fix(zerogpu): bypass PyTorch CUDA caching allocator

app.py

@@ -17,6 +17,8 @@
 
 import os
 
+# ZeroGPU/cu128 can hit an NVML assert inside PyTorch CUDACachingAllocator.
+os.environ.setdefault("PYTORCH_NO_CUDA_MEMORY_CACHING", "1")
 os.environ["GRADIO_APP"] = "texture_edit"
 import gradio as gr
 from app_style import custom_theme, image_css, lighting_css