Update modeling_GOT.py

modeling_GOT.py

@@ -1,145 +1,16 @@
-from transformers import Qwen2Config, Qwen2Model, Qwen2ForCausalLM, StoppingCriteria, TextStreamer
+from transformers import AutoConfig, AutoModelForCausalLM, \
+                         Qwen2Config, Qwen2Model, Qwen2ForCausalLM, \
+                         CLIPVisionModel, CLIPImageProcessor
 from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
 from typing import List, Optional, Tuple, Union
-from transformers.cache_utils import Cache
-import requests
-from PIL import Image
-from io import BytesIO
+from transformers.cache_utils import Cache, DynamicCache
 import torch
 import torch.nn as nn
+import torch.nn.functional as F
 from torch.nn import CrossEntropyLoss
-from .got_vision_b import build_GOT_vit_b
-from torchvision import transforms
-from torchvision.transforms.functional import InterpolationMode
-import dataclasses
-from megfile import smart_open
-
-DEFAULT_IMAGE_TOKEN = "<image>"
-DEFAULT_IMAGE_PATCH_TOKEN = '<imgpad>'
-DEFAULT_IM_START_TOKEN = '<img>'
-DEFAULT_IM_END_TOKEN = '</img>'
-
-from enum import auto, Enum
-class SeparatorStyle(Enum):
-    """Different separator style."""
-    SINGLE = auto()
-    TWO = auto()
-    MPT = auto()
-
-
-@dataclasses.dataclass
-class Conversation:
-    """A class that keeps all conversation history."""
-    system: str
-    roles: List[str]
-    messages: List[List[str]]
-    offset: int
-    sep_style: SeparatorStyle = SeparatorStyle.SINGLE
-    sep: str = "<|im_end|>"
-    sep2: str = None
-    version: str = "Unknown"
-
-    skip_next: bool = False
-
-    def get_prompt(self):
-        if self.sep_style == SeparatorStyle.SINGLE:
-            ret = self.system + self.sep + '\n'
-            for role, message in self.messages:
-                if message:
-                    if type(message) is tuple:
-                        message, _, _ = message
-                    ret += role + ": " + message + self.sep
-                else:
-                    ret += role + ":"
-            return ret
-        elif self.sep_style == SeparatorStyle.TWO:
-            seps = [self.sep, self.sep2]
-            ret = self.system + seps[0]
-            for i, (role, message) in enumerate(self.messages):
-                if message:
-                    if type(message) is tuple:
-                        message, _, _ = message
-                    ret += role + ": " + message + seps[i % 2]
-                else:
-                    ret += role + ":"
-            return ret
-        if self.sep_style == SeparatorStyle.MPT:
-            if self.system:
-                ret = self.system + self.sep 
-            else:
-                ret = ''
-            for role, message in self.messages:
-                if message:
-                    if type(message) is tuple:
-                        message, _, _ = message
-                    ret += role + message + self.sep
-                else:
-                    ret += role
-            return ret
-        else:
-            raise ValueError(f"Invalid style: {self.sep_style}")
-
-
-    def append_message(self, role, message):
-        self.messages.append([role, message])
-
-    def copy(self):
-        return Conversation(
-            system=self.system,
-            roles=self.roles,
-            messages=[[x, y] for x, y in self.messages],
-            offset=self.offset,
-            sep_style=self.sep_style,
-            sep=self.sep,
-            sep2=self.sep2)
-
-
-
-class KeywordsStoppingCriteria(StoppingCriteria):
-    def __init__(self, keywords, tokenizer, input_ids):
-        self.keywords = keywords
-        self.keyword_ids = [tokenizer(keyword).input_ids for keyword in keywords]
-        self.keyword_ids = [keyword_id[0] for keyword_id in self.keyword_ids if type(keyword_id) is list and len(keyword_id) == 1]
-        self.tokenizer = tokenizer
-        self.start_len = None
-        self.input_ids = input_ids
-
-    def __call__(self, output_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
-        if self.start_len is None:
-            self.start_len = self.input_ids.shape[1]
-        else:
-            for keyword_id in self.keyword_ids:
-                if output_ids[0, -1] == keyword_id:
-                    return True
-            outputs = self.tokenizer.batch_decode(output_ids[:, self.start_len:], skip_special_tokens=True)[0]
-            for keyword in self.keywords:
-                if keyword in outputs:
-                    return True
-        return False
-    
-
-class GOTImageEvalProcessor:
-    def __init__(self, image_size=384, mean=None, std=None):
-        if mean is None:
-            mean = (0.48145466, 0.4578275, 0.40821073)
-        if std is None:
-            std = (0.26862954, 0.26130258, 0.27577711)
-
-        self.normalize = transforms.Normalize(mean, std)
-
-        self.transform = transforms.Compose(
-            [
-                transforms.Resize(
-                    (image_size, image_size), interpolation=InterpolationMode.BICUBIC
-                ),
-                transforms.ToTensor(),
-                self.normalize,
-            ]
-        )
-    def __call__(self, item):
-        return self.transform(item)
-
-
+from GOT.utils.constants import *
+from GOT.model.vision_encoder.vary_b import build_vary_vit_b
+from GOT.model.plug.blip_process import BlipImageEvalProcessor
 
