_start_docstrings(
    "The bare TraVision Model outputting raw hidden-states without any specific head on top.",
    TRAVISIONLM_START_DOCSTRING,
)

class TraVisionPreTrainedModel(PreTrainedModel):
    config_class = TraVisionLMConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["TraVisionMultiModalProjector"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn_2 = True
    _supports_sdpa = True

    def _init_weights(self, module):
        # Do NOT init the weights of the model using this class call, this is a ported version, 
        # hence not intended to be trained from scratch.
        std = (
            self.config.initializer_range
            if hasattr(self.config, "initializer_range")
            else self.config.text_config.initializer_range
        )

        if hasattr(module, "class_embedding"):
            module.class_embedding.data.normal_(mean=0.0, std=std)

        if isinstance(module, (nn.Linear, nn.Conv2d)):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()

    @property
    def _supports_sdpa(self):
        """
        Retrieve language_model's attribute to check whether the model supports
        SDPA or not.
        """
        return self.language_model._supports_sdpa


@add_start_docstrings(
    """The TraVisionLM model which consists of a vision backbone and a language model.""",
    TRAVISIONLM_START_DOCSTRING,
)
class TraVisionForCausalLM(TraVisionPreTrainedModel):
    def __init__(self, config: TraVisionLMConfig):
        super(TraVisionForCausalLM, self).__init__(config)
        self.vocab_size = config.vocab_size
        self.pad_token_id = -1 if config.pad_token_id == None else config.pad_token_id
        self._attn_implementation = config._attn_implementation
        self.gradient_checkpointing = False

        self.vision_tower = AutoModel.from_config(config=config.vision_config)
        self.vision_projector = TraVisionMultiModalProjector(config)

        language_model = AutoModelForCausalLM.from_config(
            config=config.text_config, attn_implementation=self._attn_implementation
        )

        if language_model._tied_weights_keys is not None:
            self._tied_weights_keys = [f"language_model.{k}" for k in language_model._tied_weights_keys]

        self.language_model = language_model

        self.post_init()

    # Copied from transformers.models.paligemma.modeling_paligemma.PaliGemmaForConditionalGeneration.get_input_embeddings with PaliGemma->TraVisionLM
    def get_input_embeddings(self):
        return self.language_model.get_input_embeddings()

    # Copied from transformers.models.paligemma.modeling_paligemma.PaliGemmaForConditionalGeneration.set_input_embeddings with PaliGemma->TraVisionLM
    def set_input_embeddings(self, value):
        self.language_model.set_input_embeddings(value)

    # Copied from transformers.models.paligemma.modeling_paligemma.PaliGemmaForConditionalGeneration.get_output_embeddings with PaliGemma->TraVisionLM
    def get_output_embeddings(self):
        return self.language_model.get_output_embeddings()

    # Copied from transformers.models.paligemma.modeling_paligemma.PaliGemmaForConditionalGeneration.set_output_embeddings with PaliGemma->TraVisionLM
    def set_output_embeddings(self, new_embeddings):
        self.language_model.set_output_embeddings(new_embeddings)

    # Copied from transformers.models.paligemma.modeling_paligemma.PaliGemmaForConditionalGeneration.set_decoder with PaliGemma->TraVisionLM
    def set_decoder(self, decoder):
        self.language_model.set_decoder(decoder)

    # Copied from transformers.models.paligemma.modeling_paligemma.PaliGemmaForConditionalGeneration.get_decoder with PaliGemma->TraVisionLM
    def get_decoder(self):
        return self.language_model.get_decoder()

    # Copied from transformers.models.paligemma.modeling_paligemma.PaliGemmaForConditionalGeneration.tie_weights with PaliGemma->TraVisionLM
    def tie_weights(self):
        return self.language_model.tie_weights()
    
    def resize_token_embeddings(self, new_num_tokens: Optional[int] = None, pad_to_multiple_of=None) -> nn.Embedding:
        # TODO: config.vocab_size is deprecated and will be removed in v4.43.
        # `resize_token_embeddings` should work from `modeling_utils.py``
        model_embeds = self.language_model.resize_token_embeddings(new_num_tokens, pad_to_multiple_of)
        self.config.text_config.vocab_size = model_embeds.num_embeddings
        self.config.vocab_size = model_embeds.num_embeddings
        self.vocab_size = model_embeds.num_embeddings
        return model_embeds

