Add debug logging for Blackwell worker_init CUDA failure

app.py

@@ -1,5 +1,22 @@
 import spaces
 
+# Patch spaces torch init to log debug info before the failing cuda call
+try:
+    import spaces.zero.torch.patching as _sp
+    _orig_init = _sp.init
+    def _debug_init(nvidia_uuid):
+        import os, torch
+        print(f"[DEBUG worker_init] nvidia_uuid={nvidia_uuid!r}")
+        print(f"[DEBUG worker_init] CUDA_VISIBLE_DEVICES={os.environ.get('CUDA_VISIBLE_DEVICES')!r}")
+        try:
+            print(f"[DEBUG worker_init] torch.cuda.device_count() before init={torch.cuda.device_count()}")
+        except Exception as e:
+            print(f"[DEBUG worker_init] device_count() raised: {e}")
+        return _orig_init(nvidia_uuid)
+    _sp.init = _debug_init
+except Exception as e:
+    print(f"[DEBUG] Could not patch spaces.zero.torch.patching: {e}")
+
 import torch
 import torchvision.transforms.functional as TF
 import numpy as np
@@ -7,6 +24,13 @@ import random
 import os
 import sys
 
+# Log CUDA state at startup (no GPU expected here in ZeroGPU)
+print(f"[DEBUG startup] CUDA_VISIBLE_DEVICES={os.environ.get('CUDA_VISIBLE_DEVICES')!r}")
+try:
+    print(f"[DEBUG startup] torch.cuda.device_count()={torch.cuda.device_count()}")
+except Exception as e:
+    print(f"[DEBUG startup] device_count() raised: {e}")
+
 from diffusers.utils import load_image
 from diffusers import EulerDiscreteScheduler, T2IAdapter