 class GOTConfig(Qwen2Config):
     model_type = "GOT"
@@ -151,7 +22,7 @@ class GOTQwenModel(Qwen2Model):
     def __init__(self, config: Qwen2Config):
         super(GOTQwenModel, self).__init__(config)
 
-        self.vision_tower_high = build_GOT_vit_b()
+        self.vision_tower_high = build_vary_vit_b()
 
         self.mm_projector_vary =  nn.Linear(1024, 1024)
 
@@ -167,8 +38,13 @@ class GOTQwenModel(Qwen2Model):
         device="cuda"
     ):
 
+        # Vary old codes, not use in GOT
+        image_processor = BlipImageEvalProcessor(image_size=1024)
+        # 1024*1024
+
+        image_processor_high = BlipImageEvalProcessor(image_size=1024)
+
 
-        image_processor_high = GOTImageEvalProcessor(image_size=1024)
       
         self.vision_tower_high = self.vision_tower_high.to(dtype=dtype, device=device)
 
@@ -179,17 +55,20 @@ class GOTQwenModel(Qwen2Model):
 
         self.config.vision_tower = vision_tower
         self.config.image_token_len = image_token_len
-
+        # self.config.use_im_start_end = use_im_start_end
         self.config.use_im_start_end = True
 
         self.config.vision_select_layer = vision_select_layer
         self.config.freeze_vision_tower = freeze_vision_tower
         
         return dict(
+            image_processor=image_processor,
             image_processor_high=image_processor_high,
             image_token_len=image_token_len,
         )
          
+    # def get_input_embeddings(self, x):
+    #     return self.wte(x)
     
     def forward(
         self,
@@ -219,6 +98,9 @@ class GOTQwenModel(Qwen2Model):
 
 
         if vision_tower_high is not None and (input_ids.shape[1] != 1 or self.training) and images is not None:
+        # if True:
+            # assert type(images) is list, ValueError("To fit both interleave and conversation, images must be list of batches of images")
+            # print(im)
             use_im_start_end = getattr(self.config, "use_im_start_end", -1)
 
             vision_select_layer = getattr(self.config, "vision_select_layer", -1)
@@ -233,20 +115,31 @@ class GOTQwenModel(Qwen2Model):
 
             im_end_token = 151858
             
+
+
             image_features = []
             
-
+            print(images.shape)
             for image in images:
-                P, C, H, W = image.shape
+                P, C, H, W = image[1].shape
+                # with torch.set_grad_enabled(True):
+                #     # print(image[1].shape)
+                #     cnn_feature = vision_tower_high(image[1])
+                #     cnn_feature = cnn_feature.flatten(2).permute(0, 2, 1) # 256  1024
+                #     # image_features.append(cnn_feature)
+                # image_features_2.append(cnn_feature)
                 if P == 1:
                     with torch.set_grad_enabled(False):
-                        cnn_feature = vision_tower_high(image)
+                        # print(image[1].shape)
+                        cnn_feature = vision_tower_high(image[1])
                         cnn_feature = cnn_feature.flatten(2).permute(0, 2, 1) # 256*1024
+                        # image_features.append(cnn_feature)
+                    # image_features_2.append(cnn_feature)
                     image_feature = self.mm_projector_vary(cnn_feature)
                     image_features.append(image_feature)
 
