utils/llava_qwen.py

#    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.


from typing import List, Optional, Tuple, Union

import PIL
from PIL import Image
PIL.Image.MAX_IMAGE_PIXELS=500000000

import re
import copy
# from timm.models import create_model
from abc import ABC, abstractmethod

import torch
import torch.nn as nn
from torch import Tensor
import torch.nn.functional as F

from transformers import AutoConfig, AutoModelForCausalLM, Qwen2Config, Qwen2Model, Qwen2ForCausalLM

from transformers.modeling_outputs import CausalLMOutputWithPast
from transformers.generation.utils import GenerateOutput

from typing import List, Tuple, Optional, Union, Dict, Any

# from timm.models import register_model
# from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
# from timm.layers import DropPath, SqueezeExcite

CONTROLLER_HEART_BEAT_EXPIRATION = 30
WORKER_HEART_BEAT_INTERVAL = 15
LOGDIR = "."
# 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>"
IMAGE_PLACEHOLDER = "<image-placeholder>"

class LlavaConfig(Qwen2Config):
    model_type = "llava_qwen2"

class LlavaMetaModel:

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

        if hasattr(config, "mm_vision_tower"):
            self.vision_tower = build_vision_tower(config, delay_load=True)
            self.mm_projector = build_vision_projector(config)

            if 'unpad' in getattr(config, 'mm_patch_merge_type', ''):
                self.image_newline = nn.Parameter(
                    torch.empty(config.hidden_size, dtype=self.dtype)
                )

    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
        mm_patch_merge_type = model_args.mm_patch_merge_type

        self.config.mm_vision_tower = vision_tower

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

        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
        self.config.mm_patch_merge_type = mm_patch_merge_type

        if getattr(self, 'mm_projector', None) is None:
            self.mm_projector = build_vision_projector(self.config)

            if 'unpad' in mm_patch_merge_type:
                embed_std = 1 / torch.sqrt(torch.tensor(self.config.hidden_size, dtype=self.dtype))
                self.image_newline = nn.Parameter(
                    torch.randn(self.config.hidden_size, dtype=self.dtype) * embed_std
                )
        else:
            # In case it is frozen by LoRA
            for p in self.mm_projector.parameters():
                p.requires_grad = True

        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'))

def select_best_resolution(original_size, possible_resolutions):
    """
    Selects the best resolution from a list of possible resolutions based on the original size.

    Args:
        original_size (tuple): The original size of the image in the format (width, height).
        possible_resolutions (list): A list of possible resolutions in the format [(width1, height1), (width2, height2), ...].

    Returns:
        tuple: The best fit resolution in the format (width, height).
    """
    original_width, original_height = original_size
    best_fit = None
    max_effective_resolution = 0
    min_wasted_resolution = float('inf')

    for width, height in possible_resolutions:
        scale = min(width / original_width, height / original_height)
        downscaled_width, downscaled_height = int(original_width * scale), int(original_height * scale)
        effective_resolution = min(downscaled_width * downscaled_height, original_width * original_height)
        wasted_resolution = (width * height) - effective_resolution

        if effective_resolution > max_effective_resolution or (effective_resolution == max_effective_resolution and wasted_resolution < min_wasted_resolution):
            max_effective_resolution = effective_resolution
            min_wasted_resolution = wasted_resolution
            best_fit = (width, height)

    return best_fit

def get_anyres_image_grid_shape(image_size, grid_pinpoints, patch_size):
    """
    Calculate the shape of the image patch grid after the preprocessing for images of any resolution.

    Args:
        image_size (tuple): The size of the input image in the format (width, height).
        grid_pinpoints (str): A string representation of a list of possible resolutions.
        patch_size (int): The size of each image patch.

