api.py

import argparse
from flask import Flask, request, jsonify
from PIL import Image
from io import BytesIO
import base64
import torch
from transformers import AutoProcessor, AutoModelForCausalLM
import threading

app = Flask(__name__)

# Parse command line arguments
parser = argparse.ArgumentParser(description='Start the Flask server with specified model and device.')
parser.add_argument('--model-path', type=str, required=True, help='Path to the pretrained model')
parser.add_argument('--device', type=str, choices=['cpu', 'gpu'], default='auto', help='Device to use: "cpu", "gpu", or "auto"')
args = parser.parse_args()

# Determine the device
if args.device == 'auto':
    device = "cuda:0" if torch.cuda.is_available() else "cpu"
elif args.device == 'gpu':
    if torch.cuda.is_available():
        device = "cuda:0"
    else:
        raise ValueError("GPU option specified but no GPU is available.")
else:
    device = "cpu"

torch_dtype = torch.float16 if device.startswith("cuda") else torch.float32

from unittest.mock import patch
from transformers.dynamic_module_utils import get_imports
import os

def fixed_get_imports(filename: str | os.PathLike) -> list[str]:
    if not str(filename).endswith("modeling_florence2.py"):
        return get_imports(filename)
    imports = get_imports(filename)
    imports.remove("flash_attn")
    return imports

# Initialize the model and processor
with patch("transformers.dynamic_module_utils.get_imports", fixed_get_imports): #workaround for unnecessary flash_attn requirement
            model = AutoModelForCausalLM.from_pretrained(args.model_path, attn_implementation="sdpa", torch_dtype=torch_dtype,trust_remote_code=True).to(device)
processor = AutoProcessor.from_pretrained(args.model_path, trust_remote_code=True)

lock = threading.Lock()  # Use a lock to ensure thread safety when accessing the model

def predict_image(image, task: str = "<OD>", prompt: str = None):
    prompt = task + " " + prompt if prompt else task
    print(f"Prompt: {prompt}")
    with lock:
        inputs = processor(text=prompt, images=image, return_tensors="pt").to(device, torch_dtype)
        generated_ids = model.generate(
            input_ids=inputs["input_ids"],
            pixel_values=inputs["pixel_values"],
            max_new_tokens=1024,
            do_sample=False,
            num_beams=3
        )
        generated_text = processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
        parsed_answer = processor.post_process_generation(generated_text, task=task, image_size=(image.width, image.height))
    return parsed_answer

@app.route('/predict', methods=['POST'])
def predict():
    if request.is_json:
        data = request.get_json()
        if 'image' not in data:
            return jsonify({'error': 'No image found in JSON'}), 400
        image_data = base64.b64decode(data['image'].split(",")[1])
        image = Image.open(BytesIO(image_data))
    else:
        return jsonify({'error': 'No image file or JSON payload'}), 400
    task = data.get('task', "<OD>")
    prompt = data.get('prompt', None)
    prediction = predict_image(image, task, prompt)
    msgid = data.get('msgid', None)
    response = {
        'msgid': msgid,
        'prediction': prediction
    }
    return jsonify(response)

if __name__ == '__main__':
    app.run(host='0.0.0.0', port=5000, threaded=True)  # Enable multi-threading