models/qwen.py

import os
import json
from PIL import Image
import torch
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from transformers import Pix2StructForConditionalGeneration, Pix2StructProcessor
from qwen_vl_utils import process_vision_info

# Globals (lazy-loaded at runtime)
qwen_model = None
qwen_processor = None
ocr_model = None
ocr_processor = None


def load_prompt():
    #with open("prompts/prompt.txt", "r", encoding="utf-8") as f:
    #    return f.read()
    return os.getenv("PROMPT_TEXT", "⚠️ PROMPT_TEXT not found in secrets.")


def try_extract_json(text):
    try:
        return json.loads(text)
    except json.JSONDecodeError:
        start = text.find('{')
        if start == -1:
            return text
        brace_count = 0
        json_candidate = ''
        for i in range(start, len(text)):
            if text[i] == '{':
                brace_count += 1
            elif text[i] == '}':
                brace_count -= 1
            json_candidate += text[i]
            if brace_count == 0:
                break
        try:
            return json.loads(json_candidate)
        except json.JSONDecodeError:
            return text


def extract_all_text_pix2struct(image: Image.Image):
    global ocr_model, ocr_processor

    if ocr_model is None or ocr_processor is None:
        ocr_processor = Pix2StructProcessor.from_pretrained("google/pix2struct-textcaps-base")
        ocr_model = Pix2StructForConditionalGeneration.from_pretrained("google/pix2struct-textcaps-base")
        device = "cuda" if torch.cuda.is_available() else "cpu"
        ocr_model = ocr_model.to(device)

    inputs = ocr_processor(images=image, return_tensors="pt").to(ocr_model.device)
    predictions = ocr_model.generate(**inputs, max_new_tokens=512)
    return ocr_processor.decode(predictions[0], skip_special_tokens=True).strip()


def assign_event_gateway_names_from_ocr(json_data: dict, ocr_text: str):
    if not ocr_text or not json_data:
        return json_data

    def assign_best_guess(obj):
        if not obj.get("name") or obj["name"].strip() == "":
            obj["name"] = "(label unknown)"

    for evt in json_data.get("events", []):
        assign_best_guess(evt)

    for gw in json_data.get("gateways", []):
        assign_best_guess(gw)

    return json_data


def run_model(image: Image.Image):
    global qwen_model, qwen_processor

    if qwen_model is None or qwen_processor is None:
        qwen_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
            "Qwen/Qwen2.5-VL-7B-Instruct",
            torch_dtype=torch.bfloat16,
            device_map="auto"
            # You can enable flash attention here if needed:
            # attn_implementation="flash_attention_2"
        )

        min_pixels = 256 * 28 * 28
        max_pixels = 1080 * 28 * 28
        qwen_processor = AutoProcessor.from_pretrained(
            "Qwen/Qwen2.5-VL-7B-Instruct",
            min_pixels=min_pixels,
            max_pixels=max_pixels
        )

    prompt = load_prompt()

    messages = [
        {
            "role": "user",
            "content": [
                {"type": "image", "image": image},
                {"type": "text", "text": prompt}
            ]
        }
    ]

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

    inputs = qwen_processor(
        text=[text],
        images=image_inputs,
        videos=video_inputs,
        padding=True,
        return_tensors="pt"
    ).to(qwen_model.device)

    generated_ids = qwen_model.generate(**inputs, max_new_tokens=5000)
    generated_ids_trimmed = [out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]

    output_text = qwen_processor.batch_decode(
        generated_ids_trimmed,
        skip_special_tokens=True,
        clean_up_tokenization_spaces=False
    )[0]

    parsed_json = try_extract_json(output_text)

    # OCR post-processing
    ocr_text = extract_all_text_pix2struct(image)
    parsed_json = assign_event_gateway_names_from_ocr(parsed_json, ocr_text)

    return {
        "json": parsed_json,
        "raw": output_text
    }