                 else:
-                    image_patches = torch.unbind(image)
+                    image_patches = torch.unbind(image[1])
                     image_patches_features = []
                     for image_patch in image_patches:
                         image_p = torch.stack([image_patch])
@@ -256,15 +149,21 @@ class GOTQwenModel(Qwen2Model):
                         image_feature_p = self.mm_projector_vary(cnn_feature_p)
                         image_patches_features.append(image_feature_p)
                     image_feature = torch.cat(image_patches_features, dim=1)
+                    # print(P)
+                    # print(image_feature.shape)
+                    # exit()
                     image_features.append(image_feature)
 
 
+
             dummy_image_features_2 = torch.zeros(256, 1024, device=inputs_embeds.device, dtype=inputs_embeds.dtype)
+            # dummy_image_features_2 = self.mm_projector_vary(dummy_image_features_2)
             dummy_image_features = dummy_image_features_2
             use_im_start_end = True
             new_input_embeds = []
             for cur_input_ids, cur_input_embeds, cur_image_features in zip(input_ids, inputs_embeds, image_features):
                 if (cur_input_ids == im_patch_token).sum() == 0:
+                    # multimodal LLM, but the current sample is not multimodal
                     cur_input_embeds = cur_input_embeds + (0. * dummy_image_features).sum()
                     new_input_embeds.append(cur_input_embeds)
                     continue
@@ -323,6 +222,11 @@ class GOTQwenForCausalLM(Qwen2ForCausalLM):
     def get_model(self):
         return self.model
 
+    # def _set_gradient_checkpointing(self, module, value=False):
+    #     if isinstance(module, GOTQwenModel):
+    #         module.gradient_checkpointing = value
+    # @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
+    # @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
     def forward(
         self,
         input_ids: torch.LongTensor = None,
@@ -344,6 +248,12 @@ class GOTQwenForCausalLM(Qwen2ForCausalLM):
         )
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        # print(input_ids)
+        # print(len(images))
+
+        # print(inputs_embeds)
+
         outputs  = self.model(
             input_ids=input_ids,
             past_key_values=past_key_values,
@@ -358,6 +268,7 @@ class GOTQwenForCausalLM(Qwen2ForCausalLM):
             
         )
 
+
         hidden_states = outputs[0]
         logits = self.lm_head(hidden_states)
         logits = logits.float()
@@ -457,389 +368,24 @@ class GOTQwenForCausalLM(Qwen2ForCausalLM):
     ):
         config = self.get_model().config
 
-
+        # add image patch token <image>
+        # tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
         self.resize_token_embeddings(len(tokenizer))
+        # config.im_patch_token = tokenizer.convert_tokens_to_ids([DEFAULT_IMAGE_PATCH_TOKEN])[0]
 
