chartinstruct_flant5_modeling.py

from typing import List, Optional, Tuple, Union
from dataclasses import dataclass
import copy, os
import torch
import torch.nn as nn
from torch.nn import CrossEntropyLoss
from transformers import AutoConfig, AutoModelForSeq2SeqLM, \
                         T5Config, T5Model, T5ForConditionalGeneration

from transformers.models.t5.modeling_t5 import T5Stack
from transformers.modeling_outputs import CausalLMOutputWithPast, Seq2SeqLMOutput, BaseModelOutput
from transformers.utils import ModelOutput
from transformers import DonutSwinModel, DonutImageProcessor, DonutSwinConfig

from abc import ABC, abstractmethod
import re

# Model Constants
IGNORE_INDEX = -100
IMAGE_TOKEN_INDEX = -200
DEFAULT_IMAGE_TOKEN = "<image>"
DEFAULT_IMAGE_PATCH_TOKEN = "<im_patch>"
DEFAULT_IM_START_TOKEN = "<im_start>"
DEFAULT_IM_END_TOKEN = "<im_end>"

class UniChartVisionTower(nn.Module):
    def __init__(self, vision_tower, args, delay_load=False):
        super().__init__()

        self.is_loaded = False

        self.vision_tower_name = vision_tower
        self.select_layer = args.mm_vision_select_layer
        self.select_feature = getattr(args, 'mm_vision_select_feature', 'patch')

        if not delay_load:
            self.load_model()
        else:
            self.cfg_only = DonutSwinConfig.from_pretrained(self.vision_tower_name)

    def load_model(self):
        self.image_processor = DonutImageProcessor.from_pretrained(self.vision_tower_name)
        self.vision_tower = DonutSwinModel.from_pretrained(self.vision_tower_name)

        # Changed. Check for this variable. It's false by default.
        if not self.tune_vision_encoder:
            self.vision_tower.requires_grad_(False)

        self.is_loaded = True

    def feature_select(self, image_forward_outs):
        image_features = image_forward_outs.hidden_states[self.select_layer]
        if self.select_feature == 'patch':
            image_features = image_features[:, 1:]
        elif self.select_feature == 'cls_patch':
            image_features = image_features
        else:
            raise ValueError(f'Unexpected select feature: {self.select_feature}')
        return image_features

    @torch.no_grad()
    def forward(self, images):
        if type(images) is list:
            image_features = []
            for image in images:
                image_forward_out = self.vision_tower(image.to(device=self.device, dtype=self.dtype).unsqueeze(0), output_hidden_states=True)
                image_feature = self.feature_select(image_forward_out).to(image.dtype)
                image_features.append(image_feature)
        else:
            image_forward_outs = self.vision_tower(images.to(device=self.device, dtype=self.dtype), output_hidden_states=True)
            image_features = self.feature_select(image_forward_outs).to(images.dtype)

        return image_features

    @property
    def dummy_feature(self):
        return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype)

    @property
    def dtype(self):
        return self.vision_tower.dtype

    @property
    def device(self):
        return self.vision_tower.device

    @property
    def config(self):
        if self.is_loaded:
            return self.vision_tower.config
        else:
            return self.cfg_only

    @property
    def hidden_size(self):
        return self.config.hidden_size

    @property
    def num_patches(self):
        return (self.config.image_size // self.config.patch_size) ** 2

def build_vision_tower(vision_tower_cfg, **kwargs):
    vision_tower = getattr(vision_tower_cfg, 'mm_vision_tower', getattr(vision_tower_cfg, 'vision_tower', None))
    is_absolute_path_exists = os.path.exists(vision_tower)
    if is_absolute_path_exists:
        if 'unichart' in vision_tower:
            return UniChartVisionTower(vision_tower, args=vision_tower_cfg, **kwargs)

    raise ValueError(f'Unknown vision tower: {vision_tower}')

def build_vision_projector(config, delay_load=False, **kwargs):
    projector_type = getattr(config, 'mm_projector_type', 'mlp3x_gelu')

    if projector_type == 'linear':
        return nn.Linear(config.mm_hidden_size, config.hidden_size)

    mlp_gelu_match = re.match(r'^mlp(\d+)x_gelu$', 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))
        return nn.Sequential(*modules)

    raise ValueError(f'Unknown projector type: {projector_type}')