    # Copied from transformers.models.paligemma.modeling_paligemma.PaliGemmaForConditionalGeneration._merge_input_ids_with_image_features with PaliGemma->TraVisionLM
    """ !!! Two significant modifications are made to the original code:
    ------> 1) The pad and eos tokens are set to be the same in TraVisionProcessor. Hence, only the features corresponding to the padding mask are filtered out 
        using the attention mask.
    ------> 2) The features corresponding to both the prompts (called prefixes in PaliGemma) and labels (called suffixes in PaliGemma) are added the final embedding tensor
        and the tokens of both the prompts and labels are applied causal attention mask. All the image tokens are attended using full-attention mask. 
        NOTE: In the original PaliGemma implementation, only the suffix tokens are applied causal masking. Check out [PaliGemma arXiv Paper](https://arxiv.org/pdf/2407.07726) 
        for the details.  
    """
    def _merge_input_ids_with_image_features(
        self, image_features, inputs_embeds, input_ids, attention_mask, labels, token_type_ids, cache_position
    ):
        _, _, embed_dim = image_features.shape
        batch_size, sequence_length = input_ids.shape
        dtype, device = inputs_embeds.dtype, inputs_embeds.device
        min_dtype = torch.finfo(dtype).min

        scaled_image_features = image_features / (self.config.hidden_size**0.5)
        final_embedding = torch.zeros(
            batch_size, sequence_length, embed_dim, dtype=inputs_embeds.dtype, device=inputs_embeds.device
        )

        text_mask = (input_ids != self.config.image_token_index) & (attention_mask | input_ids != self.config.text_config.pad_token_id)
        image_mask = input_ids == self.config.image_token_index
        pad_mask = (attention_mask == 0) & (input_ids == self.config.text_config.pad_token_id)

        # expand masks to match embedding dimension
        text_mask_expanded = text_mask.unsqueeze(-1).expand(-1, -1, embed_dim).to(inputs_embeds.device)
        pad_mask_expanded = pad_mask.unsqueeze(-1).expand(-1, -1, embed_dim).to(inputs_embeds.device)
        # insert padding and text token embeddings
        final_embedding = torch.where(text_mask_expanded, inputs_embeds, final_embedding)
        final_embedding = torch.where(pad_mask_expanded, torch.zeros_like(final_embedding), final_embedding) 
        # insert image embeddings - the image mask is always less or equal to the sentence in length
        final_embedding = final_embedding.masked_scatter(
            image_mask.unsqueeze(-1).expand_as(final_embedding).to(device=final_embedding.device),
            scaled_image_features.to(device=final_embedding.device, dtype=final_embedding.dtype),
        )
        final_embedding = torch.where(pad_mask_expanded, torch.zeros_like(final_embedding), final_embedding) 
        if attention_mask is not None:
            position_ids = (attention_mask.cumsum(-1)).masked_fill_((attention_mask == 0), 1)
        else:
            position_ids = None

        if token_type_ids is not None:
            # we are training thus we need to create a full mask on the image, but causal on prompt and suffix
            target_length = cache_position[-1] + 1
            causal_mask = torch.full(
                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
            )
            if sequence_length != 1:
                causal_mask = torch.triu(causal_mask, diagonal=1)
            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
            causal_mask = causal_mask[None, None, :, :].expand(inputs_embeds.shape[0], 1, -1, -1)
            if attention_mask is not None:
                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
                mask_length = attention_mask.shape[-1]
                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
                    causal_mask.device
                )
                # unmask the prefill
                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
                    token_type_ids[:, None, None, :].to(causal_mask.device) == 0, 0
                )
                padding_mask = padding_mask == 0
                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
                    padding_mask, min_dtype
                )

            final_labels = None
            if labels is not None:
                final_labels = torch.full(
                    (batch_size, sequence_length), self.config.ignore_index, dtype=input_ids.dtype, device=input_ids.device
                )
                final_labels = torch.where((attention_mask | input_ids != self.config.text_config.pad_token_id), labels, final_labels)
        else:
            causal_mask = attention_mask.unsqueeze(1).unsqueeze(2) * attention_mask.unsqueeze(1).unsqueeze(-1)
            # invert causal mask
            causal_mask = torch.where(causal_mask == 0, min_dtype, 0).to(dtype)
            final_labels = None

