Merge pull request #3213 from myhloli/dev

fix: add support for additional keyword arguments in modeling_unimer_mbart.py to adaptation transformers 4.54.1

README.md

@@ -43,6 +43,15 @@
 </div>
 
 # Changelog
+- 2025/07/30 version 2.1.9 Released
+  - `transformers` 4.54.1 version adaptation
+- 2025/07/28 version 2.1.8 Released
+  - `sglang` 0.4.9.post5 version adaptation
+- 2025/07/27 version 2.1.7 Released
+  - `transformers` 4.54.0 version adaptation
+- 2025/07/26 2.1.6 Released
+  - Fixed table parsing issues in handwritten documents when using `vlm` backend
+  - Fixed visualization box position drift issue when document is rotated #3175
 - 2025/07/24 2.1.5 Released
   - `sglang` 0.4.9 version adaptation, synchronously upgrading the dockerfile base image to sglang 0.4.9.post3
 - 2025/07/23 2.1.4 Released

README_zh-CN.md

@@ -43,13 +43,22 @@
 </div>
 
 # 更新记录
-- 2025/07/24 2.1.5发布
+- 2025/07/30 2.1.9 发布
+  - `transformers` 4.54.1 版本适配
+- 2025/07/28 2.1.8 发布
+  - `sglang` 0.4.9.post5 版本适配
+- 2025/07/27 2.1.7 发布
+  - `transformers` 4.54.0 版本适配
+- 2025/07/26 2.1.6 发布
+  - 修复`vlm`后端解析部分手写文档时的表格异常问题
+  - 修复文档旋转时可视化框位置漂移问题 #3175
+- 2025/07/24 2.1.5 发布
   - `sglang` 0.4.9 版本适配，同步升级dockerfile基础镜像为sglang 0.4.9.post3
-- 2025/07/23 2.1.4发布
+- 2025/07/23 2.1.4 发布
   - bug修复
     - 修复`pipeline`后端中`MFR`步骤在某些情况下显存消耗过大的问题 #2771
     - 修复某些情况下`image`/`table`与`caption`/`footnote`匹配不准确的问题 #3129
-- 2025/07/16 2.1.1发布
+- 2025/07/16 2.1.1 发布
   - bug修复 
     - 修复`pipeline`在某些情况可能发生的文本块内容丢失问题 #3005
     - 修复`sglang-client`需要安装`torch`等不必要的包的问题 #2968
@@ -57,7 +66,7 @@
   - 易用性更新
     - 更新`compose.yaml`，便于用户直接启动`sglang-server`、`mineru-api`、`mineru-gradio`服务
     - 启用全新的[在线文档站点](https://opendatalab.github.io/MinerU/zh/)，简化readme，提供更好的文档体验
-- 2025/07/05 2.1.0发布
+- 2025/07/05 2.1.0 发布
   - 这是 MinerU 2 的第一个大版本更新，包含了大量新功能和改进，包含众多性能优化、体验优化和bug修复，具体更新内容如下： 
   - 性能优化： 
     - 大幅提升某些特定分辨率（长边2000像素左右）文档的预处理速度

docker/china/Dockerfile

@@ -1,7 +1,7 @@
 # Use the official sglang image
-FROM lmsysorg/sglang:v0.4.9.post3-cu126
+FROM lmsysorg/sglang:v0.4.9.post6-cu126
 # For blackwell GPU, use the following line instead:
-# FROM lmsysorg/sglang:v0.4.9.post3-cu128-b200
+# FROM lmsysorg/sglang:v0.4.9.post6-cu128-b200
 
 # Install libgl for opencv support & Noto fonts for Chinese characters
 RUN apt-get update && \

docker/global/Dockerfile

@@ -1,7 +1,7 @@
 # Use the official sglang image
-FROM lmsysorg/sglang:v0.4.9.post3-cu126
+FROM lmsysorg/sglang:v0.4.9.post6-cu126
 # For blackwell GPU, use the following line instead:
-# FROM lmsysorg/sglang:v0.4.9.post3-cu128-b200
+# FROM lmsysorg/sglang:v0.4.9.post6-cu128-b200
 
