tonibirat/MinerU2.5-2509-1.2B_8GB_Optiomation_VM_Azure

license: apache-2.0 language:

• ne
• en tags:
• ocr
• devanagari
• qwen2-vl
• bitsandbytes
• 4-bit
• onnx
• cpu-optimized metrics:
• latency
• memory-usage pipeline_tag: visual-question-answering

SynapOCR MinerU-Hybrid-8GB Optimization

This repository contains optimized models for the SynapOCR MinerU Hybrid Auto Engine, specifically tailored for low-resource CPU-only environments (e.g., 8GB RAM Cloud VMs / Azure VM).

🚀 Key Optimizations

1. VLM Quantization (VLM_4bit/)

• Base Model: Qwen2-VL-1.2B (part of MinerU 2.5)
• Quantization: 4-bit NormalFloat (NF4) via bitsandbytes with Double Quantization.
• RAM Footprint: Reduced from ~2.3 GB to ~800 MB.
• Performance: Optimized for CPU inference under tight memory constraints while maintaining high accuracy for Devanagari (Nepali) and English text.

2. Layout & Detection ONNX (ONNX/Layout/ & ONNX/MFD/)

• Architecture: YOLO-based structural analysis models from PDF-Extract-Kit.
• Optimization: Exported to ONNX Runtime format.
• Benefits:
  • ~2x faster detection latency compared to PyTorch on CPU.
  • Reduced overhead and improved stability across different OS environments.

3. OCR Weights (ONNX/OCR/)

• Engine: PaddleOCR-v3/v4 weights.
• Format: ONNX-compatible weights for the detection and recognition heads.

💻 Hardware Requirements

🛠️ Usage with SynapOCR

This model is designed to be used with the SynapOCR MinerU Hybrid Auto Engine.

Installation

pip install bitsandbytes onnxruntime transformers accelerate

Loading the 4-bit VLM

from transformers import AutoModelForVision2Seq, BitsAndBytesConfig
import torch

# Configuration for 4-bit loading
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_use_double_quant=True
)

model = AutoModelForVision2Seq.from_pretrained(
    "tonibirat/MinerU2.5-2509-1.2B_8GB_Optiomation_VM_Azure",
    subfolder="VLM_4bit",
    quantization_config=bnb_config,
    device_map="cpu"
)

⚖️ License

The weights are distributed under the Apache 2.0 License, inherited from the original MinerU components and Qwen2-VL architecture.

Maintained by: ToniBirat7 Project: SynapOCR MinerU Hybrid Auto Engine