add modeling_magma

modeling_magma.py

@@ -0,0 +1,1460 @@
+# coding=utf-8
+# Copyright 2024 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.
+"""PyTorch Magma model."""
+
+import math
+import re
+import os
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.utils.checkpoint
+from torch import nn
+import wandb
+import torch.distributed as dist
+from transformers.modeling_utils import PreTrainedModel
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.utils import ModelOutput
+from transformers.utils import (
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from transformers import AutoConfig, AutoModelForCausalLM
+from .configuration_magma import MagmaConfig
+from .image_tower_magma import MagmaImageTower
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "MagmaConfig"
+    
+@dataclass
+# Copied from transformers.models.idefics.modeling_idefics.IdeficsCausalLMOutputWithPast with Idefics->Magma
+class MagmaCausalLMOutputWithPast(ModelOutput):
+    """
+    Base class for Magma causal 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[List[torch.FloatTensor]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+    image_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+
+
+class MagmaMultiModalProjector(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        
+        dim_vision = {'base': 640, 'large': 768, 'xxlarge': 1024}
+        vision_backbone = config.get('vision_backbone', 'convnextxxlarge')
+        vision_backbone_size = vision_backbone.replace('convnext', '')
+        projector_type = config.get('mm_projector_type', 'linear')
+        mlp_gelu_match = re.match(r'^mlp(\d+)x_gelu_segtokv(\d+)$', projector_type)
+        if mlp_gelu_match:
+            mlp_depth = int(mlp_gelu_match.group(1))
+            modules = [nn.Linear(config['mm_hidden_size'], config['hidden_size'])]
+            for _ in range(1, mlp_depth):
+                modules.append(nn.GELU())
+                modules.append(nn.Linear(config['hidden_size'], config['hidden_size']))
+            self.proj = nn.Sequential(*modules)
+
+        # define a row seperator
+        self.row_seperator = nn.Parameter(torch.zeros(1, 1, config['hidden_size']))
+        if config.get('mm_use_im_start_end', False):
+            self.img_start_seperator = nn.Parameter(torch.zeros(1, config['hidden_size']))
+            self.img_end_seperator = nn.Parameter(torch.zeros(1, config['hidden_size']))                        
+
+    def forward(self, x):
+        return self.proj(x)
+
+
+MAGMA_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 ([`MagmaConfig`] or [`MagmaVisionConfig`]):
+            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 LLaMA Model outputting raw hidden-states without any specific head on top.",
+    MAGMA_START_DOCSTRING,
+)
+
+class MagmaPreTrainedModel(PreTrainedModel):
+    config_class = MagmaConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["MagmaVisionAttention"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
+
+    def _init_weights(self, module):
+        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
+
+
+MAGMA_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)):
+            The tensors corresponding to the input images. Pixel values can be obtained using
+            [`AutoImageProcessor`]. See [`MagmaImageProcessor.__call__`] for details. [`MagmaProcessor`] uses
+            [`MagmaImageProcessor`] for processing images.
+        image_sizes (`torch.LongTensor` of shape `(batch_size, 2)`, *optional*):
+            The sizes of the images in the batch, being (height, width) for each image.
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids)
+        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)`) and 2 additional tensors of shape
+            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        vision_feature_layer (`int`, *optional*, defaults to -2):
+            The index of the layer to select the vision feature.
+        vision_feature_select_strategy (`str`, *optional*, defaults to `"default"`):
+            The feature selection strategy used to select the vision feature from the vision backbone.
+            Can be one of `"default"` or `"full"`. If `"default"`, the CLS token is removed from the vision features.
+            If `"full"`, the full vision features are used.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+@add_start_docstrings(
+    """The Magma model which consists of a vision backbone and a language model.""",
+    MAGMA_START_DOCSTRING,
+)
+class MagmaForForCausalLM(MagmaPreTrainedModel):
+    def __init__(self, config: MagmaConfig):
+        super().__init__(config)
+
+        self.vision_tower = MagmaImageTower(config.vision_config, require_pretrained=False)
+        config.vision_config['mm_hidden_size'] = config.vision_config['mm_hidden_size'] \
+            if 'mm_hidden_size' in config.vision_config else self.vision_tower.hidden_size
+        config.vision_config['hidden_size'] = config.vision_config['hidden_size'] \
+            if 'hidden_size' in config.vision_config else self.config.text_config.hidden_size
+        self.multi_modal_projector = MagmaMultiModalProjector(config.vision_config)
+
+        self.vocab_size = config.text_config.vocab_size
+        if hasattr(config.text_config, 'auto_map'):
+            del config.text_config.auto_map
+
+        try:
+            self.language_model = AutoModelForCausalLM.from_config(
+                config.text_config, 
+                # attn_implementation=config._attn_implementation, 
+                trust_remote_code=True
+            )
+        except:
+            self.language_model = AutoModelForCausalLM.from_pretrained(
+                config.text_config._name_or_path, 
+                # attn_implementation=config._attn_implementation, 
+                trust_remote_code=True
+            )
+        
+        self.pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1
+        self._padding_side = "left"  # set it to left by default, user can use setter to change padding_sides
+
+        try:
+            if dist.get_rank() == 0:
+                wandb.init(project=os.environ['WANDB_PROJECT'])
+        except:
+            pass
+
+        self.post_init()
+    
+    # def from_pretrained(self, pretrained_model_name_or_path, *model_args, **kwargs):
+    #     import pdb; pdb.set_trace()
+    #     kwargs["_from_auto"] = True
+    #     return super().from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+
+    @property
+    def padding_side(self):
+        return self._padding_side
+
+    @padding_side.setter
+    def padding_side(self, padding_side: str):
+        if padding_side not in ["left", "right"]:
+            raise ValueError(f"{padding_side} is not `left` or `right`.")