        return final_embedding, causal_mask, final_labels, position_ids
    

    def forward(
        self,
        input_ids: torch.LongTensor = None,
        pixel_values: torch.FloatTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Union[List[torch.FloatTensor], Cache]] = None,
        token_type_ids: Optional[torch.LongTensor] = None,
        cache_position: Optional[torch.LongTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, TraVisionCausalLMOutputWithPast]:
        
        if labels is not None:
            use_cache = False
        
        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
        elif input_ids is not None:
            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
            input_shape = input_ids.size()
            input_ids = input_ids.view(-1, input_shape[-1])
            batch_size = input_ids.shape[0]
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
            batch_size = inputs_embeds.shape[0]
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if past_key_values is None:
            past_length = 0
            past_key_values = tuple([None] * len(self.language_model.transformer.h))
        else:
            past_length = past_key_values[0][0].size(-2)
        if position_ids is None:
            position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=input_ids.device if input_ids is not None else inputs_embeds.device)
            position_ids = position_ids.unsqueeze(0)

        # the attention mask is turned 4d after, we keep track of the original one
        input_attention_mask = attention_mask

        if inputs_embeds is None:
            # 1. Extract the input embeddings
            inputs_embeds = self.get_input_embeddings()(input_ids)

            # 2. Add the absolute positional embeddings to only text token locations in inputs_embeds
            if pixel_values is not None and inputs_embeds.shape[1] != 1:
                # Compute the initial mask for position IDs
                position_ids_mask = torch.where(input_ids != self.config.image_token_index, position_ids, 1)
                # Update the mask for positions where input_ids is not zero
                position_ids_mask[:, :-1] = torch.where(input_ids[:, :-1] != 0, position_ids_mask[:, :-1], 1)
                # Find the first position embedding locations
                first_position_embed_locs = torch.sum(position_ids_mask == 1, dim=1)
                # Adjust the mask by subtracting the first position embedding locations
                position_ids_mask.sub_(first_position_embed_locs[:, None])
                # Ensure all values in the mask are non-negative --> assign values 1 to pad and image token locations
                position_emb_ids = torch.where(position_ids_mask >= 0, position_ids_mask, 1)
                # construct position embeddings using position_emb_ids
                position_embeds = self.language_model.transformer.wpe(position_emb_ids)
            else:
                # In this case, we generate from cache with past_key_values 
                pos_emb_ind = position_ids.view(batch_size, -1)
                position_embeds = self.language_model.transformer.wpe(pos_emb_ind)

            # Directly add position_embeds to inputs_embeds to get hidden_states
            hidden_states = inputs_embeds + position_embeds

            # 3. Merge text and images
            if pixel_values is not None and input_ids.shape[1] != 1:
                # make sure that pixel values are of 4D dimensions (batch_size, num_channels, width, height)
                if pixel_values.dim() == 3:
                    pixel_values = pixel_values.unsqueeze(dim=0)

                image_outputs = self.vision_tower(pixel_values.to(inputs_embeds.dtype))
                selected_image_feature = image_outputs.last_hidden_state
                image_features = self.vision_projector(selected_image_feature)

                if cache_position is None:
                    cache_position = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device)
                hidden_states, attention_mask, labels, _ = self._merge_input_ids_with_image_features(
                    image_features, hidden_states, input_ids, attention_mask, labels, token_type_ids, cache_position
                )

            else:
                # In case input_ids.shape[1] == 1 & pixel_values==None & past_key_values != None, we are in the case of
                # generation with cache, we can use standard causal masking
                if past_key_values is not None and pixel_values is not None and input_ids.shape[1] == 1:
                    # Attention mask.
                    _use_sdpa = self._attn_implementation == "sdpa" and output_attentions is False
                    if attention_mask is not None:
                        attention_mask = attention_mask.view(batch_size, -1)
                        if self._attn_implementation == "flash_attention_2":
                            attention_mask = attention_mask if 0 in attention_mask else None
                        elif _use_sdpa:
                            attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
                                attention_mask=attention_mask,
                                input_shape=(batch_size, input_shape[-1]),
                                inputs_embeds=inputs_embeds,
                                past_key_values_length=past_length,
                            )
                        else:
                            # We create a 3D attention mask from a 2D tensor mask.
                            # Sizes are [batch_size, 1, 1, to_seq_length]
                            # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
                            # this attention mask is more simple than the triangular masking of causal attention
                            # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
                            attention_mask = attention_mask[:, None, None, :]

