Upload InternVL2 implementation

Dockerfile

@@ -6,6 +6,8 @@ ENV PYTHONUNBUFFERED=1
 ENV HF_HOME=/app/.cache/huggingface
 ENV TRANSFORMERS_CACHE=/app/.cache/huggingface/transformers
 ENV MPLCONFIGDIR=/tmp/matplotlib
+# Force PyTorch to use the NCCl backend
+ENV PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:128
 
 # Create necessary directories with proper permissions
 RUN mkdir -p /app/.cache/huggingface/transformers && \
@@ -23,11 +25,24 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     python3-pip \
     python3-dev \
     python3-setuptools \
+    nvidia-cuda-toolkit \
     && rm -rf /var/lib/apt/lists/*
 
 # Create a working directory
 WORKDIR /app
 
+# Add a script to check GPU status at startup
+RUN echo '#!/bin/bash \n\
+echo "Checking NVIDIA GPU status..." \n\
+if ! command -v nvidia-smi &> /dev/null; then \n\
+    echo "WARNING: nvidia-smi command not found. NVIDIA driver might not be installed." \n\
+else \n\
+    echo "NVIDIA driver found. Running nvidia-smi:" \n\
+    nvidia-smi \n\
+fi \n\
+exec "$@"' > /entrypoint.sh && \
+chmod +x /entrypoint.sh
+
 # Copy requirements file
 COPY requirements.txt .
 
@@ -63,5 +78,8 @@ RUN mkdir -p gradio_cached_examples && \
 # Make port 7860 available for the app
 EXPOSE 7860
 
+# Use our entrypoint script to check GPU status before starting the app
+ENTRYPOINT ["/entrypoint.sh"]
+
 # Start the application
 CMD ["python3", "app_internvl2.py"] 

app_internvl2.py

@@ -41,10 +41,31 @@ 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 = torch.cuda.is_available()
+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:
@@ -71,6 +92,12 @@ def load_internvl2_model():
         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:
@@ -91,6 +118,8 @@ def load_internvl2_model():
         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
 
@@ -104,6 +133,12 @@ def analyze_image(image, prompt):
             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."
@@ -164,9 +199,13 @@ def create_interface():
         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")
@@ -176,9 +215,15 @@ def create_interface():
                     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,
@@ -195,6 +240,17 @@ def create_interface():
         - **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(