+        self._padding_side = padding_side
+
+    def get_input_embeddings(self):
+        return self.language_model.get_input_embeddings()
+
+    def set_input_embeddings(self, value):
+        self.language_model.set_input_embeddings(value)
+
+    def get_output_embeddings(self):
+        return self.language_model.get_output_embeddings()
+
+    def set_output_embeddings(self, new_embeddings):
+        self.language_model.set_output_embeddings(new_embeddings)
+
+    def set_decoder(self, decoder):
+        self.language_model.set_decoder(decoder)
+
+    def get_decoder(self):
+        return self.language_model.get_decoder()
+
+    def tie_weights(self):
+        return self.language_model.tie_weights()
+
+    def load_special_module_from_ckpt(self, ckpt_path, torch_dtype=None):
+        from deepspeed.runtime.zero import Init
+        from deepspeed import zero
+        # Defer initialization for ZeRO-3 compatibility
+        # with Init(data_parallel_group=None):
+        #     # Initialize the special module
+        #     self.vision_tower = MagmaImageTower(self.config.vision_config, require_pretrained=False)
+
+        # Load checkpoint weights into the special module
+        checkpoint = torch.load(ckpt_path, map_location='cpu')
+        state_dict = {k.replace('visual.', ''): v for k, v in checkpoint.items() if 'visual.' in k}
+
+        # Convert checkpoint weights to match model's parameter dtype
+        if torch_dtype is None:
+            model_dtype = next(self.vision_tower.clip_vision_model.parameters()).dtype
+            for k, v in state_dict.items():
+                state_dict[k] = v.to(model_dtype)
+        else:
+            for k, v in state_dict.items():
+                state_dict[k] = v.to(torch_dtype)
+
+        # Temporarily gather parameters for loading (if ZeRO-3 is active)
+        with zero.GatheredParameters(list(self.vision_tower.parameters()), modifier_rank=0):
+            # Load the state dictionary
+            self.vision_tower.clip_vision_model.load_state_dict(state_dict, strict=False)
+            # After loading, ensure the module is on the correct device
+            for param in self.vision_tower.parameters():
+                param.data = param.data.to(self.device).to(torch_dtype)
+
+        # import pdb; pdb.set_trace()
+        # If using a DeepSpeed engine, attach the updated module
+        if hasattr(self, "deepspeed_engine"):
+            self.deepspeed_engine.module = self
+        
+    def resize_token_embeddings(self, new_num_tokens: Optional[int] = None, pad_to_multiple_of=None) -> nn.Embedding:
+        model_embeds = self.language_model.resize_token_embeddings(new_num_tokens, pad_to_multiple_of)
+        # update vocab size
+        self.config.text_config.vocab_size = model_embeds.num_embeddings
+        self.vocab_size = model_embeds.num_embeddings
+        return model_embeds
+
+    def _merge_input_ids_with_image_features(
+        self,
+        image_features,
+        feature_lens,
+        inputs_embeds,
+        input_ids,
+        attention_mask,
+        position_ids=None,
+        labels=None,
+        image_token_index=None,
+        ignore_index=-100,
+    ):
+        """
+        Merge input_ids with with image features into final embeddings
+
+        Args:
+            image_features (`torch.Tensor` of shape `(all_feature_lens, embed_dim)`):
+                All vision vectors of all images in the batch
+            feature_lens (`torch.LongTensor` of shape `(num_images)`):
+                The length of visual embeddings of each image as stacked in `image_features`
+            inputs_embeds (`torch.Tensor` of shape `(batch_size, sequence_length, embed_dim)`):
+                Token embeddings before merging with visual embeddings
+            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+                Input_ids of tokens, possibly filled with image token
+            attention_mask (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+                Mask to avoid performing attention on padding token indices.
+            position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+                Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+                config.n_positions - 1]`.
+            labels (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*)
+                :abels need to be recalculated to support training (if provided)
+            image_token_index (`int`, *optional*)
+                Token id used to indicate the special "image" token. Defaults to `config.image_token_index`
+            ignore_index (`int`, *optional*)
+                Value that is used to pad `labels` and will be ignored when calculated loss. Default: -100.
+        Returns:
+            final_embedding, final_attention_mask, position_ids, final_labels
+
+        Explanation:
+            each image has variable length embeddings, with length specified by feature_lens
+            image_features is concatenation of all visual embed vectors
+            task: fill each <image> with the correct number of visual embeddings
+            Example:
+                X (5 patches), Y (3 patches), Z (8)
+                X, Y are in the same sequence (in-context learning)
+            if right padding
+                input_ids: [
+                    a b c d e f X g h i j k Y l m
+                    o p q r Z s t u v _ _ _ _ _ _
+                ]
+                input_ids should be: [
+                    a b c d e f X X X X X g h i j k Y Y Y l m
+                    o p q r Z Z Z Z Z Z Z Z s t u v _ _ _ _ _
+                ]
+                labels should be: [
+                    a b c d e f _ _ _ _ _ g h i j k _ _ _ l m
+                    o p q r _ _ _ _ _ _ _ _ s t u v _ _ _ _ _
+                ]
+            elif left padding
+                input_ids: [
+                    a b c d e f X g h i j k Y l m
+                    _ _ _ _ _ _ o p q r Z s t u v
+                ]
+                input_ids should be: [
+                    a b c d e f X X X X X g h i j k Y Y Y l m
+                    _ _ _ _ _ o p q r Z Z Z Z Z Z Z Z s t u v
+                ]
+                labels should be: [
+                    a b c d e f _ _ _ _ _ g h i j k _ _ _ l m
+                    _ _ _ _ _ o p q r _ _ _ _ _ _ _ _ s t u v
+                ]
+            Edge cases:
+                * If tokens are same but image token sizes are different, then cannot infer left or right padding
+
+                input_ids: [
+                    a b c d X g h
+                    i j Y k l m n
+                ]
+                where X is 3 tokens while Y is 5, this mean after merge
+                if left-padding (batched generation)
+                    input_ids should be: [
+                        _ _ a b c d X X X g h
+                        i j Y Y Y Y Y k l m n
+                    ]
+                elif (right padding) (training)
+                    input_ids should be: [
+                        a b c d X X X g h _ _
+                        i j Y Y Y Y Y k l m n
+                    ]
+        """
+        image_token_index = image_token_index if image_token_index is not None else self.config.image_token_index
+        ignore_index = ignore_index if ignore_index is not None else self.config.ignore_index
+
+        with torch.no_grad():
+            num_images = feature_lens.size(0)
+            num_image_features, embed_dim = image_features.shape
+            if feature_lens.sum() != num_image_features:
+                raise ValueError(f"{feature_lens=} / {feature_lens.sum()} != {image_features.shape=}")
+            batch_size = input_ids.shape[0]
+            _left_padding = torch.any(attention_mask[:, 0] == 0)
+            _right_padding = torch.any(attention_mask[:, -1] == 0)
+
+            left_padding = True
+            if batch_size > 1:
+                if _left_padding and not _right_padding:
+                    left_padding = True
+                elif not _left_padding and _right_padding:
+                    left_padding = False
+                elif not _left_padding and not _right_padding:
+                    # both side is 1, so cannot tell
+                    left_padding = self.padding_side == "left"
+                else:
+                    # invalid attention_mask
+                    raise ValueError(f"both side of attention_mask has zero, invalid. {attention_mask}")
+
+            # Whether to turn off right padding
+            # 1. Create a mask to know where special image tokens are
+            special_image_token_mask = input_ids == image_token_index
+            # special_image_token_mask: [bsz, seqlen]
+            num_special_image_tokens = torch.sum(special_image_token_mask, dim=-1)
+            # num_special_image_tokens: [bsz]
+            # Reserve for padding of num_images
+            total_num_special_image_tokens = torch.sum(special_image_token_mask)
+            if total_num_special_image_tokens != num_images:
+                raise ValueError(
+                    f"Number of image tokens in input_ids ({total_num_special_image_tokens}) different from num_images ({num_images})."
