app.py

import gradio as gr
import torch
from transformers import AutoProcessor, AutoModelForCausalLM, tqdm

# 1. Configuration & Model IDs
FLORENCE_MODEL_ID = "microsoft/Florence-2-base"
DOLPHIN_MODEL_ID = "cognitivecomputations/dolphin-2.9.4-qwen2-1.5b" # Or your merged model path

# 2. Load Florence-2 (Vision)
print("Loading Florence-2...")
florence_model = AutoModelForCausalLM.from_pretrained(FLORENCE_MODEL_ID, trust_remote_code=True).to("cpu").eval()
florence_processor = AutoProcessor.from_pretrained(FLORENCE_MODEL_ID, trust_remote_code=True)

# 3. Load Dolphin-Qwen (Reasoning)
print("Loading Dolphin-Qwen...")
dolphin_model = AutoModelForCausalLM.from_pretrained(DOLPHIN_MODEL_ID).to("cpu").eval()
dolphin_tokenizer = AutoProcessor.from_pretrained(DOLPHIN_MODEL_ID)

def process_ui_task(image, prompt):
    # This function will eventually handle the phased logic:
    # Phase 1: Florence identifies the button
    # Phase 3: Dolphin maps it to a Hex Packet
    
    # Simple Florence Inference Example
    inputs = florence_processor(text="<OD>", images=image, return_tensors="pt").to("cpu")
    generated_ids = florence_model.generate(
        input_ids=inputs["input_ids"],
        pixel_values=inputs["pixel_values"],
        max_new_tokens=1024,
        do_sample=False,
        num_beams=3
    )
    results = florence_processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
    
    return f"Brain Output: {results}"

# 4. Gradio Interface
interface = gr.Interface(
    fn=process_ui_task,
    inputs=[gr.Image(type="pill"), gr.Textbox(label="Instruction")],
    outputs="text",
    title="AI Automation Brain (Florence + Dolphin)"
)

if __name__ == "__main__":
    interface.launch()