    Returns:
        tuple: The shape of the image patch grid in the format (width, height).
    """
    import ast
    if type(grid_pinpoints) is list:
        possible_resolutions = grid_pinpoints
    else:
        possible_resolutions = ast.literal_eval(grid_pinpoints)
    width, height = select_best_resolution(image_size, possible_resolutions)
    return width // patch_size, height // patch_size

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

        if type(images) is list or images.ndim == 5:
            if type(images) is list:
                images = [x.unsqueeze(0) if x.ndim == 3 else x for x in images]
            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)
            mm_patch_merge_type = getattr(self.config, 'mm_patch_merge_type', 'flat')
            image_aspect_ratio = getattr(self.config, 'image_aspect_ratio', 'square')
            if mm_patch_merge_type == 'flat':
                image_features = [x.flatten(0, 1) for x in image_features]
            elif mm_patch_merge_type.startswith('spatial'):
                new_image_features = []
                for image_idx, image_feature in enumerate(image_features):
                    if image_feature.shape[0] > 1:
                        base_image_feature = image_feature[0]
                        image_feature = image_feature[1:]
                        height = width = self.get_vision_tower().num_patches_per_side
                        assert height * width == base_image_feature.shape[0]
                        if image_aspect_ratio == 'anyres':
                            if hasattr(self.get_vision_tower(), 's2_image_size'):
                                img_size = self.get_vision_tower().s2_image_size
                            elif isinstance(self.get_vision_tower().config, dict):
                                img_size = self.get_vision_tower().config["image_cfg"]["image_size"]
                            else:
                                img_size = self.get_vision_tower().config.image_size

                            num_patch_width, num_patch_height = get_anyres_image_grid_shape(image_sizes[image_idx], self.config.image_grid_pinpoints, img_size)
                            image_feature = image_feature.view(num_patch_height, num_patch_width, height, width, -1)
                        else:
                            raise NotImplementedError
                        if 'unpad' in mm_patch_merge_type:
                            image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
                            image_feature = image_feature.flatten(1, 2).flatten(2, 3)
                            image_feature = unpad_image(image_feature, image_sizes[image_idx])
                            image_feature = torch.cat((
                                image_feature,
                                self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1).to(image_feature.device)
                            ), dim=-1)
                            image_feature = image_feature.flatten(1, 2).transpose(0, 1)
                        else:
                            image_feature = image_feature.permute(0, 2, 1, 3, 4).contiguous()
                            image_feature = image_feature.flatten(0, 3)
                        image_feature = torch.cat((base_image_feature, image_feature), dim=0)
                    else:
                        image_feature = image_feature[0]
                        if 'unpad' in mm_patch_merge_type:
                            image_feature = torch.cat((
                                image_feature,
                                self.model.image_newline[None].to(image_feature.device)
                            ), dim=0)
                    new_image_features.append(image_feature)
                image_features = new_image_features
            else:
                raise ValueError(f"Unexpected mm_patch_merge_type: {self.config.mm_patch_merge_type}")
        else:
            image_features = self.encode_images(images)

        # TODO: image start / end is not implemented here to support pretraining.
        if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
            raise NotImplementedError

        # Let's just add dummy tensors if they do not exist,
        # it is a headache to deal with None all the time.
        # But it is not ideal, and if you have a better idea,
        # please open an issue / submit a PR, thanks.
        _labels = labels
        _position_ids = position_ids
        _attention_mask = attention_mask
        if attention_mask is None:
            attention_mask = torch.ones_like(input_ids, dtype=torch.bool)
        else:
            attention_mask = attention_mask.bool()
        if position_ids is None:
            position_ids = torch.arange(0, input_ids.shape[1], dtype=torch.long, device=input_ids.device)
        if labels is None:
            labels = torch.full_like(input_ids, IGNORE_INDEX)

        # remove the padding using attention_mask -- FIXME
        _input_ids = input_ids
        input_ids = [cur_input_ids[cur_attention_mask] for cur_input_ids, cur_attention_mask in zip(input_ids, attention_mask)]
        labels = [cur_labels[cur_attention_mask] for cur_labels, cur_attention_mask in zip(labels, attention_mask)]

        new_input_embeds = []
        new_labels = []
        cur_image_idx = 0
        for batch_idx, cur_input_ids in enumerate(input_ids):
            num_images = (cur_input_ids == IMAGE_TOKEN_INDEX).sum()
            if num_images == 0:
                cur_image_features = image_features[cur_image_idx]
                cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids)
                cur_input_embeds = torch.cat([cur_input_embeds_1, cur_image_features[0:0]], dim=0)
                new_input_embeds.append(cur_input_embeds)
                new_labels.append(labels[batch_idx])
                cur_image_idx += 1
                continue

