app.py

import os
import torch
import gradio as gr
import spaces
from PIL import Image
from transformers import AutoModel, AutoTokenizer, AutoModelForCausalLM
import warnings
warnings.filterwarnings("ignore")

# =========================================================
# إعدادات النموذج
# =========================================================

MODEL_ID = "openbmb/MiniCPM-o-2_6"

# تحميل كسول للنموذج
model = None
tokenizer = None


def load_model():
    """تحميل النموذج عند الحاجة فقط"""
    global model, tokenizer
    
    if model is not None:
        return
    
    print(f"Loading {MODEL_ID}...")
    
    # استخدام float16 للتوافق مع ZeroGPU
    device = "cuda" if torch.cuda.is_available() else "cpu"
    dtype = torch.float16 if torch.cuda.is_available() else torch.float32
    
    try:
        # تحميل tokenizer أولاً
        tokenizer = AutoTokenizer.from_pretrained(
            MODEL_ID,
            trust_remote_code=True,
            use_fast=False
        )
        
        # تحميل النموذج مع trust_remote_code=True
        model = AutoModel.from_pretrained(
            MODEL_ID,
            trust_remote_code=True,
            torch_dtype=dtype,
            low_cpu_mem_usage=True,
            attn_implementation="eager",
        ).eval()
        
        if torch.cuda.is_available():
            model = model.cuda()
        
        print("Model loaded successfully!")
        
    except Exception as e:
        print(f"Error with AutoModel, trying AutoModelForCausalLM: {e}")
        
        # محاولة بديلة مع AutoModelForCausalLM
        try:
            model = AutoModelForCausalLM.from_pretrained(
                MODEL_ID,
                trust_remote_code=True,  # مهم جداً!
                torch_dtype=dtype,
                low_cpu_mem_usage=True,
                attn_implementation="eager"
            ).eval()
            
            if torch.cuda.is_available():
                model = model.cuda()
            
            print("Model loaded successfully with AutoModelForCausalLM!")
                
        except Exception as e2:
            print(f"Failed to load model: {e2}")
            raise RuntimeError(f"Could not load model: {e2}")


# =========================================================
# دالة معالجة الصور
# =========================================================

def process_image(image_input):
    """معالجة الصورة للنموذج"""
    if image_input is None:
        return None
    
    if isinstance(image_input, str):
        return Image.open(image_input).convert('RGB')
    else:
        return image_input.convert('RGB')


# =========================================================
# دالة الاستدلال مع ZeroGPU
# =========================================================

@spaces.GPU(duration=60)
def generate_response(
    text_input,
    image_input,
    temperature,
    top_p,
    max_new_tokens
):
    """
    معالجة النص والصور باستخدام MiniCPM-o-2_6
    """
    
    if not text_input and not image_input:
        return "Please provide text or image input."
    
    try:
        load_model()
        global model, tokenizer
        
        # إعداد المدخلات
        if image_input is not None:
            # معالجة الصورة + النص
            image = process_image(image_input)
            
            if not text_input:
                text_input = "What is shown in this image? Please describe in detail."
            
            # التحقق من وجود دالة chat في النموذج
            if hasattr(model, 'chat'):
                try:
                    # استخدام دالة chat المخصصة
                    msgs = [{"role": "user", "content": [image, text_input]}]
                    
                    with torch.no_grad():
                        response = model.chat(
                            image=image,
                            msgs=msgs,
                            tokenizer=tokenizer,
                            sampling=True,
                            temperature=temperature,
                            top_p=top_p,
                            max_new_tokens=max_new_tokens
                        )
                    
                    return response
                    
                except Exception as e:
                    print(f"Chat method failed: {e}")
                    # السقوط إلى الطريقة العادية
            
            # الطريقة البديلة للصور
            # دمج النص مع وصف الصورة
            prompt = f"Image: [Image will be processed]\n\nQuestion: {text_input}\n\nAnswer:"
            
        else:
            # نص فقط
            prompt = text_input
        
        # المعالجة العادية للنص
        inputs = tokenizer(
            prompt,
            return_tensors="pt",
            padding=True,
            truncation=True,
            max_length=2048
        )
        
        if torch.cuda.is_available():
            inputs = {k: v.cuda() for k, v in inputs.items() if v is not None}
        
        # إعدادات التوليد
        gen_kwargs = {
            "max_new_tokens": max_new_tokens,
            "temperature": temperature if temperature > 0 else 1e-7,
            "top_p": top_p,
            "do_sample": temperature > 0,
            "pad_token_id": tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id,
            "eos_token_id": tokenizer.eos_token_id,
        }
        
        # التوليد
        with torch.no_grad():
            outputs = model.generate(**inputs, **gen_kwargs)
        
        # فك التشفير
        response = tokenizer.decode(
            outputs[0][inputs['input_ids'].shape[1]:],
            skip_special_tokens=True
        )
        
        return response.strip()
        
    except Exception as e:
        import traceback
        traceback.print_exc()
        return f"Error: {str(e)}"