#    Copyright 2023 Haotian Liu
#
#    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.


class LlavaMetaModel:

    def __init__(self, config): #, embed_tokens):
        super(LlavaMetaModel, self).__init__(config) #, embed_tokens)
        if hasattr(config, "mm_vision_tower"):
            self.vision_tower = build_vision_tower(config, delay_load=True)
            self.mm_projector = build_vision_projector(self.config) #nn.Linear(config.mm_hidden_size, config.hidden_size)

    def get_vision_tower(self):
        vision_tower = getattr(self, 'vision_tower', None)
        if type(vision_tower) is list:
            vision_tower = vision_tower[0]
        return vision_tower

    def initialize_vision_modules(self, model_args, fsdp=None):
        vision_tower = model_args.vision_tower
        mm_vision_select_layer = model_args.mm_vision_select_layer
        mm_vision_select_feature = model_args.mm_vision_select_feature
        pretrain_mm_mlp_adapter = model_args.pretrain_mm_mlp_adapter

        self.config.mm_vision_tower = vision_tower

        vision_tower = build_vision_tower(model_args)

        if fsdp is not None and len(fsdp) > 0:
            self.vision_tower = [vision_tower]
        else:
            self.vision_tower = vision_tower

        self.config.use_mm_proj = True
        self.config.mm_projector_type = getattr(model_args, 'mm_projector_type', 'linear')
        self.config.mm_hidden_size = vision_tower.hidden_size
        self.config.mm_vision_select_layer = mm_vision_select_layer
        self.config.mm_vision_select_feature = mm_vision_select_feature

        if not hasattr(self, 'mm_projector'):
            self.mm_projector = build_vision_projector(self.config) #nn.Linear(self.config.mm_hidden_size, self.config.hidden_size)

        if pretrain_mm_mlp_adapter is not None:
            mm_projector_weights = torch.load(pretrain_mm_mlp_adapter, map_location='cpu')
            def get_w(weights, keyword):
                return {k.split(keyword + '.')[1]: v for k, v in weights.items() if keyword in k}

            self.mm_projector.load_state_dict(get_w(mm_projector_weights, 'mm_projector'))


class LlavaMetaForCausalLM(ABC):

    @abstractmethod
    def get_model(self):
        pass

    def get_vision_tower(self):
        return self.get_model().get_vision_tower()

    def encode_images(self, images):
        image_features = self.get_model().get_vision_tower()(images)
        image_features = self.get_model().mm_projector(image_features)
        return image_features

    def prepare_inputs_labels_for_multimodal(
        self, input_ids, attention_mask, past_key_values, labels, images
    ):
        vision_tower = self.get_vision_tower()
        if vision_tower is None or images is None or input_ids.shape[1] == 1:
            if past_key_values is not None and vision_tower is not None and images is not None and input_ids.shape[1] == 1:
                attention_mask = torch.ones((attention_mask.shape[0], past_key_values[-1][-1].shape[-2] + 1), dtype=attention_mask.dtype, device=attention_mask.device)
            return input_ids, attention_mask, past_key_values, None, labels

        if type(images) is list or images.ndim == 5:
            concat_images = torch.cat([image for image in images], dim=0)
            image_features = self.encode_images(concat_images)
            split_sizes = [image.shape[0] for image in images]
            image_features = torch.split(image_features, split_sizes, dim=0)
            image_features = [x.flatten(0, 1) for x in image_features]
        else:
            image_features = self.encode_images(images)