+                )
+            # Compute the maximum embed dimension
+            # max_image_feature_lens is max_feature_lens per batch
+            feature_lens_batch = feature_lens.split(num_special_image_tokens.tolist(), dim=0)
+            feature_lens_batch_sum = torch.tensor([x.sum() for x in feature_lens_batch], device=feature_lens.device)
+            embed_sequence_lengths = (
+                (attention_mask == 1).long().sum(-1) - num_special_image_tokens + feature_lens_batch_sum
+            )
+            max_embed_dim = embed_sequence_lengths.max()
+
+            batch_indices, non_image_indices = torch.where((input_ids != image_token_index) & (attention_mask == 1))
+            # 2. Compute the positions where text should be written
+            # Calculate new positions for text tokens in merged image-text sequence.
+            # `special_image_token_mask` identifies image tokens. Each image token will be replaced by `nb_text_tokens_per_images` text tokens.
+            # `torch.cumsum` computes how each image token shifts subsequent text token positions.
+            # - 1 to adjust for zero-based indexing, as `cumsum` inherently increases indices by one.
+            # ! instead of special_image_token_mask * (num_image_patches - 1)
+            #   special_image_token_mask * (num_feature_len - 1)
+            special_image_token_mask = special_image_token_mask.long()
+            special_image_token_mask[special_image_token_mask == 1] = feature_lens - 1
+            new_token_positions = torch.cumsum((special_image_token_mask + 1), -1) - 1
+            if left_padding:
+                # shift right token positions so that they are ending at the same number
+                # the below here was incorrect? new_token_positions += new_token_positions[:, -1].max() - new_token_positions[:, -1:]
+                new_token_positions += max_embed_dim - 1 - new_token_positions[:, -1:]
+
+            text_to_overwrite = new_token_positions[batch_indices, non_image_indices]
+
+        # 3. Create the full embedding, already padded to the maximum position
+        final_embedding = torch.zeros(
+            batch_size, max_embed_dim, embed_dim, dtype=inputs_embeds.dtype, device=inputs_embeds.device
+        )
+        final_attention_mask = torch.zeros(
+            batch_size, max_embed_dim, dtype=attention_mask.dtype, device=inputs_embeds.device
+        )
+        final_labels = None
+        if labels is not None:
+            # NOTE: this is a bug in the original code!!!
+            final_labels = torch.full_like(final_attention_mask.long(), ignore_index).to(torch.long)
+        # In case the Vision model or the Language model has been offloaded to CPU, we need to manually
+        # set the corresponding tensors into their correct target device.
+        target_device = inputs_embeds.device
+        batch_indices, non_image_indices, text_to_overwrite = (
+            batch_indices.to(target_device),
+            non_image_indices.to(target_device),
+            text_to_overwrite.to(target_device),
+        )
+        attention_mask = attention_mask.to(target_device)
+
+        # 4. Fill the embeddings based on the mask. If we have ["hey" "<image>", "how", "are"]
+        # we need to index copy on [0, 577, 578, 579] for the text and [1:576] for the image features
+        final_embedding[batch_indices, text_to_overwrite] = inputs_embeds[batch_indices, non_image_indices]
+        final_attention_mask[batch_indices, text_to_overwrite] = attention_mask[batch_indices, non_image_indices]
+        if labels is not None:
+            final_labels[batch_indices, text_to_overwrite] = labels[batch_indices, non_image_indices]
+
+        # 5. Fill the embeddings corresponding to the images. Anything that is not `text_positions` needs filling (#29835)
+        with torch.no_grad():
+            image_to_overwrite = torch.full(
+                (batch_size, max_embed_dim), True, dtype=torch.bool, device=inputs_embeds.device
+            )
+            image_to_overwrite[batch_indices, text_to_overwrite] = False
+            embed_indices = torch.arange(max_embed_dim).unsqueeze(0).to(target_device)
+            embed_indices = embed_indices.expand(batch_size, max_embed_dim)
+            embed_seq_lens = embed_sequence_lengths[:, None].to(target_device)
+
+            if left_padding:
+                # exclude padding on the left
+                val = (max_embed_dim - embed_indices) <= embed_seq_lens
+            else:
+                # exclude padding on the right
+                val = embed_indices < embed_seq_lens
+            image_to_overwrite &= val
+
+            if image_to_overwrite.sum() != num_image_features:
+                raise ValueError(
+                    f"{image_to_overwrite.sum()=} != {num_image_features=} The input provided to the model are wrong. "
+                    f"The number of image tokens is {torch.sum(special_image_token_mask)} while"
+                    f" the number of image given to the model is {num_images}. "
+                    f"This prevents correct indexing and breaks batch generation."
+                )
+        final_embedding[image_to_overwrite] = image_features.contiguous().reshape(-1, embed_dim).to(target_device)
+        final_attention_mask |= image_to_overwrite
+        position_ids = (final_attention_mask.cumsum(-1) - 1).masked_fill_((final_attention_mask == 0), 1)
+
+        return final_embedding, final_attention_mask, position_ids, final_labels
+
+    @add_start_docstrings_to_model_forward(MAGMA_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=MagmaCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        pixel_values: Union[torch.FloatTensor, List[torch.FloatTensor], List[List[torch.FloatTensor]]] = None,
+        image_sizes: Union[torch.LongTensor, List[torch.LongTensor], List[List[torch.LongTensor]]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        vision_feature_layer: Optional[int] = None,
+        vision_feature_select_strategy: Optional[str] = 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, MagmaCausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, MagmaForConditionalGeneration
+
+        >>> model = MagmaForConditionalGeneration.from_pretrained("microsoft/magma-8b-hf")
+        >>> processor = AutoProcessor.from_pretrained("microsoft/magma-8b-hf")
+
+        >>> prompt = "[INST] <image>\nWhat is shown in this image? [/INST]"
+        >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(text=prompt, images=image, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(**inputs, max_length=30)
+        >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "[INST]  \nWhat is shown in this image? [/INST] The image appears to be a radar chart, which is a type of multi-dimensional plot (...)"
