latest

app.py

@@ -11,20 +11,33 @@ from transformers import Sam3Model, Sam3Processor
 import warnings
 warnings.filterwarnings("ignore")
 
-# Initialize model and processor (matching working space exactly)
-device = "cuda" if torch.cuda.is_available() else "cpu"
-model = Sam3Model.from_pretrained(
-    "facebook/sam3",
-    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
-).to(device)
-processor = Sam3Processor.from_pretrained("facebook/sam3")
+# Global variables for lazy initialization
+_model = None
+_processor = None
+_device = None
+
+def get_model_and_processor():
+    """Lazy initialization of model and processor"""
+    global _model, _processor, _device
+    if _model is None:
+        _device = "cuda" if torch.cuda.is_available() else "cpu"
+        _model = Sam3Model.from_pretrained(
+            "facebook/sam3",
+            torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
+        ).to(_device)
+        _processor = Sam3Processor.from_pretrained("facebook/sam3")
+        print(f"Model loaded on device: {_device}")
+    return _model, _processor, _device
 
 @spaces.GPU
 def sam3_inference(image, text_prompt, confidence_threshold=0.5):
     """
-    Standalone GPU function matching working space pattern
+    Standalone GPU function with lazy model initialization for Spaces Stateless GPU
     """
     try:
+        # Initialize model inside GPU function (required for Stateless GPU)
+        model, processor, device = get_model_and_processor()
+
         # Handle base64 input (for API)
         if isinstance(image, str):
             if image.startswith('data:image'):
@@ -32,14 +45,14 @@ def sam3_inference(image, text_prompt, confidence_threshold=0.5):
             image_bytes = base64.b64decode(image)
             image = Image.open(io.BytesIO(image_bytes)).convert("RGB")
 
-        # Process with SAM3 (matching working space exactly)
+        # Process with SAM3
         inputs = processor(
             images=image,
             text=text_prompt.strip(),
             return_tensors="pt"
         ).to(device)
 
-        # Convert dtype to match model (following working space pattern)
+        # Convert dtype to match model
         for key in inputs:
             if inputs[key].dtype == torch.float32:
                 inputs[key] = inputs[key].to(model.dtype)
@@ -64,7 +77,7 @@ class SAM3Handler:
     """SAM3 handler for both UI and API access"""
 
     def __init__(self):
-        print(f"SAM3 handler initialized with device: {device}")
+        print("SAM3 handler initialized (models will be loaded lazily)")
 
     def predict(self, image, text_prompt, confidence_threshold=0.5):
         """