more CUDA initialization fixes via perplexity

app.py

@@ -1,18 +1,20 @@
+import spaces
 import gradio as gr
 from PIL import Image
-import src.depth_pro as depth_pro
 import numpy as np
 import matplotlib.pyplot as plt
 import subprocess
-import spaces
-import torch
 import tempfile
 import os
 import trimesh
 import time
+from datetime import datetime
+
+# Import potentially CUDA-initializing modules after 'spaces'
+import torch
+import src.depth_pro as depth_pro
 import timm
 import cv2
-from datetime import datetime
 
 print(f"Timm version: {timm.__version__}")
 
@@ -20,7 +22,7 @@ subprocess.run(["bash", "get_pretrained_models.sh"])
 
 @spaces.GPU(duration=20)
 def load_model_and_predict(image_path):
-    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
     model, transform = depth_pro.create_model_and_transforms()
     model = model.to(device)
     model.eval()
@@ -68,6 +70,7 @@ def resize_image(image_path, max_size=1024):
             img.save(temp_file, format="PNG")
             return temp_file.name
 
+@spaces.GPU(duration=20)
 def generate_3d_model(depth, image_path, focallength_px, simplification_factor=0.8, smoothing_iterations=1, thin_threshold=0.01):
     """
     Generate a textured 3D mesh from the depth map and the original image.
@@ -178,6 +181,7 @@ def remove_thin_features(mesh, thickness_threshold=0.01):
     
     return mesh
 
+@spaces.GPU(duration=20)
 def regenerate_3d_model(depth_csv, image_path, focallength_px, simplification_factor, smoothing_iterations, thin_threshold):
     # Load depth from CSV
     depth = np.loadtxt(depth_csv, delimiter=',')
@@ -190,6 +194,7 @@ def regenerate_3d_model(depth_csv, image_path, focallength_px, simplification_fa
     
     return view_model_path, download_model_path
 
+@spaces.GPU(duration=20)
 def predict_depth(input_image):
     temp_file = None
     try: