# Qwen2-VL-2B-Instruct – Stage 3 (Multi-turn)

## 1. Model Overview

This model is part of a Vision-Language AI system designed for chest X-ray analysis in Vietnamese clinical settings.

The full pipeline consists of 3 stages:
- Stage 1: Findings generation (image → radiology findings)
- Stage 2: Impression generation (image → clinical impression)
- Stage 3: Multi-turn conversation (findings + impression + dialogue)

This repository corresponds to:

- Stage: 3 (Multi-turn)
- Task: Multi-turn reasoning with findings and impression
- Domain: Vietnamese medical imaging (Chest X-ray)

The model supports **multi-turn dialogue**, where:
- Turn 1: Generate findings  
- Turn 2: Generate clinical impression based on previous context  

---

## 2. Installation

```bash
pip install torch torchvision transformers qwen-vl-utils pillow
```

---

## 3. Inference

GPU is recommended.

```python
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info
import torch

model = Qwen2VLForConditionalGeneration.from_pretrained(
    "THP2903/weight_qwen2-2b_instruct_multi",
    torch_dtype="auto",
    device_map="auto"
)

processor = AutoProcessor.from_pretrained("THP2903/weight_qwen2-2b_instruct_multi")

# Turn 1: Findings
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "your_image.jpg"},
            {"type": "text", "text": "Ảnh chụp xray bệnh nhân nam, 48 tuổi PA. Mô tả thông tin benh nhân."},
        ],
    }
]

text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
image_inputs, video_inputs = process_vision_info(messages)

inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
).to("cuda")

generated_ids = model.generate(**inputs, max_new_tokens=512)

generated_ids_trimmed = [
    out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]

response1 = processor.batch_decode(
    generated_ids_trimmed,
    skip_special_tokens=True,
    clean_up_tokenization_spaces=False
)[0]

print("Turn 1:", response1)

# Turn 2: Impression (reuse previous response)
messages.append(
    {"role": "assistant", "content": [{"type": "text", "text": response1}]}
)

messages.append(
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "Kết luận bệnh gì?"}
        ],
    }
)

text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
image_inputs, video_inputs = process_vision_info(messages)

inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
).to("cuda")

generated_ids = model.generate(**inputs, max_new_tokens=512)

generated_ids_trimmed = [
    out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]

response2 = processor.batch_decode(
    generated_ids_trimmed,
    skip_special_tokens=True,
    clean_up_tokenization_spaces=False
)[0]

print("Turn 2:", response2)
```

---

## 4. Notes

- Input must be a chest X-ray image  
- Turn 1 generates findings  
- Turn 2 generates clinical impression using previous conversation context  
- Conversation history is maintained via messages list  
- This model follows the original Qwen2-VL multi-turn inference pipeline  
- For best performance, consider using Qwen2-VL-7B