            image_token_indices = [-1] + torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0].tolist() + [cur_input_ids.shape[0]]
            cur_input_ids_noim = []
            cur_labels = labels[batch_idx]
            cur_labels_noim = []
            for i in range(len(image_token_indices) - 1):
                cur_input_ids_noim.append(cur_input_ids[image_token_indices[i]+1:image_token_indices[i+1]])
                cur_labels_noim.append(cur_labels[image_token_indices[i]+1:image_token_indices[i+1]])
            split_sizes = [x.shape[0] for x in cur_labels_noim]
            cur_input_embeds = self.get_model().embed_tokens(torch.cat(cur_input_ids_noim))
            cur_input_embeds_no_im = torch.split(cur_input_embeds, split_sizes, dim=0)
            cur_new_input_embeds = []
            cur_new_labels = []

            for i in range(num_images + 1):
                cur_new_input_embeds.append(cur_input_embeds_no_im[i])
                cur_new_labels.append(cur_labels_noim[i])
                if i < num_images:
                    cur_image_features = image_features[cur_image_idx]
                    cur_image_idx += 1
                    cur_new_input_embeds.append(cur_image_features)
                    cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=cur_labels.device, dtype=cur_labels.dtype))

            cur_new_input_embeds = [x.to(self.device) for x in cur_new_input_embeds]

            cur_new_input_embeds = torch.cat(cur_new_input_embeds)
            cur_new_labels = torch.cat(cur_new_labels)

            new_input_embeds.append(cur_new_input_embeds)
            new_labels.append(cur_new_labels)

        # Truncate sequences to max length as image embeddings can make the sequence longer
        tokenizer_model_max_length = getattr(self.config, 'tokenizer_model_max_length', None)
        if tokenizer_model_max_length is not None:
            new_input_embeds = [x[:tokenizer_model_max_length] for x in new_input_embeds]
            new_labels = [x[:tokenizer_model_max_length] for x in new_labels]

        # Combine them
        max_len = max(x.shape[0] for x in new_input_embeds)
        batch_size = len(new_input_embeds)

        new_input_embeds_padded = []
        new_labels_padded = torch.full((batch_size, max_len), IGNORE_INDEX, dtype=new_labels[0].dtype, device=new_labels[0].device)
        attention_mask = torch.zeros((batch_size, max_len), dtype=attention_mask.dtype, device=attention_mask.device)
        position_ids = torch.zeros((batch_size, max_len), dtype=position_ids.dtype, device=position_ids.device)

        for i, (cur_new_embed, cur_new_labels) in enumerate(zip(new_input_embeds, new_labels)):
            cur_len = cur_new_embed.shape[0]
            if getattr(self.config, 'tokenizer_padding_side', 'right') == "left":
                new_input_embeds_padded.append(torch.cat((
                    torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device),
                    cur_new_embed
                ), dim=0))
                if cur_len > 0:
                    new_labels_padded[i, -cur_len:] = cur_new_labels
                    attention_mask[i, -cur_len:] = True
                    position_ids[i, -cur_len:] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
            else:
                new_input_embeds_padded.append(torch.cat((
                    cur_new_embed,
                    torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)
                ), dim=0))
                if cur_len > 0:
                    new_labels_padded[i, :cur_len] = cur_new_labels
                    attention_mask[i, :cur_len] = True
                    position_ids[i, :cur_len] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)

        new_input_embeds = torch.stack(new_input_embeds_padded, dim=0)

        if _labels is None:
            new_labels = None
        else:
            new_labels = new_labels_padded

        if _attention_mask is None:
            attention_mask = None
        else:
            attention_mask = attention_mask.to(dtype=_attention_mask.dtype)

        if _position_ids is None:
            position_ids = None

        return None, position_ids, 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 LlavaQwen2Model(LlavaMetaModel, Qwen2Model):
    config_class = LlavaConfig

    def __init__(self, config: Qwen2Config):
        super(LlavaQwen2Model, self).__init__(config)


class LlavaQwen2ForCausalLM(Qwen2ForCausalLM, LlavaMetaForCausalLM):
    config_class = LlavaConfig

    def __init__(self, config):
        super(Qwen2ForCausalLM, self).__init__(config)
        self.model = LlavaQwen2Model(config)
        # self.pretraining_tp = config.pretraining_tp
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

