disable expandable segments: they abort the ZeroGPU worker

app.py

@@ -21,9 +21,13 @@ the GitHub repo locally for free text.
 
 import os
 
-# Set before torch is imported anywhere: the decoder allocates and frees a
-# pixel-space buffer every block, which is the pattern expandable segments exist for.
-os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True")
+# Deliberately NOT `expandable_segments:True`. That is the usual fix for allocator
+# trouble under transient spikes, but here it *causes* it: expandable segments grow
+# through the CUDA VMM path, and the first growth inside a ZeroGPU worker aborts with
+#   NVML_SUCCESS == r INTERNAL ASSERT FAILED ... CUDACachingAllocator.cpp
+# on an allocation of ~49 MB, while the same allocation succeeds with the default
+# allocator. Set it explicitly so a platform default cannot turn it back on.
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:False"
 
 import queue
 import shutil