        new_input_embeds = []
        new_labels = [] if labels is not None else None
        cur_image_idx = 0
        for batch_idx, cur_input_ids in enumerate(input_ids):
            if (cur_input_ids == IMAGE_TOKEN_INDEX).sum() == 0:
                # multimodal LLM, but the current sample is not multimodal
                cur_input_embeds = self.get_model().embed_tokens(cur_input_ids)
                cur_input_embeds = cur_input_embeds + (0. * self.get_model().mm_projector(vision_tower.dummy_feature)).sum()
                new_input_embeds.append(cur_input_embeds)
                if labels is not None:
                    new_labels.append(labels[batch_idx])
                cur_image_idx += 1
                continue
            image_token_indices = torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0]
            cur_new_input_embeds = []
            if labels is not None:
                cur_labels = labels[batch_idx]
                cur_new_labels = []
                assert cur_labels.shape == cur_input_ids.shape
            while image_token_indices.numel() > 0:
                cur_image_features = image_features[cur_image_idx]
                image_token_start = image_token_indices[0]
                if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
                    cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[:image_token_start-1]).detach())
                    cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[image_token_start-1:image_token_start]))
                    cur_new_input_embeds.append(cur_image_features)
                    cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[image_token_start+1:image_token_start+2]))
                    if labels is not None:
                        cur_new_labels.append(cur_labels[:image_token_start])
                        cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=labels.device, dtype=labels.dtype))
                        cur_new_labels.append(cur_labels[image_token_start:image_token_start+1])
                        cur_labels = cur_labels[image_token_start+2:]
                else:
                    cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[:image_token_start]))
                    cur_new_input_embeds.append(cur_image_features)
                    if labels is not None:
                        cur_new_labels.append(cur_labels[:image_token_start])
                        cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=labels.device, dtype=labels.dtype))
                        cur_labels = cur_labels[image_token_start+1:]
                cur_image_idx += 1
                if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
                    cur_input_ids = cur_input_ids[image_token_start+2:]
                else:
                    cur_input_ids = cur_input_ids[image_token_start+1:]
                image_token_indices = torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0]
            if cur_input_ids.numel() > 0:
                if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
                    cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids).detach())
                else:
                    cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids))
                if labels is not None:
                    cur_new_labels.append(cur_labels)
            cur_new_input_embeds = [x.to(device=self.device) for x in cur_new_input_embeds]
            cur_new_input_embeds = torch.cat(cur_new_input_embeds, dim=0)
            new_input_embeds.append(cur_new_input_embeds)
            if labels is not None:
                cur_new_labels = torch.cat(cur_new_labels, dim=0)
                new_labels.append(cur_new_labels)

        if any(x.shape != new_input_embeds[0].shape for x in new_input_embeds):
            max_len = max(x.shape[0] for x in new_input_embeds)

            new_input_embeds_align = []
            for cur_new_embed in new_input_embeds:
                cur_new_embed = torch.cat((cur_new_embed, torch.zeros((max_len - cur_new_embed.shape[0], cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)), dim=0)
                new_input_embeds_align.append(cur_new_embed)
            new_input_embeds = torch.stack(new_input_embeds_align, dim=0)

            if labels is not None:
                new_labels_align = []
                _new_labels = new_labels
                for cur_new_label in new_labels:
                    cur_new_label = torch.cat((cur_new_label, torch.full((max_len - cur_new_label.shape[0],), IGNORE_INDEX, dtype=cur_new_label.dtype, device=cur_new_label.device)), dim=0)
                    new_labels_align.append(cur_new_label)
                new_labels = torch.stack(new_labels_align, dim=0)

            if attention_mask is not None:
                new_attention_mask = []
                for cur_attention_mask, cur_new_labels, cur_new_labels_align in zip(attention_mask, _new_labels, new_labels):
                    new_attn_mask_pad_left = torch.full((cur_new_labels.shape[0] - labels.shape[1],), True, dtype=attention_mask.dtype, device=attention_mask.device)
                    new_attn_mask_pad_right = torch.full((cur_new_labels_align.shape[0] - cur_new_labels.shape[0],), False, dtype=attention_mask.dtype, device=attention_mask.device)
                    cur_new_attention_mask = torch.cat((new_attn_mask_pad_left, cur_attention_mask, new_attn_mask_pad_right), dim=0)
                    new_attention_mask.append(cur_new_attention_mask)
                attention_mask = torch.stack(new_attention_mask, dim=0)
                assert attention_mask.shape == new_labels.shape
        else:
            new_input_embeds = torch.stack(new_input_embeds, dim=0)
            if labels is not None:
                new_labels  = torch.stack(new_labels, dim=0)

