from typing import Dict, List, Any from transformers import AutoProcessor, AutoModel import torch import torch.nn.functional as F from PIL import Image import requests import base64 from io import BytesIO class EndpointHandler: def __init__(self, path=""): self.processor = AutoProcessor.from_pretrained(path, trust_remote_code=True, use_fast=True) self.model = AutoModel.from_pretrained( path, trust_remote_code=True, torch_dtype=torch.float16, device_map="auto" ) self.model.eval() self.query_text = "This image shows a component that is used in HVAC system" def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]: inputs_data = data.get("inputs", []) if not isinstance(inputs_data, list): inputs_data = [inputs_data] results = [] for item in inputs_data: if isinstance(item, dict): image = item.get("image") text = item.get("text", self.query_text) else: image = None text = item if isinstance(item, str) else self.query_text if image: if isinstance(image, str) and image.startswith("data:image"): img_data = image.split(",")[1] pil_img = Image.open(BytesIO(base64.b64decode(img_data))) elif isinstance(image, str): pil_img = Image.open(requests.get(image, stream=True).raw) else: pil_img = image messages = [{"role": "user", "content": [ {"type": "image", "image": pil_img}, {"type": "text", "text": text}, ]}] prompt = self.processor.apply_chat_template( messages, tokenize=False, add_generation_prompt=False ) inputs = self.processor( text=[prompt], images=[pil_img], return_tensors="pt", padding=True ).to(self.model.device) else: messages = [{"role": "user", "content": [{"type": "text", "text": text}]}] prompt = self.processor.apply_chat_template( messages, tokenize=False, add_generation_prompt=False ) inputs = self.processor( text=[prompt], return_tensors="pt", padding=True ).to(self.model.device) with torch.no_grad(): out = self.model(**inputs, output_hidden_states=True) hidden = out.last_hidden_state if hasattr(out, 'last_hidden_state') else out[0] emb = hidden.mean(dim=1) emb = F.normalize(emb, p=2, dim=-1) results.append({"embedding": emb[0].cpu().float().tolist()}) return results