Qwen2-VL-2B-Instruct-OpenVINO-INT4-v2

This repository contains the v2 (Gold Series) optimized OpenVINO™ IR version of Qwen2-VL-2B-Instruct, quantized to INT4 precision using NNCF.

---

💎 The Difference: v1 vs. v2 (Gold Series)

This v2 release represents a significant architectural upgrade over the original v1 port.

Feature	v1 (Standard)	v2 (Gold Series)
C# Integration	Basic / Manual Logic	Native OpenVINO.GenAI (VLMPipeline)
Quantization	Initial INT4	Latest NNCF (85% INT4 / 15% INT8)
Exporter	Legacy Optimum-Intel	Optimum-Intel v1.20.0 (Latest Trace)
Metadata	Standard Tags	Gold-Series Branded / Verified
OCR Depth	Standard	Enhanced Dynamic Resolution Support

---

🐍 Python Inference (Optimum-Intel)

To run this vision engine locally using the optimum-intel library:

from optimum.intel import OVModelForVisualCausalLM
from transformers import AutoProcessor
from PIL import Image

model_id = "CelesteImperia/Qwen2-VL-2B-Instruct-OpenVINO-INT4-v2"
processor = AutoProcessor.from_pretrained(model_id)
model = OVModelForVisualCausalLM.from_pretrained(model_id)

image = Image.open("path/to/your/image.jpg")
prompt = "Describe this image in detail."

inputs = processor(text=[prompt], images=[image], return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=128)
print(processor.decode(outputs[0], skip_special_tokens=True))

---

💻 For C# / .NET Users (OpenVINO.GenAI)

The v2 release is optimized for the native OpenVINO.GenAI C# bindings, making production deployment in Windows automation systems seamless.

using OpenVino.GenAI;

// 1. Initialize the Visual-LLM Pipeline
var device = "CPU"; // Use "GPU" for RTX 3090/A4000 acceleration
using var pipe = new VLMPipeline("path/to/qwen2-vl-v2-model", device);

// 2. Prepare the visual input
var image = OpenVino.GenAI.Utils.LoadImage("automation_capture.png");
var prompt = "Perform OCR on this technical drawing and return the components as JSON.";

// 3. Execute Multimodal Inference
var result = pipe.Generate(prompt, image);
Console.WriteLine(result.Texts[0]);

---

🏗️ Technical Details

• Optimization Tool: NNCF (Neural Network Compression Framework)
• Quantization: INT4 Asymmetric (Group Size: 128)
• Multimodal Stack: Language Model, Visual Encoder, Merger Pipeline.
• Workstation Validation: Dual-GPU (RTX 3090 + RTX A4000)

---

☕ Support the Forge

Maintaining the infrastructure for high-bandwidth model hosting and multimodal AI research requires significant resources. If this v2 Gold Series model powers your industrial automation, consider supporting our development:

Platform	Support Link
Global & India	Support via Razorpay

Scan to support via UPI (India Only):

---

📜 License

This model is released under the Apache 2.0 License.

---

Connect with the architect: Abhishek Jaiswal on LinkedIn