            if attention_mask is not None:
                new_attn_mask_pad_left = torch.full((attention_mask.shape[0], new_input_embeds.shape[1] - input_ids.shape[1]), True, dtype=attention_mask.dtype, device=attention_mask.device)
                attention_mask = torch.cat((new_attn_mask_pad_left, attention_mask), dim=1)
                assert attention_mask.shape == new_input_embeds.shape[:2]

        return None, attention_mask, past_key_values, new_input_embeds, new_labels

    def initialize_vision_tokenizer(self, model_args, tokenizer):
        if model_args.mm_use_im_patch_token:
            tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
            self.resize_token_embeddings(len(tokenizer))

        if model_args.mm_use_im_start_end:
            num_new_tokens = tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)
            self.resize_token_embeddings(len(tokenizer))

            if num_new_tokens > 0:
                input_embeddings = self.get_input_embeddings().weight.data
                output_embeddings = self.get_output_embeddings().weight.data

                input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(
                    dim=0, keepdim=True)
                output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(
                    dim=0, keepdim=True)

                input_embeddings[-num_new_tokens:] = input_embeddings_avg
                output_embeddings[-num_new_tokens:] = output_embeddings_avg

            if model_args.tune_mm_mlp_adapter:
                for p in self.get_input_embeddings().parameters():
                    p.requires_grad = True
                for p in self.get_output_embeddings().parameters():
                    p.requires_grad = False

            if model_args.pretrain_mm_mlp_adapter:
                mm_projector_weights = torch.load(model_args.pretrain_mm_mlp_adapter, map_location='cpu')
                embed_tokens_weight = mm_projector_weights['model.embed_tokens.weight']
                assert num_new_tokens == 2
                if input_embeddings.shape == embed_tokens_weight.shape:
                    input_embeddings[-num_new_tokens:] = embed_tokens_weight[-num_new_tokens:]
                elif embed_tokens_weight.shape[0] == num_new_tokens:
                    input_embeddings[-num_new_tokens:] = embed_tokens_weight
                else:
                    raise ValueError(f"Unexpected embed_tokens_weight shape. Pretrained: {embed_tokens_weight.shape}. Current: {input_embeddings.shape}. Numer of new tokens: {num_new_tokens}.")
        elif model_args.mm_use_im_patch_token:
            if model_args.tune_mm_mlp_adapter:
                for p in self.get_input_embeddings().parameters():
                    p.requires_grad = False
                for p in self.get_output_embeddings().parameters():
                    p.requires_grad = False



class LlavaMetaForConditionalGeneration(ABC):

    def get_vision_tower(self):
        return self.get_encoder().get_vision_tower()

    def encode_images(self, images):
        image_features = self.get_encoder().get_vision_tower()(images)
        image_features = self.get_encoder().mm_projector(image_features)
        return image_features

    def prepare_inputs_labels_for_multimodal(
        self, input_ids, attention_mask, labels, images
    ):
        vision_tower = self.get_vision_tower()
        if vision_tower is None or images is None or input_ids.shape[1] == 1:
            return input_ids, attention_mask, None

        if type(images) is list or images.ndim == 5:
            concat_images = torch.cat([image for image in images], dim=0)
            image_features = self.encode_images(concat_images)
            split_sizes = [image.shape[0] for image in images]
            image_features = torch.split(image_features, split_sizes, dim=0)
            image_features = [x.flatten(0, 1) for x in image_features]
        else:
            image_features = self.encode_images(images)

        # TODO: double check.
        if labels is None:
            labels = torch.full_like(input_ids, IGNORE_INDEX)
        ######