+        ```"""
+        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
+        vision_feature_layer = (
+            vision_feature_layer if vision_feature_layer is not None else self.config.vision_config['vision_feature_layer']
+        )
+        
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        
+        if inputs_embeds is None:
+            # 1. Extract the input embeddings
+            # In case image_token_index is not in the embeddings (extra token but embedding don't have it)
+            for_inputs_embeds_ids = input_ids.clone()
+            for_inputs_embeds_ids[(input_ids == self.config.image_token_index)] = 0
+            inputs_embeds = self.get_input_embeddings()(for_inputs_embeds_ids)
+
+            # 2. Merge text and images
+            if pixel_values is not None and input_ids.shape[1] != 1 and len(pixel_values) > 0:
+                # ! infer image_num_patches from image_sizes
+                if type(pixel_values) == list:
+                    # nested list of pixel_values, each element is a list of pixel_values for each training instance, it could be multiple for video or interleaved setting
+                    # e.g., pixel_values = [[img1, img2], [img1, img2, img3]]
+                    n_imgs_per_sample = [len(pv) for pv in pixel_values]
+                    pixels_values_list = sum(pixel_values, [])
+                    image_sizes_list = sum(image_sizes, [])
+                else:
+                    image_num_patches = [(imsize[imsize.sum(1) > 0,0] * imsize[imsize.sum(1) > 0,1]).tolist() for imsize in image_sizes]       
+                    # image_num_patches = [(imsize[:,0]*imsize[:,1]).tolist() for imsize in image_sizes]             
+                    # figure out if pixel_values is concatenated or stacked
+                    if pixel_values.dim() == 5:
+                        # stacking when input is (batch_size, num_patches, num_channels, height, width)
+                        _pixel_values_list = [
+                            pix_val[:sum(num_patch)].split(num_patch, dim=0) for pix_val, num_patch in zip(pixel_values, image_num_patches)
+                        ]
+                        _image_sizes_list = [image_size[image_size.sum(-1) > 0].tolist() for image_size in image_sizes]
+                    elif pixel_values.dim() != 4:
+                        # otherwise has to be stacked from list of (num_patches, num_channels, height, width)
+                        raise ValueError(f"pixel_values of shape {pixel_values.shape}, expect to be of 4 or 5 dimensions")
+
+                if self.config.vision_config['img_anyres_strategy'] == "global":
+                    selected_image_features = []
+                    # NOTE: both _image_sizes_list and _pixel_values_list are lists of lists, each item represents an training instance with one or multiple images
+                    for idx, (image_size_for_instance, pixel_values_for_instance) in enumerate(zip(_image_sizes_list, _pixel_values_list)):
+                        assert len(image_size_for_instance) == len(pixel_values_for_instance), f"{len(image_size_for_instance)} != {len(pixel_values_for_instance)}"
+                        for image_size, pixel_values_for_image in zip(image_size_for_instance, pixel_values_for_instance):
+                            pixel_values_for_image = pixel_values_for_image.view(image_size[0], image_size[1], *pixel_values_for_image.shape[1:])
+                            pixel_values_for_image = pixel_values_for_image.permute(2, 0, 3, 1, 4).flatten(3, 4).flatten(1, 2).unsqueeze(0)
+                            image_features = self.vision_tower(pixel_values_for_image)
+                            selected_image_feature = image_features[vision_feature_layer][0].permute(1, 2, 0)
+                            selected_image_feature = self.multi_modal_projector((selected_image_feature, None))
+                            selected_image_feature = torch.cat((selected_image_feature, self.multi_modal_projector.row_seperator.repeat(selected_image_feature.shape[0],1,1)), dim=1)
+                            selected_image_features.append(selected_image_feature.flatten(0, 1))
+                elif self.config.vision_config['img_anyres_strategy'] == "crop":
+                    # calculate number of crops for each instance in the batch given _image_sizes_list
+                    _image_sizes_list_temp = sum(_image_sizes_list, [])
+                    # concate nate all images in _pixel_values_list
+                    _pixel_values_list_temp = sum(_pixel_values_list, ())
+                    _pixel_values_list_temp = torch.cat(_pixel_values_list_temp, dim=0)
+                    image_features = self.vision_tower(_pixel_values_list_temp)[vision_feature_layer].permute(0, 2, 3, 1)
+                    image_features = self.multi_modal_projector((image_features, None))
+
+                    num_crops_list = [_image_size[0]*_image_size[1] for _image_size in _image_sizes_list_temp]
+                    image_features_split = torch.split(image_features, num_crops_list, dim=0)
+                    selected_image_features = []
+                    for image_feature, image_size in zip(image_features_split, _image_sizes_list_temp):
+                        image_feature = image_feature.view(image_size[0], image_size[1], *image_feature.shape[1:])
+                        image_feature = image_feature.permute(0, 2, 1, 3, 4).flatten(2, 3).flatten(0, 1)
+                        image_feature = torch.cat((image_feature, self.multi_modal_projector.row_seperator.repeat(image_feature.shape[0],1,1)), dim=1)
+                        selected_image_features.append(image_feature.flatten(0, 1))
+
+                    # raise NotImplementedError("crop strategy is not implemented yet")
+                    # image_features = self.vision_tower(pixel_values)
+                    # selected_image_feature = image_features[vision_feature_layer]            
+                    # image_features = torch.split(image_features, image_num_patches, dim=0)
+
+                # NOTE we only support multimodal_patch_merge_type == "spatial_unpad"
+                feature_lens = [elem.shape[0] for elem in selected_image_features]        
+                image_features = torch.cat(selected_image_features, 0)
+                feature_lens = torch.tensor(feature_lens, dtype=torch.long, device=image_features.device)
+                
+                # inputs_embeds = inputs_embeds.to(image_features.dtype)
+                inputs_embeds, attention_mask, position_ids, labels = self._merge_input_ids_with_image_features(
+                    image_features,
+                    feature_lens,
+                    inputs_embeds,
+                    input_ids,
+                    attention_mask,
+                    position_ids,
+                    labels=labels,
+                )
+
+            # pixel_values is not None but is empty ---> text only cases
+            elif pixel_values is not None and input_ids.shape[1] != 1 and pixel_values.size(0) == 0:
+                # there are no images
+                pass
+
+            # In case input_ids.shape[1] == 1 & pixel_values==None & past_key_values != None, we are in the case of
+            # generation with cache
+            elif past_key_values is not None and pixel_values is not None and input_ids.shape[1] == 1:
+                # Retrieve the first layer to inspect the logits and mask out the hidden states
+                # that are set to 0
+                first_layer_past_key_value = past_key_values[0][0][:, :, :, 0]
+
+                # Sum all dimensions of head_dim (-2) to avoid random errors such as: https://github.com/huggingface/transformers/pull/28032#issuecomment-1863691941
+                batch_index, non_attended_tokens = torch.where(first_layer_past_key_value.float().sum(-2) == 0)
+
+                # Get the target length
+                target_length = input_ids.shape[1]
+                past_length = first_layer_past_key_value.shape[-1]
+
+                extended_attention_mask = torch.ones(
+                    (attention_mask.shape[0], past_length),