         config.im_patch_token = 151859
 
         config.use_im_start_end = True
 
+        # add image start token <im_start> and end token <im_end>
         if config.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))
-            config.im_start_token, config.im_end_token = 151857, 151858
-
-    def load_image(self, image_file):
-        if image_file.startswith('http') or image_file.startswith('https'):
-            response = requests.get(image_file)
-            image = Image.open(BytesIO(response.content)).convert('RGB')
-        else:
-            image = Image.open(image_file).convert('RGB')
-        return image
-
-    def disable_torch_init(self):
-        """
-        Disable the redundant torch default initialization to accelerate model creation.
-        """
-        import torch
-        setattr(torch.nn.Linear, "reset_parameters", lambda self: None)
-        setattr(torch.nn.LayerNorm, "reset_parameters", lambda self: None)
-
-    def chat(self, tokenizer, image_file, ocr_type, ocr_box='', ocr_color='', render=False, save_render_file=None):
-
-        self.disable_torch_init()
-
-
-        image_processor_high =  GOTImageEvalProcessor(image_size=1024)
-
-        use_im_start_end = True
-
-        image_token_len = 256
-
-        image = self.load_image(image_file)
-
-        w, h = image.size
-        
-        if ocr_type == 'format':
-            qs = 'OCR with format: '
-        else:
-            qs = 'OCR: '
-
-        if ocr_box:
-            bbox = eval(ocr_box)
-            if len(bbox) == 2:
-                bbox[0] = int(bbox[0]/w*1000)
-                bbox[1] = int(bbox[1]/h*1000)
-            if len(bbox) == 4:
-                bbox[0] = int(bbox[0]/w*1000)
-                bbox[1] = int(bbox[1]/h*1000)
-                bbox[2] = int(bbox[2]/w*1000)
-                bbox[3] = int(bbox[3]/h*1000)
-            if ocr_type == 'format':
-                qs = str(bbox) + ' ' + 'OCR with format: '
-            else:
-                qs = str(bbox) + ' ' + 'OCR: '
-
-        if ocr_color:
-            if ocr_type == 'format':
-                qs = '[' + ocr_color + ']' + ' ' + 'OCR with format: '
-            else:
-                qs = '[' + ocr_color + ']' + ' ' + 'OCR: '
-
-        if use_im_start_end:
-            qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_PATCH_TOKEN*image_token_len + DEFAULT_IM_END_TOKEN + '\n' + qs 
-        else:
-            qs = DEFAULT_IMAGE_TOKEN + '\n' + qs
-
-
-        conv_mpt = Conversation(
-            system="""<|im_start|>system
-        You should follow the instructions carefully and explain your answers in detail.""",
-            # system = None,
-            roles=("<|im_start|>user\n", "<|im_start|>assistant\n"),
-            version="mpt",
-            messages=(),
-            offset=0,
-            sep_style=SeparatorStyle.MPT,
-            sep="<|im_end|>",
-        )
-
-        conv = conv_mpt.copy()
-        conv.append_message(conv.roles[0], qs)
-        conv.append_message(conv.roles[1], None)
-        prompt = conv.get_prompt()
-
-        print(prompt)
+            # config.im_start_token, config.im_end_token = tokenizer.convert_tokens_to_ids([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN])
 
