gpu support

app.py

@@ -3,6 +3,9 @@ VideoMaMa Gradio Demo
 Interactive video matting with SAM2 mask tracking
 """
 
+# CRITICAL: Import spaces FIRST before any CUDA-related imports
+import spaces
+
 import os
 import json
 import time
@@ -47,9 +50,12 @@ POINT_ALPHA = 0.9
 POINT_RADIUS = 15
 
 def initialize_models():
-    """Initialize SAM2 and VideoMaMa models"""
+    """Initialize SAM2 and VideoMaMa models (lazy loading)"""
     global sam2_tracker, videomama_pipeline
     
+    if sam2_tracker is not None and videomama_pipeline is not None:
+        return  # Already initialized
+    
     device = "cuda" if torch.cuda.is_available() else "cpu"
     print(f"Using device: {device}")
     
@@ -149,6 +155,7 @@ def load_video(video_input, video_state):
            gr.update(visible=True), gr.update(visible=False)
 
 
+@spaces.GPU
 def sam_refine(video_state, point_prompt, click_state, evt: gr.SelectData):
     """
     Add click and update mask on first frame
@@ -159,6 +166,9 @@ def sam_refine(video_state, point_prompt, click_state, evt: gr.SelectData):
         click_state: [[points], [labels]]
         evt: Gradio SelectData event with click coordinates
     """
+    # Lazy load models on first use
+    initialize_models()
+    
     if video_state is None or "frames" not in video_state:
         return None, video_state, click_state
     
@@ -264,10 +274,14 @@ def propagate_masks(video_state, click_state):
     return video_state, status_msg, gr.update(visible=True)
 
 
+@spaces.GPU(duration=120)
 def run_videomama_with_sam2(video_state, click_state):
     """
     Run SAM2 propagation and VideoMaMa inference together
     """
+    # Lazy load models on first use
+    initialize_models()
+    
     if video_state is None or "frames" not in video_state:
         return video_state, None, None, None, "⚠️ No video loaded"
     
@@ -482,8 +496,8 @@ if __name__ == "__main__":
     print("VideoMaMa Interactive Demo")
     print("=" * 60)
     
-    # Initialize models
-    initialize_models()
+    # Models will be initialized on first use (lazy loading for ZeroGPU)
+    # initialize_models()
     
     # Launch demo
     demo.queue()

requirements.txt

@@ -1,5 +1,8 @@
 # Hugging Face Space Requirements for VideoMaMa Demo
 
+# CRITICAL: Hugging Face ZeroGPU support
+spaces
+
 # Core frameworks
 torch>=2.0.0
 torchvision>=0.15.0