+                    dtype=attention_mask.dtype,
+                    device=attention_mask.device,
+                )
+
+                # Filter out only the tokens that can be un-attended, this can happen
+                # if one uses Llava + Fused modules where the cache on the
+                # first iteration is already big enough, or if one passes custom cache
+                valid_indices = non_attended_tokens < extended_attention_mask.size(-1)
+                new_batch_index = batch_index[valid_indices]
+                new_non_attended_tokens = non_attended_tokens[valid_indices]
+
+                # Zero-out the places where we don't need to attend
+                extended_attention_mask[new_batch_index, new_non_attended_tokens] = 0
+
+                attention_mask = torch.cat((extended_attention_mask, attention_mask[:, -target_length:]), dim=1)
+
+                position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1
+        
+        # outputs = self.language_model(
+        #     attention_mask=attention_mask,
+        #     position_ids=position_ids,
+        #     past_key_values=past_key_values,
+        #     inputs_embeds=inputs_embeds,
+        #     use_cache=use_cache,
+        #     output_attentions=output_attentions,
+        #     output_hidden_states=output_hidden_states,
+        #     return_dict=return_dict,
+        # )
+
+        # logits = outputs[0]
+        # loss = None
+        # if labels is not None:
+        #     # Shift so that tokens < n predict n
+        #     if attention_mask is not None:
+        #         shift_attention_mask = attention_mask[..., 1:]
+        #         shift_logits = logits[..., :-1, :][shift_attention_mask.to(logits.device) != 0].contiguous()
+        #         shift_labels = labels[..., 1:][shift_attention_mask.to(labels.device) != 0].contiguous()
+        #     else:
+        #         shift_logits = logits[..., :-1, :].contiguous()
+        #         shift_labels = labels[..., 1:].contiguous()
+        #     # Flatten the tokens
+        #     loss_fct = nn.CrossEntropyLoss()
+        #     loss = loss_fct(
+        #         shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1).to(shift_logits.device)
+        #     )
+
+        outputs = self.language_model.model(
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict
+        )
+
+        hidden_states = outputs[0]
+
+        loss = None
+
+        if labels is not None and self.training:
+            valid_mask = labels[..., 1:] != -100
+            shift_logits = self.language_model.lm_head(hidden_states[:,:-1][valid_mask]).contiguous()
+            shift_logits = shift_logits.view(-1, self.language_model.config.vocab_size)
+            logits = shift_logits # dummy logits
+            shift_labels = labels[..., 1:][valid_mask].contiguous()
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss_fct = nn.CrossEntropyLoss()
+            loss = loss_fct(shift_logits, shift_labels)
+
+            # localize the positions for shift_labels where the id is in betweek [config.tokenizer_vocab_size-256, config.tokenizer_vocab_size]
+            valid_indices = (shift_labels<self.config.tokenizer_vocab_size) & (shift_labels>=self.config.tokenizer_vocab_size-256)
+            if valid_indices.sum() > 0:
+                action_labels = shift_labels[valid_indices]
+                action_logits = shift_logits[valid_indices]
+                # calcualte the accuracy
+                action_accuracy = (action_logits.argmax(-1) == action_labels).float().mean()
+                # log the action accuracy
+            else:
+                action_accuracy = torch.tensor(0.0).to(shift_logits.device)
+            # torch distributed gather the action accuracy across all devices     
+            action_accuracy = action_accuracy.unsqueeze(0)
+            # gather the action accuracy across all devices
+            action_accuracy_gather = [torch.zeros_like(action_accuracy) for _ in range(dist.get_world_size())]
+            dist.all_gather(action_accuracy_gather, action_accuracy)
+            # concatenate the action accuracy across all devices
+            action_accuracy = torch.cat(action_accuracy_gather)            
+            
+            if dist.get_rank() == 0:            
+                # remove zero values
+                if action_accuracy.mean() == 0:
+                    wandb.log({"action_accuracy": action_accuracy.mean().item()})
+                else:
+                    action_accuracy = action_accuracy[action_accuracy != 0]
+                    wandb.log({"action_accuracy": action_accuracy.mean().item()})
+        else:
+            logits = self.language_model.lm_head(hidden_states)
+            logits = logits.float()
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return MagmaCausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        inputs_embeds=None,
+        pixel_values=None,
+        image_sizes=None,
+        attention_mask=None,
+        **kwargs,
+    ):
+        if past_key_values is not None:
+            if isinstance(past_key_values, Cache):
+                cache_length = past_key_values.get_seq_length()
+                past_length = past_key_values.seen_tokens
+            else:
+                cache_length = past_length = past_key_values[0][0].shape[2]
+
+            # Keep only the unprocessed tokens:
+            # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
+            # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
+            # input)
+            if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
+                input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
+            # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
+            # input_ids based on the past_length.
+            elif past_length < input_ids.shape[1]:
+                input_ids = input_ids[:, past_length:]
+            # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
+            elif self.config.image_token_index in input_ids:
+                input_ids = input_ids[:, input_ids.shape[1] - 1 :]
+            # If the cache has seen more tokens than it can hold, then the cache has a size limit. Let's discard the
+            # older attention values, as their corresponding values are not part of the input.
+            if cache_length < past_length and attention_mask is not None:
+                attention_mask = attention_mask[:, -(cache_length + input_ids.shape[1]) :]
+
+        position_ids = kwargs.get("position_ids", None)
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -input_ids.shape[1] :]
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+                "pixel_values": pixel_values,
+                "image_sizes": image_sizes,
+            }
+        )
+        return model_inputs
+
+    def _reorder_cache(self, *args, **kwargs):
+        return self.language_model._reorder_cache(*args, **kwargs)
+
+@add_start_docstrings(
+    """The Magma model which consists of a vision backbone and a language model.""",
+    MAGMA_START_DOCSTRING,
+)
+class MagmaForConditionalGeneration(MagmaPreTrainedModel):
+    def __init__(self, config: MagmaConfig):
+        super().__init__(config)
+
+        self.vision_tower = MagmaImageTower(config.vision_config, require_pretrained=('magma' not in config.name_or_path))
+        self.multi_modal_projector = MagmaMultiModalProjector(config.vision_config)
+
+        self.vocab_size = config.text_config.vocab_size
+        self.language_model = AutoModelForCausalLM.from_config(
+            config.text_config, 
+            # attn_implementation=config._attn_implementation, 
+            trust_remote_code=True
+        )
+        
+        self.pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1
+        self._padding_side = "left"  # set it to left by default, user can use setter to change padding_sides
+
+        self.post_init()
+
+    @property
+    def padding_side(self):
+        return self._padding_side
+
+    @padding_side.setter
+    def padding_side(self, padding_side: str):
+        if padding_side not in ["left", "right"]:
+            raise ValueError(f"{padding_side} is not `left` or `right`.")
