app_internvl2.py

import gradio as gr
from PIL import Image
import os
import time
import numpy as np
import torch
import warnings
import stat

# Set environment variables
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:128"

# Ensure all cache directories exist with proper permissions
def setup_cache_directories():
    # Gradio cache directory
    cache_dir = os.path.join(os.getcwd(), "gradio_cached_examples")
    os.makedirs(cache_dir, exist_ok=True)
    
    # HuggingFace cache directories
    hf_cache = os.path.join(os.getcwd(), ".cache", "huggingface")
    transformers_cache = os.path.join(hf_cache, "transformers")
    os.makedirs(hf_cache, exist_ok=True)
    os.makedirs(transformers_cache, exist_ok=True)
    
    # Set permissions
    try:
        for directory in [cache_dir, hf_cache, transformers_cache]:
            if os.path.exists(directory):
                os.chmod(directory, stat.S_IRWXU | stat.S_IRWXG | stat.S_IRWXO)  # 0o777
                print(f"Set permissions for {directory}")
    except Exception as e:
        print(f"Warning: Could not set permissions: {str(e)}")
    
    return cache_dir

# Set up cache directories
cache_dir = setup_cache_directories()

# Suppress specific warnings that might be caused by package version mismatches
warnings.filterwarnings("ignore", message=".*The 'nopython' keyword.*")
warnings.filterwarnings("ignore", message=".*Torch is not compiled with CUDA enabled.*")
warnings.filterwarnings("ignore", category=UserWarning)

# Check for actual GPU availability
def check_gpu_availability():
    """Check if GPU is actually available and working"""
    if not torch.cuda.is_available():
        print("CUDA is not available in PyTorch")
        return False
    
    try:
        # Try to initialize CUDA and run a simple operation
        x = torch.rand(10, device="cuda")
        y = x + x
        return True
    except Exception as e:
        print(f"GPU initialization failed: {str(e)}")
        return False

# Global variables
internvl2_pipeline = None
MODEL_LOADED = False
USE_GPU = check_gpu_availability()

if USE_GPU:
    print("GPU is available and working properly")
else:
    print("WARNING: GPU is not available or not working properly. This application requires GPU acceleration.")

# Check if lmdeploy is available and try to import
try:
    from lmdeploy import pipeline, TurbomindEngineConfig
    LMDEPLOY_AVAILABLE = True
    print("Successfully imported lmdeploy")
except ImportError:
    LMDEPLOY_AVAILABLE = False
    print("lmdeploy import failed. Will use a placeholder for demos.")

# Model configuration
MODEL_ID = "OpenGVLab/InternVL2-40B-AWQ"  # 4-bit quantized model

def load_internvl2_model():
    """Load the InternVL2 model using lmdeploy"""
    global internvl2_pipeline, MODEL_LOADED
    
    # If already loaded, return
    if internvl2_pipeline is not None:
        return True
    
    # If lmdeploy is not available, we'll use a demo placeholder
    if not LMDEPLOY_AVAILABLE:
        print("lmdeploy not available. Using demo placeholder.")
        MODEL_LOADED = False
        return False

    # Check if GPU is available
    if not USE_GPU:
        print("Cannot load InternVL2 model without GPU acceleration.")
        MODEL_LOADED = False
        return False
        
    print("Loading InternVL2 model...")
    try:
        # Configure for AWQ quantized model
        backend_config = TurbomindEngineConfig(model_format='awq')
        
        # Create pipeline
        internvl2_pipeline = pipeline(
            MODEL_ID, 
            backend_config=backend_config, 
            log_level='INFO'
        )
        
        print("InternVL2 model loaded successfully!")
        MODEL_LOADED = True
        return True
    except Exception as e:
        print(f"Error loading InternVL2 model: {str(e)}")
        if "CUDA out of memory" in str(e):
            print("Not enough GPU memory for the model")
        elif "Found no NVIDIA driver" in str(e):
            print("NVIDIA GPU driver not found or not properly configured")
        MODEL_LOADED = False
        return False

def analyze_image(image, prompt):
    """Analyze the image using InternVL2 model"""
    try:
        start_time = time.time()
        
        # Skip model loading if lmdeploy is not available
        if not LMDEPLOY_AVAILABLE:
            return ("This is a demo placeholder. The actual model couldn't be loaded because lmdeploy " 
                   "is not properly installed. Check your installation and dependencies.")
        
        # Check for GPU
        if not USE_GPU:
            return ("ERROR: This application requires a GPU to run InternVL2. "
                  "The NVIDIA driver was not detected on this system. "
                  "Please make sure this Space is using a GPU-enabled instance.")
        