        new_input_embeds = []
        new_labels = [] if labels is not None else None
        cur_image_idx = 0
        for batch_idx, cur_input_ids in enumerate(input_ids):
            if (cur_input_ids == IMAGE_TOKEN_INDEX).sum() == 0:
                # multimodal LLM, but the current sample is not multimodal
                cur_input_embeds = self.get_encoder().embed_tokens(cur_input_ids)
                cur_input_embeds = cur_input_embeds + (0. * self.get_encoder().mm_projector(vision_tower.dummy_feature)).sum()
                new_input_embeds.append(cur_input_embeds)
                if labels is not None:
                    new_labels.append(labels[batch_idx])
                cur_image_idx += 1
                continue
            image_token_indices = torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0]
            cur_new_input_embeds = []
            if labels is not None:
                cur_labels = labels[batch_idx]
                cur_new_labels = []
                assert cur_labels.shape == cur_input_ids.shape
            while image_token_indices.numel() > 0:
                cur_image_features = image_features[cur_image_idx]
                image_token_start = image_token_indices[0]
                if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
                    cur_new_input_embeds.append(self.get_encoder().embed_tokens(cur_input_ids[:image_token_start-1]).detach())
                    cur_new_input_embeds.append(self.get_encoder().embed_tokens(cur_input_ids[image_token_start-1:image_token_start]))
                    cur_new_input_embeds.append(cur_image_features)
                    cur_new_input_embeds.append(self.get_encoder().embed_tokens(cur_input_ids[image_token_start+1:image_token_start+2]))
                    if labels is not None:
                        cur_new_labels.append(cur_labels[:image_token_start])
                        cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=labels.device, dtype=labels.dtype))
                        cur_new_labels.append(cur_labels[image_token_start:image_token_start+1])
                        cur_labels = cur_labels[image_token_start+2:]
                else:
                    cur_new_input_embeds.append(self.get_encoder().embed_tokens(cur_input_ids[:image_token_start]))
                    cur_new_input_embeds.append(cur_image_features)
                    if labels is not None:
                        cur_new_labels.append(cur_labels[:image_token_start])
                        cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=labels.device, dtype=labels.dtype))
                        cur_labels = cur_labels[image_token_start+1:]
                cur_image_idx += 1
                if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
                    cur_input_ids = cur_input_ids[image_token_start+2:]
                else:
                    cur_input_ids = cur_input_ids[image_token_start+1:]
                image_token_indices = torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0]
            if cur_input_ids.numel() > 0:
                if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
                    cur_new_input_embeds.append(self.get_encoder().embed_tokens(cur_input_ids).detach())
                else:
                    cur_new_input_embeds.append(self.get_encoder().embed_tokens(cur_input_ids))
                if labels is not None:
                    cur_new_labels.append(cur_labels)
            cur_new_input_embeds = [x.to(device=self.device) for x in cur_new_input_embeds]
            cur_new_input_embeds = torch.cat(cur_new_input_embeds, dim=0)
            new_input_embeds.append(cur_new_input_embeds)
            if labels is not None:
                cur_new_labels = torch.cat(cur_new_labels, dim=0)
                new_labels.append(cur_new_labels)

        if any(x.shape != new_input_embeds[0].shape for x in new_input_embeds):
            max_len = max(x.shape[0] for x in new_input_embeds)

            new_input_embeds_align = []
            for cur_new_embed in new_input_embeds:
                cur_new_embed = torch.cat((cur_new_embed, torch.zeros((max_len - cur_new_embed.shape[0], cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)), dim=0)
                new_input_embeds_align.append(cur_new_embed)
            new_input_embeds = torch.stack(new_input_embeds_align, dim=0)

            if labels is not None:
                new_labels_align = []
                _new_labels = new_labels
                for cur_new_label in new_labels:
                    cur_new_label = torch.cat((cur_new_label, torch.full((max_len - cur_new_label.shape[0],), IGNORE_INDEX, dtype=cur_new_label.dtype, device=cur_new_label.device)), dim=0)
                    new_labels_align.append(cur_new_label)
                new_labels = torch.stack(new_labels_align, dim=0)

            if attention_mask is not None:
                new_attention_mask = []
                for cur_attention_mask, cur_new_labels, cur_new_labels_align in zip(attention_mask, _new_labels, new_labels):
                    new_attn_mask_pad_left = torch.full((cur_new_labels.shape[0] - labels.shape[1],), True, dtype=attention_mask.dtype, device=attention_mask.device)
                    new_attn_mask_pad_right = torch.full((cur_new_labels_align.shape[0] - cur_new_labels.shape[0],), False, dtype=attention_mask.dtype, device=attention_mask.device)
                    cur_new_attention_mask = torch.cat((new_attn_mask_pad_left, cur_attention_mask, new_attn_mask_pad_right), dim=0)
                    new_attention_mask.append(cur_new_attention_mask)
                attention_mask = torch.stack(new_attention_mask, dim=0)
                assert attention_mask.shape == new_labels.shape
        else:
            new_input_embeds = torch.stack(new_input_embeds, dim=0)
            if labels is not None:
                new_labels  = torch.stack(new_labels, dim=0)