+        self._padding_side = padding_side
+
+    def get_input_embeddings(self):
+        return self.language_model.get_input_embeddings()
+
+    def set_input_embeddings(self, value):
+        self.language_model.set_input_embeddings(value)
+
+    def get_output_embeddings(self):
+        return self.language_model.get_output_embeddings()
+
+    def set_output_embeddings(self, new_embeddings):
+        self.language_model.set_output_embeddings(new_embeddings)
+
+    def set_decoder(self, decoder):
+        self.language_model.set_decoder(decoder)
+
+    def get_decoder(self):
+        return self.language_model.get_decoder()
+
+    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:
+        model_embeds = self.language_model.resize_token_embeddings(new_num_tokens, pad_to_multiple_of)
+        # update vocab size
+        self.config.text_config.vocab_size = model_embeds.num_embeddings
+        self.vocab_size = model_embeds.num_embeddings
+        return model_embeds
+
+    def _merge_input_ids_with_image_features(
+        self,
+        image_features,
+        feature_lens,
+        inputs_embeds,
+        input_ids,
+        attention_mask,
+        position_ids=None,
+        labels=None,
+        image_token_index=None,
+        ignore_index=-100,
+    ):
+        """
+        Merge input_ids with with image features into final embeddings
+
+        Args:
+            image_features (`torch.Tensor` of shape `(all_feature_lens, embed_dim)`):
+                All vision vectors of all images in the batch
+            feature_lens (`torch.LongTensor` of shape `(num_images)`):
+                The length of visual embeddings of each image as stacked in `image_features`
+            inputs_embeds (`torch.Tensor` of shape `(batch_size, sequence_length, embed_dim)`):
+                Token embeddings before merging with visual embeddings
+            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+                Input_ids of tokens, possibly filled with image token
+            attention_mask (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+                Mask to avoid performing attention on padding token indices.
+            position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+                Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+                config.n_positions - 1]`.
+            labels (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*)
+                :abels need to be recalculated to support training (if provided)
+            image_token_index (`int`, *optional*)
+                Token id used to indicate the special "image" token. Defaults to `config.image_token_index`
+            ignore_index (`int`, *optional*)
+                Value that is used to pad `labels` and will be ignored when calculated loss. Default: -100.
+        Returns:
+            final_embedding, final_attention_mask, position_ids, final_labels
+
+        Explanation:
+            each image has variable length embeddings, with length specified by feature_lens
+            image_features is concatenation of all visual embed vectors
+            task: fill each <image> with the correct number of visual embeddings
+            Example:
+                X (5 patches), Y (3 patches), Z (8)
+                X, Y are in the same sequence (in-context learning)
+            if right padding
+                input_ids: [
+                    a b c d e f X g h i j k Y l m
+                    o p q r Z s t u v _ _ _ _ _ _
+                ]
+                input_ids should be: [
+                    a b c d e f X X X X X g h i j k Y Y Y l m
+                    o p q r Z Z Z Z Z Z Z Z s t u v _ _ _ _ _
+                ]
+                labels should be: [
+                    a b c d e f _ _ _ _ _ g h i j k _ _ _ l m
+                    o p q r _ _ _ _ _ _ _ _ s t u v _ _ _ _ _
+                ]
+            elif left padding
+                input_ids: [
+                    a b c d e f X g h i j k Y l m
+                    _ _ _ _ _ _ o p q r Z s t u v
+                ]
+                input_ids should be: [
+                    a b c d e f X X X X X g h i j k Y Y Y l m
+                    _ _ _ _ _ o p q r Z Z Z Z Z Z Z Z s t u v
+                ]
+                labels should be: [
+                    a b c d e f _ _ _ _ _ g h i j k _ _ _ l m
+                    _ _ _ _ _ o p q r _ _ _ _ _ _ _ _ s t u v
+                ]
+            Edge cases:
+                * If tokens are same but image token sizes are different, then cannot infer left or right padding
+
+                input_ids: [
+                    a b c d X g h
+                    i j Y k l m n
+                ]
+                where X is 3 tokens while Y is 5, this mean after merge
+                if left-padding (batched generation)
+                    input_ids should be: [
+                        _ _ a b c d X X X g h
+                        i j Y Y Y Y Y k l m n
+                    ]
+                elif (right padding) (training)
+                    input_ids should be: [
+                        a b c d X X X g h _ _
+                        i j Y Y Y Y Y k l m n
+                    ]
+        """
+        image_token_index = image_token_index if image_token_index is not None else self.config.image_token_index
+        ignore_index = ignore_index if ignore_index is not None else self.config.ignore_index
+
+        with torch.no_grad():
+            num_images = feature_lens.size(0)
+            num_image_features, embed_dim = image_features.shape
+            if feature_lens.sum() != num_image_features:
+                raise ValueError(f"{feature_lens=} / {feature_lens.sum()} != {image_features.shape=}")
+            batch_size = input_ids.shape[0]
+            _left_padding = torch.any(attention_mask[:, 0] == 0)
+            _right_padding = torch.any(attention_mask[:, -1] == 0)
+
+            left_padding = True
+            if batch_size > 1:
+                if _left_padding and not _right_padding:
+                    left_padding = True
+                elif not _left_padding and _right_padding:
+                    left_padding = False
+                elif not _left_padding and not _right_padding:
+                    # both side is 1, so cannot tell
+                    left_padding = self.padding_side == "left"
+                else:
+                    # invalid attention_mask
+                    raise ValueError(f"both side of attention_mask has zero, invalid. {attention_mask}")
+
+            # Whether to turn off right padding
+            # 1. Create a mask to know where special image tokens are
+            special_image_token_mask = input_ids == image_token_index
+            # special_image_token_mask: [bsz, seqlen]
+            num_special_image_tokens = torch.sum(special_image_token_mask, dim=-1)
+            # num_special_image_tokens: [bsz]
+            # Reserve for padding of num_images
+            total_num_special_image_tokens = torch.sum(special_image_token_mask)
+            if total_num_special_image_tokens != num_images:
+                raise ValueError(
+                    f"Number of image tokens in input_ids ({total_num_special_image_tokens}) different from num_images ({num_images})."
+                )
+            # Compute the maximum embed dimension
+            # max_image_feature_lens is max_feature_lens per batch
+            feature_lens_batch = feature_lens.split(num_special_image_tokens.tolist(), dim=0)
+            feature_lens_batch_sum = torch.tensor([x.sum() for x in feature_lens_batch], device=feature_lens.device)
+            embed_sequence_lengths = (
+                (attention_mask == 1).long().sum(-1) - num_special_image_tokens + feature_lens_batch_sum
+            )
+            max_embed_dim = embed_sequence_lengths.max()
+
+            batch_indices, non_image_indices = torch.where((input_ids != image_token_index) & (attention_mask == 1))
+            # 2. Compute the positions where text should be written
+            # Calculate new positions for text tokens in merged image-text sequence.