 # Install libgl for opencv support & Noto fonts for Chinese characters
 RUN apt-get update && \

docs/en/quick_start/docker_deployment.md

@@ -10,8 +10,8 @@ docker build -t mineru-sglang:latest -f Dockerfile .
 ```
 
 > [!TIP]
-> The [Dockerfile](https://github.com/opendatalab/MinerU/blob/master/docker/global/Dockerfile) uses `lmsysorg/sglang:v0.4.9.post3-cu126` as the base image by default, supporting Turing/Ampere/Ada Lovelace/Hopper platforms.
-> If you are using the newer `Blackwell` platform, please modify the base image to `lmsysorg/sglang:v0.4.9.post3-cu128-b200` before executing the build operation.
+> The [Dockerfile](https://github.com/opendatalab/MinerU/blob/master/docker/global/Dockerfile) uses `lmsysorg/sglang:v0.4.9.post6-cu126` as the base image by default, supporting Turing/Ampere/Ada Lovelace/Hopper platforms.
+> If you are using the newer `Blackwell` platform, please modify the base image to `lmsysorg/sglang:v0.4.9.post6-cu128-b200` before executing the build operation.
 
 ## Docker Description
 

docs/zh/quick_start/docker_deployment.md

@@ -10,8 +10,8 @@ docker build -t mineru-sglang:latest -f Dockerfile .
 ```
 
 > [!TIP]
-> [Dockerfile](https://github.com/opendatalab/MinerU/blob/master/docker/china/Dockerfile)默认使用`lmsysorg/sglang:v0.4.9.post3-cu126`作为基础镜像，支持Turing/Ampere/Ada Lovelace/Hopper平台，
-> 如您使用较新的`Blackwell`平台，请将基础镜像修改为`lmsysorg/sglang:v0.4.9.post3-cu128-b200` 再执行build操作。
+> [Dockerfile](https://github.com/opendatalab/MinerU/blob/master/docker/china/Dockerfile)默认使用`lmsysorg/sglang:v0.4.9.post6-cu126`作为基础镜像，支持Turing/Ampere/Ada Lovelace/Hopper平台，
+> 如您使用较新的`Blackwell`平台，请将基础镜像修改为`lmsysorg/sglang:v0.4.9.post6-cu128-b200` 再执行build操作。
 
 ## Docker说明
 

mineru/backend/pipeline/batch_analyze.py

@@ -256,9 +256,12 @@ class BatchAnalyze:
                 html_code, table_cell_bboxes, logic_points, elapse = table_model.predict(table_res_dict['table_img'])
                 # 判断是否返回正常
                 if html_code:
-                    expected_ending = html_code.strip().endswith('</html>') or html_code.strip().endswith('</table>')
-                    if expected_ending:
-                        table_res_dict['table_res']['html'] = html_code
+                    # 检查html_code是否包含'<table>'和'</table>'
+                    if '<table>' in html_code and '</table>' in html_code:
+                        # 选用<table>到</table>的内容，放入table_res_dict['table_res']['html']
+                        start_index = html_code.find('<table>')
+                        end_index = html_code.rfind('</table>') + len('</table>')
+                        table_res_dict['table_res']['html'] = html_code[start_index:end_index]
                     else:
                         logger.warning(
                             'table recognition processing fails, not found expected HTML table end'

mineru/backend/vlm/vlm_magic_model.py

@@ -5,7 +5,7 @@ from loguru import logger
 
 from mineru.utils.enum_class import ContentType, BlockType, SplitFlag
 from mineru.backend.vlm.vlm_middle_json_mkcontent import merge_para_with_text
-from mineru.utils.format_utils import convert_otsl_to_html
+from mineru.utils.format_utils import block_content_to_html
 from mineru.utils.magic_model_utils import reduct_overlap, tie_up_category_by_distance_v3
 
 
@@ -40,6 +40,10 @@ class MagicModel:
                 block_type = block_info[1].strip()
                 block_content = block_info[2].strip()
 
+                # 如果bbox是0,0,999,999，且type为text，按notes增加表格处理
+                if x1 == 0 and y1 == 0 and x2 == 999 and y2 == 999 and block_type == "text":
+                    block_content = block_content_to_html(block_content)
+
                 # print(f"坐标: {block_bbox}")
                 # print(f"类型: {block_type}")
                 # print(f"内容: {block_content}")
@@ -77,16 +81,7 @@ class MagicModel:
                     "type": span_type,
                 }
                 if span_type == ContentType.TABLE:
-                    if "<fcel>" in block_content or "<ecel>" in block_content:
-                        lines = block_content.split("\n\n")
-                        new_lines = []
-                        for line in lines:
-                            if "<fcel>" in line or "<ecel>" in line:
-                                line = convert_otsl_to_html(line)
-                            new_lines.append(line)
-                        span["html"] = "\n\n".join(new_lines)
-                    else:
-                        span["html"] = block_content
+                    span["html"] = block_content_to_html(block_content)
             elif span_type in [ContentType.INTERLINE_EQUATION]:
                 span = {
                     "bbox": block_bbox,

mineru/model/mfr/unimernet/unimernet_hf/unimer_mbart/modeling_unimer_mbart.py

@@ -1416,7 +1416,11 @@ class UnimerMBartDecoder(UnimerMBartPreTrainedModel):
             raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
 