# =========================================================
# دوال مساعدة للواجهة
# =========================================================

def clear_all():
    """مسح جميع المدخلات والمخرجات"""
    return "", None, ""


def update_examples_visibility(show_examples):
    """تحديث رؤية الأمثلة"""
    return gr.update(visible=show_examples)


# =========================================================
# واجهة Gradio
# =========================================================

def create_demo():
    """إنشاء واجهة Gradio البسيطة"""
    
    with gr.Blocks(title="MiniCPM-o-2.6", css="""
        .gradio-container {
            max-width: 1200px;
            margin: auto;
        }
        h1 {
            text-align: center;
        }
        .contain {
            background: white;
            border-radius: 10px;
            padding: 20px;
        }
    """) as demo:
        
        gr.Markdown(
            """
            # 🤖 MiniCPM-o-2.6 - Multimodal AI Assistant
            
            <div style="text-align: center;">
                <p>
                    <b>8B parameters model</b> with GPT-4 level performance<br>
                    Supports: Text Generation, Image Understanding, OCR, and Multi-lingual conversations
                </p>
            </div>
            """
        )
        
        with gr.Row():
            # العمود الرئيسي
            with gr.Column(scale=2):
                with gr.Group():
                    text_input = gr.Textbox(
                        label="💭 Text Input",
                        placeholder="Enter your question or prompt here...\nYou can ask about images, request text generation, or have a conversation.",
                        lines=4,
                        elem_id="text_input"
                    )
                    
                    image_input = gr.Image(
                        label="📷 Image Input (Optional)",
                        type="pil",
                        elem_id="image_input"
                    )
                
                with gr.Row():
                    submit_btn = gr.Button(
                        "🚀 Generate Response",
                        variant="primary",
                        scale=2
                    )
                    clear_btn = gr.Button(
                        "🗑️ Clear All",
                        variant="secondary",
                        scale=1
                    )
                
                output = gr.Textbox(
                    label="🤖 AI Response",
                    lines=10,
                    interactive=False,
                    elem_id="output"
                )
            
            # عمود الإعدادات
            with gr.Column(scale=1):
                with gr.Group():
                    gr.Markdown("### ⚙️ Generation Settings")
                    
                    temperature = gr.Slider(
                        label="Temperature",
                        minimum=0.0,
                        maximum=1.5,
                        value=0.7,
                        step=0.1,
                        info="Controls randomness (0=deterministic, 1.5=very creative)"
                    )
                    
                    top_p = gr.Slider(
                        label="Top-p (Nucleus Sampling)",
                        minimum=0.1,
                        maximum=1.0,
                        value=0.9,
                        step=0.05,
                        info="Controls diversity of output"
                    )
                    
                    max_new_tokens = gr.Slider(
                        label="Max New Tokens",
                        minimum=50,
                        maximum=2048,
                        value=512,
                        step=50,
                        info="Maximum length of generated response"
                    )
                
                gr.Markdown(
                    """
                    ### 📚 Quick Tips:
                    
                    **Text Generation:**
                    - Ask questions
                    - Request explanations
                    - Generate creative content
                    
                    **Image Understanding:**
                    - Upload an image
                    - Ask about contents
                    - Request OCR/text extraction
                    - Get detailed descriptions
                    
                    **Languages:**
                    - English, Chinese, Arabic
                    - And many more!
                    """
                )
        
        # أمثلة
        with gr.Group():
            gr.Markdown("### 💡 Example Prompts")
            gr.Examples(
                examples=[
                    ["Explain quantum computing in simple terms for a beginner.", None],
                    ["Write a short story about a robot learning to paint.", None],
                    ["What are the main differences between Python and JavaScript?", None],
                    ["Create a healthy meal plan for one week.", None],
                    ["Translate 'Hello, how are you?' to French, Spanish, and Arabic.", None],
                ],
                inputs=[text_input, image_input],
                outputs=output,
                fn=lambda t, i: generate_response(t, i, 0.7, 0.9, 512),
                cache_examples=False,
                label="Click any example to try it"
            )
        
        # ربط الأحداث
        submit_btn.click(
            fn=generate_response,
            inputs=[text_input, image_input, temperature, top_p, max_new_tokens],
            outputs=output,
            api_name="generate"
        )
        
        text_input.submit(
            fn=generate_response,
            inputs=[text_input, image_input, temperature, top_p, max_new_tokens],
            outputs=output
        )
        
        clear_btn.click(
            fn=clear_all,
            inputs=[],
            outputs=[text_input, image_input, output]
        )
        
        # رسالة ترحيبية عند التحميل
        demo.load(
            lambda: gr.Info("Model is loading... This may take a moment on first use."),
            inputs=None,
            outputs=None
        )
    
    return demo


# =========================================================
# تشغيل التطبيق
# =========================================================

if __name__ == "__main__":
    demo = create_demo()
    demo.launch(
        ssr_mode=False,
        show_error=True,
        share=False
    )