info endpoint

app.py

@@ -66,8 +66,50 @@ except ImportError:
 # Configuration
 # ============================================================================
 
-DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
-print(f"Device: {DEVICE}")
+# For ZeroGPU: Don't initialize CUDA at module level
+# Device will be determined when needed within GPU tasks
+_SPACES_ZERO_GPU = SPACES_AVAILABLE  # From spaces import above
+
+# Create a device proxy that works like torch.device but lazy-loads on ZeroGPU
+class _DeviceProxy:
+    """Proxy for torch.device that lazy-loads CUDA on ZeroGPU"""
+
+    def __init__(self):
+        self._device = None
+
+    @property
+    def type(self):
+        self._ensure_device()
+        return self._device.type
+
+    def __str__(self):
+        self._ensure_device()
+        return str(self._device)
+
+    def __repr__(self):
+        self._ensure_device()
+        return repr(self._device)
+
+    def _ensure_device(self):
+        """Lazy device initialization - only calls torch.cuda.is_available() when needed"""
+        if self._device is None:
+            if torch.cuda.is_available():
+                self._device = torch.device('cuda')
+            else:
+                self._device = torch.device('cpu')
+
+    def __eq__(self, other):
+        return str(self) == str(other)
+
+
+DEVICE = _DeviceProxy()
+
+if _SPACES_ZERO_GPU:
+    print(f"Device: Will use CUDA within GPU tasks (ZeroGPU mode)")
+else:
+    # Only access device if not ZeroGPU to avoid CUDA init
+    DEVICE._ensure_device()
+    print(f"Device: {DEVICE}")
 if SPACES_AVAILABLE:
     print("ZeroGPU support enabled")
 
@@ -242,7 +284,7 @@ class VGGFeatureExtractor(nn.Module):
 _vgg_extractor = None
 
 def get_vgg_extractor():
-    """Lazy load VGG feature extractor"""
+    """Lazy load VGG feature extractor (with ZeroGPU support)"""
     global _vgg_extractor
     if _vgg_extractor is None:
         _vgg_extractor = VGGFeatureExtractor().to(DEVICE)
@@ -528,16 +570,26 @@ def load_model(style: str, backend: str = 'auto') -> TransformerNet:
 print("=" * 50)
 print("StyleForge - Initializing...")
 print("=" * 50)
-print(f"Device: {DEVICE.type.upper()}")
+if _SPACES_ZERO_GPU:
+    print("Device: CUDA (ZeroGPU mode - lazy initialization)")
+else:
+    DEVICE._ensure_device()
+    print(f"Device: {DEVICE.type.upper()}")
 print(f"CUDA Kernels: {'Available' if CUDA_KERNELS_AVAILABLE else 'Not Available (will compile on first GPU task)'}")
-print("Preloading models...")
-for style in STYLES.keys():
-    try:
-        load_model(style, 'auto')
-        print(f"  {STYLES[style]}: Ready")
-    except Exception as e:
-        print(f"  {STYLES[style]}: Failed - {e}")
-print("All models loaded!")
+
+# Skip model preloading on ZeroGPU to avoid CUDA init in main process
+if not _SPACES_ZERO_GPU:
+    print("Preloading models...")
+    for style in STYLES.keys():
+        try:
+            load_model(style, 'auto')
+            print(f"  {STYLES[style]}: Ready")
+        except Exception as e:
+            print(f"  {STYLES[style]}: Failed - {e}")
+    print("All models loaded!")
+else:
+    print("ZeroGPU mode: Models will be loaded on-demand within GPU tasks")
+
 print("=" * 50)
 
 # ============================================================================