        # Make sure the model is loaded
        if not load_internvl2_model():
            return "Couldn't load InternVL2 model. See logs for details."
            
        # Convert numpy array to PIL Image
        if isinstance(image, np.ndarray):
            image_pil = Image.fromarray(image).convert('RGB')
        else:
            # If somehow it's already a PIL Image
            image_pil = image.convert('RGB')
        
        # Run inference with the model
        response = internvl2_pipeline((prompt, image_pil))
        
        # Get the response text
        result = response.text
        
        elapsed_time = time.time() - start_time
        return result
        
    except Exception as e:
        print(f"Error in image analysis: {str(e)}")
        # Try to clean up memory in case of error
        if USE_GPU:
            torch.cuda.empty_cache()
        return f"Error in image analysis: {str(e)}"

def process_image(image, analysis_type="general"):
    """Process the image and return the analysis"""
    if image is None:
        return "Please upload an image."
    
    # Define prompt based on analysis type
    if analysis_type == "general":
        prompt = "Describe this image in detail."
    elif analysis_type == "text":
        prompt = "What text can you see in this image? Please transcribe it accurately."
    elif analysis_type == "chart":
        prompt = "Analyze any charts, graphs or diagrams in this image in detail, including trends, data points, and conclusions."
    elif analysis_type == "people":
        prompt = "Describe the people in this image - their appearance, actions, and expressions."
    elif analysis_type == "technical":
        prompt = "Provide a technical analysis of this image, including object identification, spatial relationships, and any technical elements present."
    else:
        prompt = "Describe this image in detail."
    
    start_time = time.time()
    
    # Get analysis from the model
    analysis = analyze_image(image, prompt)
    
    elapsed_time = time.time() - start_time
    return f"{analysis}\n\nAnalysis completed in {elapsed_time:.2f} seconds."

# Define the Gradio interface
def create_interface():
    with gr.Blocks(title="Image Analysis with InternVL2") as demo:
        gr.Markdown("# Image Analysis with InternVL2-40B")
        gr.Markdown("Upload an image to analyze it using the InternVL2-40B model.")
        
        # Show warnings based on system status
        if not LMDEPLOY_AVAILABLE:
            gr.Markdown("⚠️ **WARNING**: lmdeploy is not properly installed. This demo will not function correctly.", elem_classes=["warning-message"])
        
        if not USE_GPU:
            gr.Markdown("🚫 **ERROR**: NVIDIA GPU not detected. This application requires GPU acceleration to run InternVL2 model.", elem_classes=["error-message"])
        
        with gr.Row():
            with gr.Column(scale=1):
                input_image = gr.Image(type="pil", label="Upload Image")
                analysis_type = gr.Radio(
                    ["general", "text", "chart", "people", "technical"],
                    label="Analysis Type",
                    value="general"
                )
                submit_btn = gr.Button("Analyze Image")
                
                # Disable button if GPU is not available
                if not USE_GPU:
                    submit_btn.interactive = False
            
            with gr.Column(scale=2):
                output_text = gr.Textbox(label="Analysis Result", lines=20)
                if not USE_GPU:
                    output_text.value = "ERROR: NVIDIA GPU driver not detected. This application requires GPU acceleration to run the InternVL2 model. Please ensure this Space is using a GPU-enabled instance."
        
        submit_btn.click(
            fn=process_image,
            inputs=[input_image, analysis_type],
            outputs=output_text
        )
        
        gr.Markdown("""
        ## Analysis Types
        - **General**: General description of the image
        - **Text**: Focus on identifying and transcribing text in the image
        - **Chart**: Detailed analysis of charts, graphs, and diagrams
        - **People**: Description of people, their appearance and actions
        - **Technical**: Technical analysis identifying objects and spatial relationships
        """)
        
        # Hardware requirements notice
        gr.Markdown("""
        ## System Requirements
        This application requires:
        - NVIDIA GPU with CUDA support
        - At least 16GB of GPU memory recommended
        - GPU drivers properly installed and configured
        
        If you're running this on Hugging Face Spaces, make sure to select a GPU-enabled hardware type.
        """)
        
        # Examples
        try:
            gr.Examples(
                examples=[
                    ["data_temp/page_2.png", "general"],
                    ["data_temp/page_2.png", "text"],
                    ["data_temp/page_2.png", "chart"]
                ],
                inputs=[input_image, analysis_type],
                outputs=output_text,
                fn=process_image,
                cache_examples=True
            )
        except Exception as e:
            print(f"Warning: Could not load examples: {str(e)}")
    
    return demo

# Main function
if __name__ == "__main__":
    # Create the Gradio interface
    demo = create_interface()
    
    # Launch the interface (removed incompatible parameters)
    demo.launch(share=False, server_name="0.0.0.0")