routers/donut_inference.py

import re
import time
import torch
from transformers import DonutProcessor, VisionEncoderDecoderModel
from config import settings
from functools import lru_cache
import os
import requests


@lru_cache(maxsize=1)
def load_model(model_url: str):
    model_name = model_url.replace("https://huggingface.co/", "")
    print(f"[Model Loader] Loading model: {model_name}")

    processor = DonutProcessor.from_pretrained(model_name)
    model = VisionEncoderDecoderModel.from_pretrained(model_name)

    device = "cuda" if torch.cuda.is_available() else "cpu"
    model.to(device)

    return processor, model, device


def process_document_donut(image, model_url: str):
    """
    Process the document using the DONUT model.

    :param image: The input image to process.
    :param model_url: The model URL to use for inference.
    :return: A tuple of the result and processing time.
    """
    worker_pid = os.getpid()
    print(f"[Inference] Handling request with worker PID: {worker_pid}")

    start_time = time.time()

    # Load the model dynamically based on the model_url
    processor, model, device = load_model(model_url)

    # Prepare encoder inputs
    pixel_values = processor(image, return_tensors="pt").pixel_values

    # Prepare decoder inputs
    task_prompt = "<s_cord-v2>"
    decoder_input_ids = processor.tokenizer(task_prompt, add_special_tokens=False, return_tensors="pt").input_ids

    # Generate answer
    outputs = model.generate(
        pixel_values.to(device),
        decoder_input_ids=decoder_input_ids.to(device),
        max_length=model.decoder.config.max_position_embeddings,
        early_stopping=True,
        pad_token_id=processor.tokenizer.pad_token_id,
        eos_token_id=processor.tokenizer.eos_token_id,
        use_cache=True,
        num_beams=1,
        bad_words_ids=[[processor.tokenizer.unk_token_id]],
        return_dict_in_generate=True,
    )

    # Postprocess the result
    sequence = processor.batch_decode(outputs.sequences)[0]
    sequence = sequence.replace(processor.tokenizer.eos_token, "").replace(processor.tokenizer.pad_token, "")
    sequence = re.sub(r"<.*?>", "", sequence, count=1).strip()  # Remove first task start token

    end_time = time.time()
    processing_time = end_time - start_time

    print(f"[Inference] Done. PID: {worker_pid} | Time taken: {processing_time:.2f} sec")

    return processor.token2json(sequence), processing_time