-        inputs = tokenizer([prompt])
-
-        image_tensor_1 = image_processor_high(image)
-
-        input_ids = torch.as_tensor(inputs.input_ids).cuda()
-
-        stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
-        keywords = [stop_str]
-        stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids)
-        streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
-
-
-        with torch.autocast("cuda", dtype=torch.bfloat16):
-            output_ids = self.generate(
-                input_ids,
-                images=[image_tensor_1.unsqueeze(0).half().cuda()],
-                do_sample=False,
-                num_beams = 1,
-                no_repeat_ngram_size = 20,
-                streamer=streamer,
-                max_new_tokens=4096,
-                stopping_criteria=[stopping_criteria]
-                )
-            
-
-            if render:
-                print('==============rendering===============')
-                from .render_tools import svg_to_html, content_mmd_to_html, tik_html, translation_table
-
-                outputs = tokenizer.decode(output_ids[0, input_ids.shape[1]:]).strip()
-                
-                if outputs.endswith(stop_str):
-                    outputs = outputs[:-len(stop_str)]
-                outputs = outputs.strip()
-
-                if '**kern' in outputs:
-                    import verovio
-                    from cairosvg import svg2png
-                    import cv2
-                    import numpy as np
-                    tk = verovio.toolkit()
-                    tk.loadData(outputs)
-                    tk.setOptions({"pageWidth": 2100, "footer": 'none',
-                'barLineWidth': 0.5, 'beamMaxSlope': 15,
-                'staffLineWidth': 0.2, 'spacingStaff': 6})
-                    tk.getPageCount()
-                    svg = tk.renderToSVG()
-                    svg = svg.replace("overflow=\"inherit\"", "overflow=\"visible\"")
-
-                    svg_to_html(svg, save_render_file)
-
-                if ocr_type == 'format' and '**kern' not in outputs:
-
-                    
-                    if  '\\begin{tikzpicture}' not in outputs:
-                        html_path_2 = save_render_file
-                        right_num = outputs.count('\\right')
-                        left_num = outputs.count('\left')
-
-                        if right_num != left_num:
-                            outputs = outputs.replace('\left(', '(').replace('\\right)', ')').replace('\left[', '[').replace('\\right]', ']').replace('\left{', '{').replace('\\right}', '}').replace('\left|', '|').replace('\\right|', '|').replace('\left.', '.').replace('\\right.', '.')
-
-
-                        outputs = outputs.replace('"', '``').replace('$', '')
-
-                        outputs_list = outputs.split('\n')
-                        gt= ''
-                        for out in outputs_list:
-                            gt +=  '"' + out.replace('\\', '\\\\') + r'\n' + '"' + '+' + '\n' 
-                        
-                        gt = gt[:-2]
-
-
-                        lines = content_mmd_to_html
-                        lines = lines.split("const text =")
-                        new_web = lines[0] + 'const text ='  + gt  + lines[1]
-
-                    else:
-                        html_path_2 = save_render_file
-                        outputs = outputs.translate(translation_table)
-                        outputs_list = outputs.split('\n')
-                        gt= ''
-                        for out in outputs_list:
-                            if out:
-                                if '\\begin{tikzpicture}' not in out and '\\end{tikzpicture}' not in out:
-                                    while out[-1] == ' ':
-                                        out = out[:-1]
-                                        if out is None:
-                                            break
-        
-                                    if out:
-                                        if out[-1] != ';':
-                                            gt += out[:-1] + ';\n'
-                                        else:
-                                            gt += out + '\n'
-                                else:
-                                    gt += out + '\n'
-
-
-                        lines = tik_html
-                        lines = lines.split("const text =")
-                        new_web = lines[0] + gt + lines[1]
-
-                    with smart_open(html_path_2, 'w') as web_f_new:
-                        web_f_new.write(new_web)
-
-    def dynamic_preprocess(self, image, min_num=1, max_num=6, image_size=1024, use_thumbnail=True):
-        
-        def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
-            best_ratio_diff = float('inf')
-            best_ratio = (1, 1)
-            area = width * height
-            for ratio in target_ratios:
-                target_aspect_ratio = ratio[0] / ratio[1]
-                ratio_diff = abs(aspect_ratio - target_aspect_ratio)
-                if ratio_diff < best_ratio_diff:
-                    best_ratio_diff = ratio_diff
-                    best_ratio = ratio
-                elif ratio_diff == best_ratio_diff:
-                    if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
-                        best_ratio = ratio
-            # print(f'width: {width}, height: {height}, best_ratio: {best_ratio}')
-            return best_ratio
-        
-        orig_width, orig_height = image.size
-        aspect_ratio = orig_width / orig_height
-
-        # calculate the existing image aspect ratio
-        target_ratios = set(
-            (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
-            i * j <= max_num and i * j >= min_num)
-        # print(target_ratios)
-        target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
-
-        # find the closest aspect ratio to the target
