app.py

import gradio as gr
from transformers import AutoModelForImageTextToText, AutoProcessor
from peft import PeftModel
import torch

# --- CONFIGURATION ---
BASE_MODEL = "unsloth/Qwen3-VL-2B-Instruct-unsloth-bnb-4bit"
LORA_ID = "EthanCastro/qwen3-vl-2b-quickdraw" 

print("Loading model and processor...")
model = AutoModelForImageTextToText.from_pretrained(
    BASE_MODEL,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True
)

model = PeftModel.from_pretrained(model, LORA_ID)
processor = AutoProcessor.from_pretrained("Qwen/Qwen3-VL-2B-Instruct", trust_remote_code=True)
print("Model Ready!")

def respond(message, image, history):
    # History is now a list of dictionaries
    # Format: [{"role": "user", "content": "hi"}, {"role": "assistant", "content": "hello"}]
    messages = []
    
    # 1. Convert history to Qwen's multimodal format
    for msg in history:
        # We need to ensure content is treated as text for the history buffer
        content = msg["content"]
        # If content is a list (multimodal), extract just the text for simplicity
        if isinstance(content, list):
            text_content = next((item['text'] for item in content if item['type'] == 'text'), "")
        else:
            text_content = content
            
        messages.append({
            "role": msg["role"], 
            "content": [{"type": "text", "text": text_content}]
        })

    # 2. Add current user turn with the new image
    user_content = []
    if image is not None:
        user_content.append({"type": "image", "image": image})
    user_content.append({"type": "text", "text": message})
    messages.append({"role": "user", "content": user_content})

    # 3. Tokenize and Generate
    text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    
    if image is not None:
        inputs = processor(text=[text], images=[image], return_tensors="pt").to("cuda")
    else:
        inputs = processor(text=[text], return_tensors="pt").to("cuda")

    with torch.no_grad():
        outputs = model.generate(**inputs, max_new_tokens=1500, temperature=0.3)
    
    generated_text = processor.batch_decode(outputs, skip_special_tokens=True)[0]
    
    if "assistant" in generated_text:
        response = generated_text.split("assistant")[-1].strip()
    else:
        response = generated_text

    return response

# --- GRADIO INTERFACE ---
# Note: 'theme' removed from here per Gradio 6 migration guide
with gr.Blocks() as demo:
    gr.Markdown("# 🎨 QuickDraw → tldraw JSON")
    
    # Chatbot using default "messages" format (no type argument needed)
    chatbot = gr.Chatbot(height=500)
    
    with gr.Row():
        img_input = gr.Image(type="pil", label="Upload Sketch", scale=1)
        with gr.Column(scale=3):
            txt_input = gr.Textbox(
                show_label=False, 
                placeholder="Convert this sketch to tldraw JSON format...", 
                container=False
            )
            submit_btn = gr.Button("Send", variant="primary")

    def chat_wrapper(message, image, history):
        # 1. Get response
        bot_res = respond(message, image, history)
        
        # 2. Update history using DICTIONARIES
        history.append({"role": "user", "content": message})
        history.append({"role": "assistant", "content": bot_res})
        
        return "", None, history

    # Initialize state as an empty list
    submit_btn.click(chat_wrapper, [txt_input, img_input, chatbot], [txt_input, img_input, chatbot])
    txt_input.submit(chat_wrapper, [txt_input, img_input, chatbot], [txt_input, img_input, chatbot])

# Theme is now applied here in launch()
# Disable SSR to help prevent 503 errors on resource-constrained Spaces
demo.launch(theme=gr.themes.Soft(), ssr_mode=False)