+            # `special_image_token_mask` identifies image tokens. Each image token will be replaced by `nb_text_tokens_per_images` text tokens.
+            # `torch.cumsum` computes how each image token shifts subsequent text token positions.
+            # - 1 to adjust for zero-based indexing, as `cumsum` inherently increases indices by one.
+            # ! instead of special_image_token_mask * (num_image_patches - 1)
+            #   special_image_token_mask * (num_feature_len - 1)
+            special_image_token_mask = special_image_token_mask.long()
+            special_image_token_mask[special_image_token_mask == 1] = feature_lens - 1
+            new_token_positions = torch.cumsum((special_image_token_mask + 1), -1) - 1
+            if left_padding:
+                # shift right token positions so that they are ending at the same number
+                # the below here was incorrect? new_token_positions += new_token_positions[:, -1].max() - new_token_positions[:, -1:]
+                new_token_positions += max_embed_dim - 1 - new_token_positions[:, -1:]
+
+            text_to_overwrite = new_token_positions[batch_indices, non_image_indices]
+
+        # 3. Create the full embedding, already padded to the maximum position
+        final_embedding = torch.zeros(
+            batch_size, max_embed_dim, embed_dim, dtype=inputs_embeds.dtype, device=inputs_embeds.device
+        )
+        final_attention_mask = torch.zeros(
+            batch_size, max_embed_dim, dtype=attention_mask.dtype, device=inputs_embeds.device
+        )
+        final_labels = None
+        if labels is not None:
+            final_labels = torch.full_like(final_attention_mask, ignore_index).to(torch.long)
+        # In case the Vision model or the Language model has been offloaded to CPU, we need to manually
+        # set the corresponding tensors into their correct target device.
+        target_device = inputs_embeds.device
+        batch_indices, non_image_indices, text_to_overwrite = (
+            batch_indices.to(target_device),
+            non_image_indices.to(target_device),
+            text_to_overwrite.to(target_device),
+        )
+        attention_mask = attention_mask.to(target_device)
+
+        # 4. Fill the embeddings based on the mask. If we have ["hey" "<image>", "how", "are"]
+        # we need to index copy on [0, 577, 578, 579] for the text and [1:576] for the image features
+        final_embedding[batch_indices, text_to_overwrite] = inputs_embeds[batch_indices, non_image_indices]
+        final_attention_mask[batch_indices, text_to_overwrite] = attention_mask[batch_indices, non_image_indices]
+        if labels is not None:
+            final_labels[batch_indices, text_to_overwrite] = labels[batch_indices, non_image_indices]
+
+        # 5. Fill the embeddings corresponding to the images. Anything that is not `text_positions` needs filling (#29835)
+        with torch.no_grad():
+            image_to_overwrite = torch.full(
+                (batch_size, max_embed_dim), True, dtype=torch.bool, device=inputs_embeds.device
+            )
+            image_to_overwrite[batch_indices, text_to_overwrite] = False
+            embed_indices = torch.arange(max_embed_dim).unsqueeze(0).to(target_device)
+            embed_indices = embed_indices.expand(batch_size, max_embed_dim)
+            embed_seq_lens = embed_sequence_lengths[:, None].to(target_device)
+
+            if left_padding:
+                # exclude padding on the left
+                val = (max_embed_dim - embed_indices) <= embed_seq_lens
+            else:
+                # exclude padding on the right
+                val = embed_indices < embed_seq_lens
+            image_to_overwrite &= val
+
+            if image_to_overwrite.sum() != num_image_features:
+                raise ValueError(
+                    f"{image_to_overwrite.sum()=} != {num_image_features=} The input provided to the model are wrong. "
+                    f"The number of image tokens is {torch.sum(special_image_token_mask)} while"
+                    f" the number of image given to the model is {num_images}. "
+                    f"This prevents correct indexing and breaks batch generation."
+                )
+        final_embedding[image_to_overwrite] = image_features.contiguous().reshape(-1, embed_dim).to(target_device)
+        final_attention_mask |= image_to_overwrite
+        position_ids = (final_attention_mask.cumsum(-1) - 1).masked_fill_((final_attention_mask == 0), 1)
+
+        return final_embedding, final_attention_mask, position_ids, final_labels
+
+    @add_start_docstrings_to_model_forward(MAGMA_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=MagmaCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        pixel_values: torch.FloatTensor = None,
+        image_sizes: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        vision_feature_layer: Optional[int] = None,
+        vision_feature_select_strategy: Optional[str] = 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, MagmaCausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, MagmaForConditionalGeneration
+
+        >>> model = MagmaForConditionalGeneration.from_pretrained("microsoft/magma-8b-hf")
+        >>> processor = AutoProcessor.from_pretrained("microsoft/magma-8b-hf")
+
+        >>> prompt = "[INST] <image>\nWhat is shown in this image? [/INST]"
+        >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(text=prompt, images=image, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(**inputs, max_length=30)
+        >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "[INST]  \nWhat is shown in this image? [/INST] The image appears to be a radar chart, which is a type of multi-dimensional plot (...)"