-        target_aspect_ratio = find_closest_aspect_ratio(
-            aspect_ratio, target_ratios, orig_width, orig_height, image_size)
-
-        # print(target_aspect_ratio)
-        # calculate the target width and height
-        target_width = image_size * target_aspect_ratio[0]
-        target_height = image_size * target_aspect_ratio[1]
-        blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
-
-        # resize the image
-        resized_img = image.resize((target_width, target_height))
-        processed_images = []
-        for i in range(blocks):
-            box = (
-                (i % (target_width // image_size)) * image_size,
-                (i // (target_width // image_size)) * image_size,
-                ((i % (target_width // image_size)) + 1) * image_size,
-                ((i // (target_width // image_size)) + 1) * image_size
-            )
-            # split the image
-            split_img = resized_img.crop(box)
-            processed_images.append(split_img)
-        assert len(processed_images) == blocks
-        if use_thumbnail and len(processed_images) != 1:
-            thumbnail_img = image.resize((image_size, image_size))
-            processed_images.append(thumbnail_img)
-        return processed_images
-
-
-    def chat_plus(self, tokenizer, image_file_list, render=False, save_render_file=None):
-        # Model
-        self.disable_torch_init()
-        multi_page=False
-
-
-        image_processor_high =  GOTImageEvalProcessor(image_size=1024)
-
-        use_im_start_end = True
-
-
-        image_token_len = 256
-
-        image_list = []
-
-        if len(image_file_list)>1:
-            multi_page = True
-
-        if multi_page:
-            qs = 'OCR with format across multi pages: '
-            # only for png files
-            import glob
-            from natsort import natsorted
-            # patches = glob.glob(image_file + '/*png')
-            patches = image_file_list
-            patches = natsorted(patches)
-            sub_images = []
-            for sub_image in patches:
-                sub_images.append(self.load_image(sub_image))
-
-            ll = len(patches)
-            print(patches)
-            print("len ll: ", ll)
-
-        else:
-            qs = 'OCR with format upon the patch reference: '
-            img = self.load_image(image_file_list[0])
-            sub_images = self.dynamic_preprocess(img)
-            ll = len(sub_images)
-
-        for image in sub_images:
-            image_tensor_1 = image_processor_high(image)
-            image_list.append(image_tensor_1)
-
-
-        image_list = torch.stack(image_list)
-
-        print('====new images batch size======:  ',image_list.shape)
-
-
-        if use_im_start_end:
-            qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_PATCH_TOKEN*image_token_len*ll + DEFAULT_IM_END_TOKEN + '\n' + qs 
-        else:
-            qs = DEFAULT_IMAGE_TOKEN + '\n' + qs
-
-
-        conv_mpt = Conversation(
-            system="""<|im_start|>system
-        You should follow the instructions carefully and explain your answers in detail.""",
-            # system = None,
-            roles=("<|im_start|>user\n", "<|im_start|>assistant\n"),
-            version="mpt",
-            messages=(),
-            offset=0,
-            sep_style=SeparatorStyle.MPT,
-            sep="<|im_end|>",
-        )
-
-        conv = conv_mpt.copy()
-        conv.append_message(conv.roles[0], qs)
-        conv.append_message(conv.roles[1], None)
-        prompt = conv.get_prompt()
-
-        print(prompt)
-
-        inputs = tokenizer([prompt])
-
-        input_ids = torch.as_tensor(inputs.input_ids).cuda()
-
-        stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
-        keywords = [stop_str]
-        stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids)
-        streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
-
-
-        with torch.autocast("cuda", dtype=torch.bfloat16):
-            output_ids = self.generate(
-                input_ids,
-                images=[image_list.half().cuda()],
-                do_sample=False,
-                num_beams = 1,
-                # no_repeat_ngram_size = 20,
-                streamer=streamer,
-                max_new_tokens=4096,
-                stopping_criteria=[stopping_criteria]
-                )
-            
-        if render:
-            print('==============rendering===============')
-            from .render_tools import content_mmd_to_html
-            outputs = tokenizer.decode(output_ids[0, input_ids.shape[1]:]).strip()
-            
-            if outputs.endswith(stop_str):
-                outputs = outputs[:-len(stop_str)]
-            outputs = outputs.strip()
-
-            html_path_2 = save_render_file
-            right_num = outputs.count('\\right')
-            left_num = outputs.count('\left')
-
-            if right_num != left_num:
-                outputs = outputs.replace('\left(', '(').replace('\\right)', ')').replace('\left[', '[').replace('\\right]', ']').replace('\left{', '{').replace('\\right}', '}').replace('\left|', '|').replace('\\right|', '|').replace('\left.', '.').replace('\\right.', '.')
+            config.im_start_token, config.im_end_token = 151857, 151858
 
 
-            outputs = outputs.replace('"', '``').replace('$', '')
+AutoConfig.register("GOT", GOTConfig)
+AutoModelForCausalLM.register(GOTConfig, GOTQwenForCausalLM)
 
-            outputs_list = outputs.split('\n')
-            gt= ''
-            for out in outputs_list:
-                gt +=  '"' + out.replace('\\', '\\\\') + r'\n' + '"' + '+' + '\n' 
-            
-            gt = gt[:-2]
-
-            lines = content_mmd_to_html
-            lines = lines.split("const text =")
-            new_web = lines[0] + 'const text ='  + gt  + lines[1]
-                
-            with smart_open(html_path_2, 'w') as web_f_new:
-                web_f_new.write(new_web)