         # past_key_values_length
-        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+        # past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+        past_key_values_length = 0
+        if past_key_values is not None:
+            if isinstance(past_key_values, (list, tuple)) and past_key_values:
+                past_key_values_length = past_key_values[0][0].shape[2]
 
         if inputs_embeds is None:
             inputs_embeds = self.embed_tokens(input_ids)
@@ -1501,7 +1505,12 @@ class UnimerMBartDecoder(UnimerMBartPreTrainedModel):
                 if dropout_probability < self.layerdrop:
                     continue
 
-            past_key_value = past_key_values[idx] if past_key_values is not None else None
+            # past_key_value = past_key_values[idx] if past_key_values is not None else None
+            past_key_value = past_key_values[idx] if (
+                    past_key_values is not None and
+                    isinstance(past_key_values, (list, tuple)) and
+                    idx < len(past_key_values)
+            ) else None
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
@@ -2182,6 +2191,7 @@ class UnimerMBartForCausalLM(UnimerMBartPreTrainedModel, GenerationMixin):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         count_gt: Optional[torch.LongTensor] = None,
+        **kwargs,
     ) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
         r"""
         Args:

mineru/model/vlm_sglang_model/image_processor.py

@@ -56,8 +56,8 @@ def process_anyres_image(image, processor, grid_pinpoints):
 
 
 class Mineru2ImageProcessor(BaseProcessor):
-    def __init__(self, hf_config, server_args, _processor):
-        super().__init__(hf_config, server_args, _processor)
+    def __init__(self, hf_config, server_args, _processor, *args, **kwargs):
+        super().__init__(hf_config, server_args, _processor, *args, **kwargs)
 
     @staticmethod
     def _process_single_image_task(

mineru/utils/draw_bbox.py

@@ -2,21 +2,64 @@ import json
 from io import BytesIO
 
 from loguru import logger
-from pypdf import PdfReader, PdfWriter
+from pypdf import PdfReader, PdfWriter, PageObject
 from reportlab.pdfgen import canvas
 
 from .enum_class import BlockType, ContentType
 
 
+def cal_canvas_rect(page, bbox):
+    """
+    Calculate the rectangle coordinates on the canvas based on the original PDF page and bounding box.
+
+    Args:
+        page: A PyPDF2 Page object representing a single page in the PDF.
+        bbox: [x0, y0, x1, y1] representing the bounding box coordinates.
+
+    Returns:
+        rect: [x0, y0, width, height] representing the rectangle coordinates on the canvas.
+    """
+    page_width, page_height = float(page.cropbox[2]), float(page.cropbox[3])
+    
+    actual_width = page_width    # The width of the final PDF display
+    actual_height = page_height  # The height of the final PDF display
+    
+    rotation = page.get("/Rotate", 0)
+    rotation = rotation % 360
+    
+    if rotation in [90, 270]:
+        # PDF is rotated 90 degrees or 270 degrees, and the width and height need to be swapped
+        actual_width, actual_height = actual_height, actual_width
+        
+    x0, y0, x1, y1 = bbox
+    rect_w = abs(x1 - x0)
+    rect_h = abs(y1 - y0)
+    
+    if 270 == rotation:
+        rect_w, rect_h = rect_h, rect_w
+        x0 = actual_height - y1
+        y0 = actual_width - x1
+    elif 180 == rotation:
+        x0 = page_width - x1
+        y0 = y0
+    elif 90 == rotation:
+        rect_w, rect_h = rect_h, rect_w
+        x0, y0 = y0, x0 
+    else:
+        # 0 == rotation:
+        x0 = x0
+        y0 = page_height - y1
+    
+    rect = [x0, y0, rect_w, rect_h]        
+    return rect
+
+
 def draw_bbox_without_number(i, bbox_list, page, c, rgb_config, fill_config):
     new_rgb = [float(color) / 255 for color in rgb_config]
     page_data = bbox_list[i]
-    page_width, page_height = page.cropbox[2], page.cropbox[3]
 
