Base model registered and uploaded

README.md

@@ -0,0 +1,199 @@
+---
+library_name: transformers
+tags: []
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]

config.json

@@ -0,0 +1,46 @@
+{
+  "_name_or_path": "ucsahin/TRaVisionLM-base-final-unregistered",
+  "architectures": [
+    "TraVisionForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_travisionlm.TraVisionLMConfig",
+    "AutoModelForCausalLM": "modeling_travisionlm.TraVisionForCausalLM"
+  },
+  "hidden_size": 1280,
+  "ignore_index": -100,
+  "image_token_index": 50257,
+  "model_type": "travisionlm",
+  "num_image_tokens": 256,
+  "projection_dim": 768,
+  "text_config": {
+    "architectures": [
+      "GPT2LMHeadModel"
+    ],
+    "bos_token_id": 0,
+    "eos_token_id": 0,
+    "model_type": "gpt2",
+    "n_ctx": 1024,
+    "n_embd": 1280,
+    "n_head": 20,
+    "n_layer": 36,
+    "pad_token_id": 0,
+    "reorder_and_upcast_attn": true,
+    "scale_attn_by_inverse_layer_idx": true,
+    "task_specific_params": {
+      "text-generation": {
+        "do_sample": true,
+        "max_length": 50
+      }
+    },
+    "torch_dtype": "float32",
+    "vocab_size": 51282
+  },
+  "torch_dtype": "float32",
+  "transformers_version": "4.44.0.dev0",
+  "vision_config": {
+    "image_size": 256,
+    "model_type": "siglip_vision_model",
+    "projection_dim": 768
+  }
+}

configuration_travisionlm.py

@@ -0,0 +1,94 @@
+"""TraVisionLM configuration"""
+
+from transformers import PretrainedConfig
+from transformers import logging, CONFIG_MAPPING
+import warnings
+
+logger = logging.get_logger(__name__)
+
+class TraVisionLMConfig(PretrainedConfig):
+    model_type = "travisionlm"
+    is_composition = False
+
+    def __init__(
+            self,
+            vision_config=None,
+            text_config=None,
+            ignore_index=-100,
+            image_token_idx=50257,
+            vocab_size=51282,
+            projection_dim=768,
+            hidden_size=1280,
+            **kwargs, 
+    ):
+        self.ignore_index = ignore_index
+        self.image_token_index = image_token_idx
+        self._vocab_size = vocab_size
+        self.projection_dim = projection_dim
+        self.hidden_size = hidden_size
+        self.vision_config = vision_config
+        self.is_encoder_decoder = False
+        if isinstance(self.vision_config, dict):
+            vision_config["model_type"] = (
+                vision_config["model_type"] if "model_type" in vision_config else "siglip_vision_model"
+            )
+            self.vision_config = CONFIG_MAPPING[vision_config["model_type"]](**vision_config)
+        elif vision_config is None:
+            self.vision_config = CONFIG_MAPPING["siglip_vision_model"](
+                attention_dropout=0.0,
+                hidden_act="gelu_pytorch_tanh",
+                hidden_size=768,
+                image_size=256,
+                intermediate_size=3072,
+                layer_norm_eps=1e-06,
+                num_attention_heads=12,
+                num_channels=3,
+                num_hidden_layers=12,
+                patch_size=16,
+            )
+        self.vocab_size = vocab_size
+
+        self.text_config = text_config
+
+        if isinstance(self.text_config, dict):
+            text_config["model_type"] = text_config["model_type"] if "model_type" in text_config else "gpt2"
+            self.text_config = CONFIG_MAPPING[text_config["model_type"]](**text_config)
+        elif text_config is None:
+            self.text_config = CONFIG_MAPPING["gpt2"](
+                activation_function="gelu_new",
+                attn_pdrop=0.1,
+                embd_pdrop=0.1,
+                initializer_range=0.02,
+                layer_norm_epsilon=1e-05,
+                n_ctx=1024,
+                n_embd=1280,
+                n_head=20,
+                n_layer=36,
+                n_positions=1024,
+                reorder_and_upcast_attn=False,
+                resid_pdrop=0.1,
+                scale_attn_by_inverse_layer_idx=False,
+                scale_attn_weights=True,
+                vocab_size=vocab_size
+            )
+        self.num_image_tokens = (self.vision_config.image_size // self.vision_config.patch_size) ** 2
+        self.pad_token_id = self.text_config.pad_token_id
+        self.vision_config.projection_dim = projection_dim
+        super().__init__(**kwargs)
+
+    @property
+    def vocab_size(self):
+        warnings.warn(
+            "The `vocab_size` attribute is deprecated and will be removed in v4.44, Please use `text_config.vocab_size` instead.",
+            FutureWarning,
+        )
+        return self._vocab_size
+
+    @vocab_size.setter
+    def vocab_size(self, value):
+        self._vocab_size = value
+
+    def to_dict(self):
+        output = super().to_dict()
+        output.pop("_vocab_size", None)
+        return output

generation_config.json

@@ -0,0 +1,7 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 0,
+  "eos_token_id": 0,
+  "pad_token_id": 0,
+  "transformers_version": "4.44.0.dev0"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:84db9d17f7b2db7e2fa107682bf1a214f1d6a2615c2b0b8dd295db0d0d721ff8
+size 3498353344

modeling_travisionlm.py

@@ -0,0 +1,520 @@
+"""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