            if attention_mask is not None:
                new_attn_mask_pad_left = torch.full((attention_mask.shape[0], new_input_embeds.shape[1] - input_ids.shape[1]), True, dtype=attention_mask.dtype, device=attention_mask.device)
                attention_mask = torch.cat((new_attn_mask_pad_left, attention_mask), dim=1)
                assert attention_mask.shape == new_input_embeds.shape[:2]

        return None, attention_mask, new_input_embeds, new_labels

    def initialize_vision_tokenizer(self, model_args, tokenizer):
        if model_args.mm_use_im_patch_token:
            tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
            self.resize_token_embeddings(len(tokenizer))

        if model_args.mm_use_im_start_end:
            num_new_tokens = tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)
            self.resize_token_embeddings(len(tokenizer))

            if num_new_tokens > 0:
                input_embeddings = self.get_input_embeddings().weight.data
                output_embeddings = self.get_output_embeddings().weight.data

                input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(
                    dim=0, keepdim=True)
                output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(
                    dim=0, keepdim=True)

                input_embeddings[-num_new_tokens:] = input_embeddings_avg
                output_embeddings[-num_new_tokens:] = output_embeddings_avg

            if model_args.tune_mm_mlp_adapter:
                for p in self.get_input_embeddings().parameters():
                    p.requires_grad = True
                for p in self.get_output_embeddings().parameters():
                    p.requires_grad = False

            if model_args.pretrain_mm_mlp_adapter:
                mm_projector_weights = torch.load(model_args.pretrain_mm_mlp_adapter, map_location='cpu')
                embed_tokens_weight = mm_projector_weights['model.embed_tokens.weight']
                assert num_new_tokens == 2
                if input_embeddings.shape == embed_tokens_weight.shape:
                    input_embeddings[-num_new_tokens:] = embed_tokens_weight[-num_new_tokens:]
                elif embed_tokens_weight.shape[0] == num_new_tokens:
                    input_embeddings[-num_new_tokens:] = embed_tokens_weight
                else:
                    raise ValueError(f"Unexpected embed_tokens_weight shape. Pretrained: {embed_tokens_weight.shape}. Current: {input_embeddings.shape}. Numer of new tokens: {num_new_tokens}.")
        elif model_args.mm_use_im_patch_token:
            if model_args.tune_mm_mlp_adapter:
                for p in self.get_input_embeddings().parameters():
                    p.requires_grad = False
                for p in self.get_output_embeddings().parameters():
                    p.requires_grad = False

class LlavaMetaT5Model:

    def __init__(self, config, embed_tokens):
        super(LlavaMetaT5Model, self).__init__(config, embed_tokens)
        if hasattr(config, "mm_vision_tower"):
            self.vision_tower = build_vision_tower(config, delay_load=True)
            self.mm_projector = nn.Linear(config.mm_hidden_size, config.hidden_size)

    def get_vision_tower(self):
        vision_tower = getattr(self, 'vision_tower', None)
        if type(vision_tower) is list:
            vision_tower = vision_tower[0]
        return vision_tower

    def initialize_vision_modules(self, model_args, fsdp=None):
        vision_tower = model_args.vision_tower
        mm_vision_select_layer = model_args.mm_vision_select_layer
        mm_vision_select_feature = model_args.mm_vision_select_feature
        pretrain_mm_mlp_adapter = model_args.pretrain_mm_mlp_adapter

        self.config.mm_vision_tower = vision_tower

        vision_tower = build_vision_tower(model_args)

        if fsdp is not None and len(fsdp) > 0:
            self.vision_tower = [vision_tower]
        else:
            self.vision_tower = vision_tower

        self.config.use_mm_proj = True
        self.config.mm_projector_type = getattr(model_args, 'mm_projector_type', 'linear')
        self.config.mm_hidden_size = vision_tower.hidden_size
        self.config.mm_vision_select_layer = mm_vision_select_layer
        self.config.mm_vision_select_feature = mm_vision_select_feature

        if not hasattr(self, 'mm_projector'):
            self.mm_projector = build_vision_projector(self.config) #nn.Linear(self.config.mm_hidden_size, self.config.hidden_size)

        if pretrain_mm_mlp_adapter is not None:
            mm_projector_weights = torch.load(pretrain_mm_mlp_adapter, map_location='cpu')
            def get_w(weights, keyword):
                return {k.split(keyword + '.')[1]: v for k, v in weights.items() if keyword in k}

            self.mm_projector.load_state_dict(get_w(mm_projector_weights, 'mm_projector'))