     for bbox in page_data:
-        width = bbox[2] - bbox[0]
-        height = bbox[3] - bbox[1]
-        rect = [bbox[0], page_height - bbox[3], width, height]  # Define the rectangle
+        rect = cal_canvas_rect(page, bbox)  # Define the rectangle  
 
         if fill_config:  # filled rectangle
             c.setFillColorRGB(new_rgb[0], new_rgb[1], new_rgb[2], 0.3)
@@ -35,10 +78,8 @@ def draw_bbox_with_number(i, bbox_list, page, c, rgb_config, fill_config, draw_b
 
     for j, bbox in enumerate(page_data):
         # 确保bbox的每个元素都是float
-        x0, y0, x1, y1 = map(float, bbox)
-        width = x1 - x0
-        height = y1 - y0
-        rect = [x0, page_height - y1, width, height]
+        rect = cal_canvas_rect(page, bbox)  # Define the rectangle  
+        
         if draw_bbox:
             if fill_config:
                 c.setFillColorRGB(*new_rgb, 0.3)
@@ -48,8 +89,23 @@ def draw_bbox_with_number(i, bbox_list, page, c, rgb_config, fill_config, draw_b
                 c.rect(rect[0], rect[1], rect[2], rect[3], stroke=1, fill=0)
         c.setFillColorRGB(*new_rgb, 1.0)
         c.setFontSize(size=10)
-        # 这里也要用float
-        c.drawString(x1 + 2, page_height - y0 - 10, str(j + 1))
+        
+        c.saveState()
+        rotation = page.get("/Rotate", 0)
+        rotation = rotation % 360
+    
+        if 0 == rotation:
+            c.translate(rect[0] + rect[2] + 2, rect[1] + rect[3] - 10)
+        elif 90 == rotation:
+            c.translate(rect[0] + 10, rect[1] + rect[3] + 2)
+        elif 180 == rotation:
+            c.translate(rect[0] - 2, rect[1] + 10)
+        elif 270 == rotation:
+            c.translate(rect[0] + rect[2] - 10, rect[1] - 2)
+            
+        c.rotate(rotation)
+        c.drawString(0, 0, str(j + 1))
+        c.restoreState()
 
     return c
 
@@ -185,6 +241,9 @@ def draw_layout_bbox(pdf_info, pdf_bytes, out_path, filename):
 
         # 添加检查确保overlay_pdf.pages不为空
         if len(overlay_pdf.pages) > 0:
+            new_page = PageObject(pdf=None)
+            new_page.update(page)
+            page = new_page
             page.merge_page(overlay_pdf.pages[0])
         else:
             # 记录日志并继续处理下一个页面
@@ -300,6 +359,9 @@ def draw_span_bbox(pdf_info, pdf_bytes, out_path, filename):
 
         # 添加检查确保overlay_pdf.pages不为空
         if len(overlay_pdf.pages) > 0:
+            new_page = PageObject(pdf=None)
+            new_page.update(page)
+            page = new_page
             page.merge_page(overlay_pdf.pages[0])
         else:
             # 记录日志并继续处理下一个页面

mineru/utils/format_utils.py

@@ -317,3 +317,26 @@ def convert_otsl_to_html(otsl_content: str):
     )
 
     return export_to_html(table_data)
+
+
+def block_content_to_html(block_content: str) -> str:
+    """
+    Converts block content containing OTSL (Open Table Structure Language) tags into HTML.
+
+    This function processes a block of text, splitting it into lines and converting any lines
+    containing OTSL table tags (e.g., <fcel>, <ecel>) into HTML tables. Lines without these
+    tags are left unchanged.
+
+    Parameters:
+        block_content (str): A string containing block content with potential OTSL tags.
+
+    Returns:
+        str: The processed block content with OTSL tags converted to HTML tables.
+    """
+    lines = block_content.split("\n\n")
+    new_lines = []
+    for line in lines:
+        if "<fcel>" in line or "<ecel>" in line:
+            line = convert_otsl_to_html(line)
+        new_lines.append(line)
+    return "\n\n".join(new_lines)

mineru/version.py

@@ -1 +1 @@
-__version__ = "2.1.4"
+__version__ = "2.1.8"