                            # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
                            # masked positions, this operation will create a tensor which is 0.0 for
                            # positions we want to attend and the dtype's smallest value for masked positions.
                            # Since we are adding it to the raw scores before the softmax, this is
                            # effectively the same as removing these entirely.
                            attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility
                            attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min

        if attention_mask is not None:
            attention_mask = attention_mask.to(inputs_embeds.dtype)

        hidden_states = self.language_model.transformer.drop(hidden_states)
        output_shape = (-1,) + input_shape[1:] + (hidden_states.size(-1),)
        
        presents = () if use_cache else None
        all_self_attentions = () if output_attentions else None
        all_hidden_states = () if output_hidden_states else None
        for i, (block, layer_past) in enumerate(zip(self.language_model.transformer.h, past_key_values)):
            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)
            outputs = block(
                hidden_states,
                layer_past=layer_past,
                attention_mask=attention_mask,
                use_cache=use_cache,
                output_attentions=output_attentions,
            )
            hidden_states = outputs[0]
            if use_cache is True:
                presents = presents + (outputs[1],)

            if output_attentions:
                all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)

        hidden_states = self.language_model.transformer.ln_f(hidden_states)

        hidden_states = hidden_states.view(output_shape)
        # Add last hidden state
        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        logits = self.language_model.lm_head(hidden_states)
        logits = logits.float()
        loss = None
        if labels is not None:
            shift_logits = logits[..., :-1, :] 
            shift_labels = labels[..., 1:] # shift to right
            if input_attention_mask is not None:
                # we use the input attention mask to shift the logits and labels, because it is 2D.
                shift_attention_mask = input_attention_mask[..., 1:]
                shift_logits = shift_logits[shift_attention_mask.to(logits.device) != 0].contiguous()
                shift_labels = shift_labels[shift_attention_mask.to(shift_labels.device) != 0].contiguous()
            else:
                shift_logits = shift_logits.contiguous()
                shift_labels = shift_labels.contiguous()
            # Flatten the tokens
            loss_fct = nn.CrossEntropyLoss()

            flat_logits = shift_logits.view(-1, self.config.vocab_size)
            flat_labels = shift_labels.view(-1).to(shift_logits.device)
            loss = loss_fct(flat_logits, flat_labels)
        if not return_dict:
            output = (logits, presents, all_hidden_states, all_self_attentions)
            return (loss,) + output if loss is not None else output

        return TraVisionCausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=presents,
            hidden_states=all_hidden_states,
            attentions=all_self_attentions,
        )
    
    def prepare_inputs_for_generation(
            self,
            input_ids,
            past_key_values=None,
            inputs_embeds=None,
            cache_position=None,
            position_ids=None,
            pixel_values=None,
            attention_mask=None,
            token_type_ids=None,
            use_cache=True,
            **kwargs,
        ):
        # set position inds here: we are going to use absolute position embeddings, hence carefully track the locs of the past position embedding indices
        if attention_mask is not None and position_ids is None:
            if past_key_values:
                position_ids_mask = (input_ids != self.config.image_token_index)
                position_ids_mask[:, :-1] &= (input_ids[:, :-1] != self.config.text_config.pad_token_id)
                last_index = position_ids_mask.sum(dim=1) - 1
                position_ids = torch.stack([torch.arange(start, start+cache_position.shape[0], device=input_ids.device) for start in last_index])

        # If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
        # Exception 1: when passing input_embeds, input_ids may be missing entries
        # Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
        if past_key_values is not None:
            if inputs_embeds is not None:  # Exception 1
                input_ids = input_ids[:, -cache_position.shape[0] :]
            elif input_ids.shape[1] != cache_position.shape[0]:  # Default case (the "else", a no op, is Exception 2)
                input_ids = input_ids[:, cache_position]

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and cache_position[0] == 0:
            model_inputs = {"inputs_embeds": inputs_embeds}
        else:
            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases

        model_inputs.update(
            {
                "position_ids": position_ids,
                "past_key_values": past_key_values,
                "cache_position": cache_position,
                "use_cache": use_cache,
                "attention_mask": attention_mask,
                "pixel_values": pixel_values,
                "token_type_ids": token_type_ids,
            }
        )
        return model_inputs