handler.py

from typing import Dict, List, Any
from transformers import AutoModelForVision2Seq, AutoTokenizer, AutoProcessor
from PIL import Image
import torch
import io
import base64
from peft import PeftModel

class EndpointHandler():
    def __init__(self, model_dir: str):
        self.path = model_dir
        # Load base model and tokenizer
        base_model_id = "Qwen/Qwen2-VL-2B-Instruct"
        
        # Load tokenizer/processor
        self.processor = AutoProcessor.from_pretrained(
            self.path,
            trust_remote_code=True
        )
        
        # Load base model
        self.model = AutoModelForVision2Seq.from_pretrained(
            base_model_id,
            torch_dtype=torch.float16,
            device_map="auto",
            trust_remote_code=True
        )
        
        # Load LoRA adapter
        self.model = PeftModel.from_pretrained(
            self.model,
            self.path,
            device_map="auto"
        )
        
        # Merge and unload for faster inference
        self.model = self.model.merge_and_unload()
        
        # Set to eval mode
        self.model.eval()
        
        # Store the instruction template
        self.instruction = """
A conversation between a Healthcare Provider and an AI Medical Image Analysis Assistant. The provider shares a medical image, and the Assistant generates a clear description/report. The assistant first analyzes the image systematically, then provides a concise report. The analysis process and report are enclosed within <thinking> </thinking><answer> </answer>.
Always respond in this format:
<thinking>
1. Initial Assessment:
   - What type of image is this? (X-ray, CT, MRI, etc.)
   - Which body part/region is shown?
   - Is the image quality adequate?
2. Key Findings:
   - What are the normal structures visible?
   - Are there any abnormalities?
   - What are the important measurements?
3. Clinical Significance:
   - What are the main clinical findings?
   - Are there any critical findings?
</thinking>
<answer>
Brief Structured Report:
1. EXAM TYPE: [imaging type and body region]
2. FINDINGS: [key observations and abnormalities]
3. IMPRESSION: [summary and clinical significance]
</answer>
"""

    def __call__(self, data: Dict[str, Any]) -> List[Dict[str, str]]:
        """
        data args:
            inputs (:obj: `str` | `PIL.Image` | `np.array`)
            parameters (:obj: `Dict[str, Any]`, *optional*)
        Return:
            A :obj:`list` | `dict`: will be serialized and returned
        """
        # Extract inputs and parameters
        inputs = data.pop("inputs", data)
        parameters = data.pop("parameters", {})
        
        # Handle different input formats
        if isinstance(inputs, str):
            # Base64 encoded image
            image_bytes = base64.b64decode(inputs)
            image = Image.open(io.BytesIO(image_bytes)).convert("RGB")
        elif isinstance(inputs, dict):
            # Dictionary with image key
            image_data = inputs.get("image", inputs.get("inputs", ""))
            if isinstance(image_data, str):
                image_bytes = base64.b64decode(image_data)
                image = Image.open(io.BytesIO(image_bytes)).convert("RGB")
            else:
                image = image_data
        else:
            # Direct image
            image = inputs
            
        # Ensure image is RGB
        if image.mode != "RGB":
            image = image.convert("RGB")
        
        # Prepare messages in Qwen format
        messages = [
            {
                "role": "user",
                "content": [
                    {"type": "image", "image": image},
                    {"type": "text", "text": self.instruction}
                ]
            }
        ]
        
        # Process inputs
        text = self.processor.apply_chat_template(
            messages,
            tokenize=False,
            add_generation_prompt=True
        )
        
        # Prepare inputs for model
        inputs = self.processor(
            text=[text],
            images=[image],
            padding=True,
            return_tensors="pt"
        ).to(self.model.device)
        
        # Generate response
        with torch.no_grad():
            output_ids = self.model.generate(
                **inputs,
                max_new_tokens=parameters.get("max_new_tokens", 512),
                temperature=parameters.get("temperature", 0.7),
                top_p=parameters.get("top_p", 0.9),
                do_sample=True,
                pad_token_id=self.processor.tokenizer.pad_token_id,
                eos_token_id=self.processor.tokenizer.eos_token_id,
            )
        
        # Decode output - only the generated part
        output_text = self.processor.batch_decode(
            output_ids[:, inputs.input_ids.shape[1]:],
            skip_special_tokens=True
        )[0]
        
        return [{"generated_text": output_text}]