Durgaram/dots.mocr-4bit

Dots MOCR – 4-bit Quantized (NF4)

🔍 Introduction

This repository provides a 4-bit quantized version of dots.mocr, optimized using BitsAndBytes (NF4 precision) for efficient, low-memory inference.

The original model is a powerful multimodal OCR system capable of:

• Document parsing
• Layout understanding
• Multilingual OCR
• Structured outputs (JSON / Markdown / SVG)

This version enables deployment on low-VRAM GPUs while maintaining strong performance.

---

⚙️ Key Features

• 4-bit quantization (NF4)
• Reduced VRAM usage (~70–80%)
• Faster inference
• Compatible with Hugging Face Transformers
• Supports OCR and document parsing
• Suitable for edge and local deployments

---

🛠️ Installation (Base Setup)

⚠️ This model depends on the original dots.mocr repository.

conda create -n dots_mocr python=3.12
conda activate dots_mocr

git clone https://github.com/rednote-hilab/dots.mocr.git
cd dots.mocr

pip install -e .
pip install flash-attn==2.8.0.post2

---

🚀 Usage (Quantized Inference)

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

model_id = "rednote-hilab/dots.mocr"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype="float16",
    bnb_4bit_use_double_quant=True
)

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True
)

# Example usage
inputs = tokenizer("Extract text from image", return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

📊 Quantization Details

Parameter	Value
Precision	4-bit
Quant Type	NF4
Compute Dtype	float16
Double Quant	Enabled
Library	BitsAndBytes

---

📌 Use Cases

• Document OCR
• PDF parsing
• Layout detection
• Structured data extraction
• AI-powered document understanding
• Edge deployment of large OCR models

---

⚠️ Limitations

• Slight accuracy drop compared to full precision
• GPU recommended for optimal performance
• Some layers remain in higher precision
• Not fully optimized for CPU inference

---

🔮 Future Work

• GGUF conversion for CPU inference
• FlashAttention optimization improvements
• Integration with full OCR pipelines
• Web UI (Gradio / Streamlit demo)
• Benchmark comparisons (VRAM vs accuracy)

---

🙌 Acknowledgement

• Base Model: rednote-hilab/dots.mocr
• Quantization: BitsAndBytes
• Framework: Hugging Face Transformers

---

📄 License

MIT License