deepseekvl2 pdf ocr.stub

---

id: deepseekvl2-pdf-ocr name: deepseekvl2-PDF-OCR-private repo_path: ~/Documents/sw30labs/repos/deepseekvl2-PDF-OCR-private remote_url: https://github.com/sw30labs/deepseekvl2-PDF-OCR-private.git primary_language: Python framework: DeepSeek-VL2 + PyTorch + NVIDIA CUDA ingested: 2026-04-13 last_commit_date: 2025-02-09

Purpose

Local PDF OCR engine using state-of-the-art DeepSeek-VL2 mixture-of-experts vision-language model on NVIDIA CUDA GPUs. Supports single and batch PDF processing with automatic multi-GPU parallel distribution for faster extraction.

Architecture

• Single + batch modes: Processes one PDF or entire directory with auto-GPU distribution
• Multi-GPU parallel processing: Custom modified DeepSeek-VL2 inference.py for device management and tensor handling
• Patching system: apply_patches.py automatically backs up and applies multi-GPU modifications
• PDF → Image conversion: pdf2image for page extraction to temporary JPEG sequence
• Markdown output: Structured markdown with page sections and extracted text
• Model variants: tiny (~8GB), small (~12GB), base (~16GB) with configurable batch sizes
• Progress tracking: Per-PDF status and detailed logging
• Cleanup: Automatic temporary image removal post-processing

Dependencies

• torch - PyTorch deep learning framework (CUDA)
• transformers - DeepSeek-VL2 model loading and inference
• pdf2image - PDF page to image conversion
• Pillow - image processing and manipulation
• tqdm - progress bar tracking
• Python stdlib (argparse, os, subprocess, pathlib)

Key Features

• State-of-the-art DeepSeek-VL2 MoE vision-language model
• Local processing (no cloud API, full privacy)
• Single PDF or batch directory processing
• Multi-GPU parallel processing with custom modifications
• Model variant selection (tiny, small, base)
• CUDA-optimized inference
• Markdown output with page structure
• Progress tracking and detailed logging
• Memory optimization strategies
• Error handling with recovery

Tags

• python
• ocr
• local-inference
• torch
• security
• developer-tools

Entities and Topics

• local-inference-mlx
• developer-tools
• ollama-pdf2markdown