Upload app.py with huggingface_hub

app.py

@@ -176,7 +176,7 @@ def diagnose():
     lines.append(f"config: {Path(CONFIG_PATH).exists()}")
     return "\n".join(lines)
 
-@spaces.GPU(duration=300)
+@spaces.GPU(duration=600)
 def reconstruct_objects(image: np.ndarray):
     if image is None:
         return None, None, "No image"
@@ -186,8 +186,9 @@ def reconstruct_objects(image: np.ndarray):
         print(f"GPU: {torch.cuda.get_device_name()}")
 
         from sam2.automatic_mask_generator import SAM2AutomaticMaskGenerator
-        sam2_gen = SAM2AutomaticMaskGenerator.from_pretrained("facebook/sam2-hiera-large")
-        print(f"  SAM2 loaded ({time.time()-t0:.0f}s)")
+        print(f"  Loading SAM2... (VRAM: {torch.cuda.memory_allocated()/1e9:.1f}GB)")
+        sam2_gen = SAM2AutomaticMaskGenerator.from_pretrained("facebook/sam2-hiera-small")
+        print(f"  SAM2 loaded ({time.time()-t0:.0f}s, VRAM: {torch.cuda.memory_allocated()/1e9:.1f}GB)")
 
         image_np = np.array(image) if not isinstance(image, np.ndarray) else image
         masks = sam2_gen.generate(image_np)
@@ -200,12 +201,19 @@ def reconstruct_objects(image: np.ndarray):
         preview[best_mask] = (preview[best_mask] * 0.5 + np.array([0, 255, 0]) * 0.5).astype(np.uint8)
         print(f"  {len(masks)} masks ({time.time()-t0:.0f}s)")
 
+        # Free SAM2 to save VRAM for SAM3D
+        del sam2_gen
+        torch.cuda.empty_cache()
+        print(f"  SAM2 freed (VRAM: {torch.cuda.memory_allocated()/1e9:.1f}GB)")
+
         from inference import Inference
+        print(f"  Loading SAM3D... (VRAM: {torch.cuda.memory_allocated()/1e9:.1f}GB)")
         sam3d = Inference(CONFIG_PATH, compile=False)
-        print(f"  SAM3D loaded ({time.time()-t0:.0f}s)")
+        print(f"  SAM3D loaded ({time.time()-t0:.0f}s, VRAM: {torch.cuda.memory_allocated()/1e9:.1f}GB)")
 
+        print(f"  Running reconstruction... (VRAM: {torch.cuda.memory_allocated()/1e9:.1f}GB)")
         result = sam3d(image=image_np, mask=best_mask, seed=42)
-        print(f"  Reconstructed ({time.time()-t0:.0f}s)")
+        print(f"  Reconstructed ({time.time()-t0:.0f}s, VRAM: {torch.cuda.memory_allocated()/1e9:.1f}GB)")
         if result is None:
             return None, preview, "Reconstruction returned None"