+        ```"""
+        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
+        vision_feature_layer = (
+            vision_feature_layer if vision_feature_layer is not None else self.config.vision_config['vision_feature_layer']
+        )
+
+        if inputs_embeds is None:
+            # 1. Extract the input embeddings
+            # In case image_token_index is not in the embeddings (extra token but embedding don't have it)
+            for_inputs_embeds_ids = input_ids.clone()
+            for_inputs_embeds_ids[(input_ids == self.config.image_token_index)] = 0
+            inputs_embeds = self.get_input_embeddings()(for_inputs_embeds_ids)
+
+            # 2. Merge text and images
+            if pixel_values is not None and input_ids.shape[1] != 1 and pixel_values.size(0) > 0:
+                # ! infer image_num_patches from image_sizes
+                # figure out if pixel_values is concatenated or stacked
+                if pixel_values.dim() == 5:
+                    image_num_patches = [(imsize[:,0]*imsize[:,1]).tolist() for imsize in image_sizes]
+                    # stacking when input is (batch_size, num_patches, num_channels, height, width)
+                    _pixel_values_list = [
+                        pix_val[:num_patch] for pix_val, num_patch in zip(pixel_values, image_num_patches)
+                    ]
+                    pixel_values = torch.cat(_pixel_values_list, dim=0)
+                elif pixel_values.dim() != 4:
+                    # otherwise has to be stacked from list of (num_patches, num_channels, height, width)
+                    raise ValueError(f"pixel_values of shape {pixel_values.shape}, expect to be of 4 or 5 dimensions")
+
+                if self.config.vision_config['img_anyres_strategy'] == "global":
+                    num_patches_for_images = [(imsize[0]*imsize[1]).item() for imsize in image_sizes]
+                    pixel_values_for_images = pixel_values.split(num_patches_for_images, dim=0)
+                    selected_image_features = []
+                    for idx, (image_size, pixel_values_for_image) in enumerate(zip(image_sizes, pixel_values_for_images)):
+                        pixel_values_for_image = pixel_values_for_image.view(image_size[0], image_size[1], *pixel_values_for_image.shape[1:])
+                        pixel_values_for_image = pixel_values_for_image.permute(2, 0, 3, 1, 4).flatten(3, 4).flatten(1, 2).unsqueeze(0)
+                        image_features = self.vision_tower(pixel_values_for_image)
+                        selected_image_feature = image_features[vision_feature_layer][0].permute(1, 2, 0)
+                        selected_image_feature = self.multi_modal_projector((selected_image_feature, None))
+                        selected_image_feature = torch.cat((selected_image_feature, self.multi_modal_projector.row_seperator.repeat(selected_image_feature.shape[0],1,1)), dim=1)
+                        selected_image_features.append(selected_image_feature)
+                elif self.config.vision_config['img_anyres_strategy'] == "crop":
+                    image_features = self.vision_tower(pixel_values)[vision_feature_layer].permute(0, 2, 3, 1)
+                    image_features = self.multi_modal_projector((image_features, None))
+                    num_patches_for_images = [(imsize[0]*imsize[1]).item() for imsize in image_sizes]
+                    image_features_split = torch.split(image_features, num_patches_for_images, dim=0)
+                    selected_image_features = []
+                    for image_feature, image_size in zip(image_features_split, image_sizes):
+                        image_feature = image_feature.view(image_size[0], image_size[1], *image_feature.shape[1:])
+                        image_feature = image_feature.permute(0, 2, 1, 3, 4).flatten(2, 3).flatten(0, 1)
+                        image_feature = torch.cat((image_feature, self.multi_modal_projector.row_seperator.repeat(image_feature.shape[0],1,1)), dim=1)
+                        selected_image_features.append(image_feature)
+
+                # NOTE we only support multimodal_patch_merge_type == "spatial_unpad"
+                feature_lens = [elem.shape[0]*elem.shape[1] for elem in selected_image_features]        
+                image_features = torch.cat([elem.flatten(0, 1) for elem in selected_image_features], 0)
+                feature_lens = torch.tensor(feature_lens, dtype=torch.long, device=image_features.device)
+
+                # inputs_embeds = inputs_embeds.to(image_features.dtype)
+                inputs_embeds, attention_mask, position_ids, labels = self._merge_input_ids_with_image_features(
+                    image_features,
+                    feature_lens,
+                    inputs_embeds,
+                    input_ids,
+                    attention_mask,
+                    position_ids,
+                    labels=labels,
+                )
+
+            # pixel_values is not None but is empty ---> text only cases
+            elif pixel_values is not None and input_ids.shape[1] != 1 and pixel_values.size(0) == 0:
+                # there are no images
+                pass
+
+            # In case input_ids.shape[1] == 1 & pixel_values==None & past_key_values != None, we are in the case of
+            # generation with cache
+            elif past_key_values is not None and pixel_values is not None and input_ids.shape[1] == 1:
+                # Retrieve the first layer to inspect the logits and mask out the hidden states
+                # that are set to 0
+                first_layer_past_key_value = past_key_values[0][0][:, :, :, 0]
+
+                # Sum all dimensions of head_dim (-2) to avoid random errors such as: https://github.com/huggingface/transformers/pull/28032#issuecomment-1863691941
+                batch_index, non_attended_tokens = torch.where(first_layer_past_key_value.float().sum(-2) == 0)
+
+                # Get the target length
+                target_length = input_ids.shape[1]
+                past_length = first_layer_past_key_value.shape[-1]
+
+                extended_attention_mask = torch.ones(
+                    (attention_mask.shape[0], past_length),
+                    dtype=attention_mask.dtype,
+                    device=attention_mask.device,
+                )
+
+                # Filter out only the tokens that can be un-attended, this can happen
+                # if one uses Llava + Fused modules where the cache on the
+                # first iteration is already big enough, or if one passes custom cache
+                valid_indices = non_attended_tokens < extended_attention_mask.size(-1)
+                new_batch_index = batch_index[valid_indices]
+                new_non_attended_tokens = non_attended_tokens[valid_indices]
+
+                # Zero-out the places where we don't need to attend
+                extended_attention_mask[new_batch_index, new_non_attended_tokens] = 0
+
+                attention_mask = torch.cat((extended_attention_mask, attention_mask[:, -target_length:]), dim=1)
+
+                position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1
+
+        outputs = self.language_model(
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        logits = outputs[0]
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            if attention_mask is not None:
+                shift_attention_mask = attention_mask[..., 1:]
+                shift_logits = logits[..., :-1, :][shift_attention_mask.to(logits.device) != 0].contiguous()
+                shift_labels = labels[..., 1:][shift_attention_mask.to(labels.device) != 0].contiguous()
+            else:
+                shift_logits = logits[..., :-1, :].contiguous()
+                shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = nn.CrossEntropyLoss()
+            loss = loss_fct(
+                shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1).to(shift_logits.device)
+            )
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return MagmaCausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        inputs_embeds=None,
+        pixel_values=None,
+        image_sizes=None,
+        attention_mask=None,
+        **kwargs,
+    ):
+        if past_key_values is not None:
+            if isinstance(past_key_values, Cache):
+                cache_length = past_key_values.get_seq_length()
+                past_length = past_key_values.seen_tokens
+            else:
+                cache_length = past_length = past_key_values[0][0].shape[2]
+
+            # Keep only the unprocessed tokens:
+            # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
+            # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
+            # input)
+            if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
+                input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
+            # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
+            # input_ids based on the past_length.
+            elif past_length < input_ids.shape[1]:
+                input_ids = input_ids[:, past_length:]
+            # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
+            elif self.config.image_token_index in input_ids:
+                input_ids = input_ids[:, input_ids.shape[1] - 1 :]
+            # If the cache has seen more tokens than it can hold, then the cache has a size limit. Let's discard the
+            # older attention values, as their corresponding values are not part of the input.
+            if cache_length < past_length and attention_mask is not None:
+                attention_mask = attention_mask[:, -(cache_length + input_ids.shape[1]) :]
+
+        position_ids = kwargs.get("position_ids", None)
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -input_ids.shape[1] :]
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+                "pixel_values": pixel_values,
+                "image_sizes": image_sizes,
+            }
+        )
+        return model_inputs
+
+    def _reorder_cache(self, *args, **kwargs):
+        return self.language_model._reorder_cache(*args, **kwargs)
+
+AutoConfig.register("magma", MagmaConfig)
+AutoModelForCausalLM.register(MagmaConfig, MagmaForConditionalGeneration)