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

    def get_model(self):
        return self.model

    def forward(
        self,
        input_ids: 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,
        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,
        image_sizes: Optional[List[List[int]]] = None,
        return_dict: Optional[bool] = None,
        cache_position=None,
    ) -> Union[Tuple, CausalLMOutputWithPast]:

        if inputs_embeds is None:
            (
                input_ids,
                position_ids,
                attention_mask,
                past_key_values,
                inputs_embeds,
                labels
            ) = self.prepare_inputs_labels_for_multimodal(
                input_ids,
                position_ids,
                attention_mask,
                past_key_values,
                labels,
                images,
                image_sizes
            )

        return super().forward(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            labels=labels,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict
        )

    @torch.no_grad()
    def generate(
        self,
        inputs: Optional[torch.Tensor] = None,
        images: Optional[torch.Tensor] = None,
        image_sizes: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> Union[GenerateOutput, torch.LongTensor]:
        position_ids = kwargs.pop("position_ids", None)
        attention_mask = kwargs.pop("attention_mask", None)
        if "inputs_embeds" in kwargs:
            raise NotImplementedError("`inputs_embeds` is not supported")

        if images is not None:
            (
                inputs,
                position_ids,
                attention_mask,
                _,
                inputs_embeds,
                _
            ) = self.prepare_inputs_labels_for_multimodal(
                inputs,
                position_ids,
                attention_mask,
                None,
                None,
                images,
                image_sizes=image_sizes
            )
        else:
            inputs_embeds = self.get_model().embed_tokens(inputs)

        return super().generate(
            position_ids=position_ids,
            attention_mask=attention_mask,
            inputs_embeds=inputs_embeds,
            **kwargs
        )

    def prepare_inputs_for_generation(self, input_ids, past_key_values=None,
                                      inputs_embeds=None, **kwargs):
        images = kwargs.pop("images", None)
        image_sizes = kwargs.pop("image_sizes", None)
        inputs = super().prepare_inputs_for_generation(
            input_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, **kwargs
        )
        if images is not None:
            inputs['images'] = images
        if image_sizes is not None:
            inputs['image_sizes'] = image_sizes
        return inputs


AutoConfig.register("llava_qwen2", LlavaConfig)
AutoModelForCausalLM.register(LlavaConfig, LlavaQwen2ForCausalLM)


def tokenizer_image_token(prompt, tokenizer, image_token_index=IMAGE_TOKEN_INDEX, return_tensors=None):
    prompt_chunks = [tokenizer(chunk).input_ids for chunk in prompt.split('<image>')]

    def insert_separator(X, sep):
        return [ele for sublist in zip(X, [sep]*len(X)) for ele in sublist][:-1]

    input_ids = []
    offset = 0
    if len(prompt_chunks) > 0 and len(prompt_chunks[0]) > 0 and prompt_chunks[0][0] == tokenizer.bos_token_id:
        offset = 1
        input_ids.append(prompt_chunks[0][0])

    for x in insert_separator(prompt_chunks, [image_token_index] * (offset + 1)):
        input_ids.extend(x[offset:])

    # if return_tensors is not None:
    #     if return_tensors == 'pt':
    #         return torch.tensor(input_ids, dtype=torch.long)
    #     raise ValueError(f'Unsupported tensor type: {return_tensors}')
    return input_ids

def expand2square(pil_img, background_color):
    width, height = pil_img.size
    if width == height:
        return pil_img
    elif width > height:
        result = Image.new(pil_img.mode, (width, width), background_color)
        result.paste(pil_img, (0, (width - height) // 2))
        return result
    else:
        result = Image.new(pil_img.mode, (height, height), background_color)
        result.paste(pil_img, ((height - width) // 2, 0))
        return result