#    Copyright 2023 Haotian Liu
#
#    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.



@dataclass
class BaseModelOutputWithPastAndCrossAttentionsWithAttentionMask(ModelOutput):
    """
    Base class for model's outputs that may also contain a past key/values (to speed up sequential decoding).

    Args:
        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
            Sequence of hidden-states at the output of the last layer of the model.

            If `past_key_values` is used only the last hidden-state of the sequences of shape `(batch_size, 1,
            hidden_size)` is output.
        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 optionally if
            `config.is_encoder_decoder=True` 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 optionally if
            `config.is_encoder_decoder=True` in the cross-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.
        cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` and `config.add_cross_attention=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 of the decoder's cross-attention layer, after the attention softmax, used to compute the
            weighted average in the cross-attention heads.
    """

    last_hidden_state: torch.FloatTensor = None
    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[torch.FloatTensor]] = None
    cross_attentions: Optional[Tuple[torch.FloatTensor]] = None
    attention_mask: Optional[torch.LongTensor] = None

class LlavaT5Config(T5Config):
    model_type = "llava_t5"

class LlavaT5Stack(LlavaMetaT5Model, T5Stack):
    config_class = LlavaT5Config

    def __init__(self, config: T5Config, embed_tokens=None):
        super(LlavaT5Stack, self).__init__(config, embed_tokens)


class LlavaT5ForConditionalGeneration(T5ForConditionalGeneration, LlavaMetaForConditionalGeneration):
    config_class = LlavaT5Config

    def __init__(self, config):
        super(T5ForConditionalGeneration, self).__init__(config)

        self.model_dim = config.d_model

        self.shared = nn.Embedding(config.vocab_size, config.d_model)

        encoder_config = copy.deepcopy(config)
        encoder_config.is_decoder = False
        encoder_config.use_cache = False
        encoder_config.is_encoder_decoder = False
        self.encoder = LlavaT5Stack(encoder_config, self.shared)

        decoder_config = copy.deepcopy(config)
        decoder_config.is_decoder = True
        decoder_config.is_encoder_decoder = False
        decoder_config.num_layers = config.num_decoder_layers
        self.decoder = T5Stack(decoder_config, self.shared)

        self.lm_head = nn.Linear(config.d_model, config.vocab_size, bias=False)

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

        # Model parallel
        self.model_parallel = False
        self.device_map = None

    def get_model(self):
        return self.encoder
    def get_encoder(self):
        return self.encoder
    def get_decoder(self):
        return self.decoder

    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = 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,
        images: Optional[torch.FloatTensor] = None,
        return_dict: Optional[bool] = None,

        decoder_input_ids: Optional[torch.LongTensor] = None,
        decoder_attention_mask: Optional[torch.BoolTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        decoder_head_mask: Optional[torch.FloatTensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        encoder_outputs: Optional[Tuple[Tuple[torch.Tensor]]] = None,
        decoder_inputs_embeds: Optional[torch.FloatTensor] = None,

    ) -> Union[Tuple, Seq2SeqLMOutput]:

        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict


        # FutureWarning: head_mask was separated into two input args - head_mask, decoder_head_mask
        if head_mask is not None and decoder_head_mask is None:
            if self.config.num_layers == self.config.num_decoder_layers:
                #warnings.warn(__HEAD_MASK_WARNING_MSG, FutureWarning)
                decoder_head_mask = head_mask

