GitLab

- Move environment variable settings for NPU, MPS, and other configurations to the global scope in doc_analyze_by_custom_model.py
- Remove redundant environment variable settings in pdf_extract_kit.py
- This change ensures consistent configuration across the application and avoids potential conflicts or duplicate settings

- Remove redundant imports for StructTableModel and TableMasterPaddleModel
- Reorder imports to group related modules together
- Update import structure for better readability and maintainability

- Remove unused utility functions
- Update import statements for better readability
- Add conditional imports for Ascend plugin
- Refactor table model initialization to support NPU

- Increase batch size from 8 to 256 for language detection inference
- Add timing measurement for language detection process

- Update batch ratio calculation logic to better utilize available GPU memory
- Improve logging for all GPU memory sizes

- Update batch ratio calculation for GPU memory range
- Increase upper limit for batch ratio 16 from 24 to 32 GB

- Update conditions for batch ratio assignment:
  -8 <= gpu_memory < 10: batch_ratio = 2 - 10 <= gpu_memory <= 12: batch_ratio =4
- This fix ensures proper batch ratio selection for GPU memory sizes

- Improve batch ratio calculation based on GPU memory
- Enhance performance for devices with 8GB or more VRAM

- Reduce batch_ratio by 1 for better performance and stability
- This change ensures more consistent memory usage when processing documents

refactor(magic_pdf): adjust VRAM allocation and MFR batch size- Update VRAM allocation logic to use 'VIRTUAL_VRAM_SIZE' environment variable
- Reduce MFR (Math Formula Recognition) batch size from 64 to 32

- Update GPU memory check and batch ratio calculation logic
- Add support for virtual VRAM size environment variable
- Improve logging for GPU memory and batch ratio

- Reduce YOLO_LAYOUT_BASE_BATCH_SIZE from 4 to 1
- Simplify batch ratio calculation for formula detection
- Remove unused conditional logic in batch ratio determination

- Add key length validation for ONNX model initialization
- Move import statements to the top of the file
- Wrap model initialization in a try-except block for better error handling
- Refactor code to improve readability and maintainability

- Commented out the original batch ratio calculation
- Set a fixed batch ratio of 2 for GPUs with less than 8 GB memory
- Increased batch ratio to 4 for GPUs with 8 GB or more memory

- Import get_device function from magic_pdf.libs.config_reader- Update RapidTableModel initialization to include device parameter for Unitable model

- Modify the batch analyze process to handle the rapid table model's output
- Add logic_points variable to capture additional output from rapid table prediction

- Update RapidTable dependency to version 1.0.3
- Add support for sub-models in RapidTable
- Update magic-pdf configuration to include table sub-model
- Modify table model initialization to support sub-models
- Update table prediction logic to handle new output format

- Adjust end_page_id calculation to prevent IndexError when accessing pages
- Enhance error handling in LLM post-processing by specifically catching JSONDecodeError

- Add support for NPU (Neural Processing Unit) when available
- Implement batch analysis for GPU and NPU devices
- Optimize memory usage and improve performance
- Update logging and error handling

- Merge title blocks that are close to each other horizontally
- Adjust line insertion logic for title blocks- Increase image size and decrease confidence threshold for layout detection
- Update DocLayoutYOLO model weights
- Refactor drawing of bounding boxes for different block types

- Add MPS support for Apple Silicon devices
- Implement empty_cache() for MPS devices
- Set PYTORCH_ENABLE_MPS_FALLBACK environment variable
- Adjust MFR model device allocation for MPS

- Remove conditional logic for OCR engine selection
- Always use RapidOCR as the OCR engine
- Simplify the __init__ method by removing unused code

- Remove YOLO v11 language detection model from model_configs.yaml
- Update language detection utils to use a fixed model path instead of dynamic configuration
- Remove unused model weight parameter for YOLO v11 language detection

- Remove LangDetectMode and related conditional logic
- Use a single model weight for language detection
- Add logging for language detection results
- Update model initialization and prediction methods

- Add language detection model initialization and integration
- Update model list to include language detection
- Refactor language detection utils for better model management

- Add separate models for Chinese/Japanese and English/French/German detection
- Implement mode-based detection to use appropriate models for different languages
- Update language detection process to use higher DPI for better accuracy
- Modify model initialization and prediction logic to support new language-specific models

- Add check for empty OCR result when using PaddleOCR model
- Assign None to ocr_result if no text is detected, preventing further errors

- Add missing 'else' statement in OCR model selection logic
- Ensure consistent formatting of 'if' statements for better readability
- Remove unnecessary empty line in the 'app.py' file

- Remove logger.info() call for additional_ocr_params to reduce log verbosity

- Implement ONNXModelSingleton to manage ONNX models
- Modify ModifiedPaddleOCR to use ONNX models on ARM CPUs without CUDA
- Update RapidTableModel to use RapidOCR with ONNXRuntime on CPU
- Add rapidocr_onnxruntime dependency in setup.py

- Remove use_npu variable initialization
- Comment out device assignment and npu check
- Comment out use_npu parameter in ModifiedPaddleOCR constructor

- Update `clean_memory.py` to use `torch_npu.npu` instead of `torch.npu`
- Update `model_utils.py` to use `torch_npu.npu` instead of `torch.npu`
- Simplify NPU availability check and bfloat16 support in `pdf_parse_union_core_v2.py`

- Update clean_memory function to support both CUDA and NPU devices
- Implement get_device function to centralize device selection logic
- Modify model initialization and memory cleaning to use the selected device
- Update RapidTableModel to support both RapidOCR and PaddleOCR engines