        # Encode if needed (training, first prediction pass)
        if encoder_outputs is None:
            input_ids, attention_mask, inputs_embeds, _ = self.prepare_inputs_labels_for_multimodal(input_ids,
                                                                                            attention_mask,
                                                                                            None,  # Important: keep it None
                                                                                            images
                                                                                            )
            # Convert encoder inputs in embeddings if needed
            encoder_outputs = self.encoder(
                input_ids=input_ids,
                attention_mask=attention_mask,
                inputs_embeds=inputs_embeds,
                head_mask=head_mask,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict=return_dict,
            )
        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
            encoder_outputs = BaseModelOutput(
                last_hidden_state=encoder_outputs[0],
                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
            )

        hidden_states = encoder_outputs[0]

        if self.model_parallel:
            torch.cuda.set_device(self.decoder.first_device)

        if labels is not None and decoder_input_ids is None and decoder_inputs_embeds is None:
            # get decoder inputs from shifting lm labels to the right
            decoder_input_ids = self._shift_right(labels)

        # Set device for model parallelism
        if self.model_parallel:
            torch.cuda.set_device(self.decoder.first_device)
            hidden_states = hidden_states.to(self.decoder.first_device)
            if decoder_input_ids is not None:
                decoder_input_ids = decoder_input_ids.to(self.decoder.first_device)
            if attention_mask is not None:
                attention_mask = attention_mask.to(self.decoder.first_device)
            if decoder_attention_mask is not None:
                decoder_attention_mask = decoder_attention_mask.to(self.decoder.first_device)


        # Decode
        decoder_outputs = self.decoder(
            input_ids=decoder_input_ids,
            attention_mask=decoder_attention_mask,
            inputs_embeds=decoder_inputs_embeds,
            past_key_values=past_key_values,
            encoder_hidden_states=hidden_states,
            encoder_attention_mask=attention_mask,
            head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        sequence_output = decoder_outputs[0]

        # Set device for model parallelism
        if self.model_parallel:
            torch.cuda.set_device(self.encoder.first_device)
            self.lm_head = self.lm_head.to(self.encoder.first_device)
            sequence_output = sequence_output.to(self.lm_head.weight.device)

        if self.config.tie_word_embeddings:
            # Rescale output before projecting on vocab
            # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/transformer.py#L586
            sequence_output = sequence_output * (self.model_dim**-0.5)

        lm_logits = self.lm_head(sequence_output)

        loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss(ignore_index=-100)
            # move labels to correct device to enable PP
            labels = labels.to(lm_logits.device)
            loss = loss_fct(lm_logits.view(-1, lm_logits.size(-1)), labels.view(-1))
            # TODO(thom): Add z_loss https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L666

        if not return_dict:
            output = (lm_logits,) + decoder_outputs[1:] + encoder_outputs
            return ((loss,) + output) if loss is not None else output

        return Seq2SeqLMOutput(
            loss=loss,
            logits=lm_logits,
            past_key_values=decoder_outputs.past_key_values,
            decoder_hidden_states=decoder_outputs.hidden_states,
            decoder_attentions=decoder_outputs.attentions,
            cross_attentions=decoder_outputs.cross_attentions,
            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
            encoder_hidden_states=encoder_outputs.hidden_states,
            encoder_attentions=encoder_outputs.attentions,
        )

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        attention_mask=None,
        head_mask=None,
        decoder_head_mask=None,
        decoder_attention_mask=None,
        cross_attn_head_mask=None,
        use_cache=None,
        encoder_outputs=None,
        **kwargs,
    ):
        # cut decoder_input_ids if past_key_values is used
        if past_key_values is not None:
            past_length = past_key_values[0][0].shape[2]

            # Some generation methods already pass only the last input ID
            if input_ids.shape[1] > past_length:
                remove_prefix_length = past_length
            else:
                # Default to old behavior: keep only final ID
                remove_prefix_length = input_ids.shape[1] - 1

            input_ids = input_ids[:, remove_prefix_length:]

        return {
            "decoder_input_ids": input_ids,
            "past_key_values": past_key_values,
            "encoder_outputs": encoder_outputs,
            "attention_mask": attention_mask,
            "head_mask": head_mask,
            "decoder_head_mask": decoder_head_mask,
            "decoder_attention_mask": decoder_attention_mask,
            "cross_attn_head_mask": cross_attn_head_mask,
            "use_cache": use_cache,
            "images": kwargs.get("images", None),
        }