Use cudaMallocAsync allocator to fix MIG GPU CUDACachingAllocator crash

PyTorch's default CUDA Caching Allocator makes NVML memory queries
(nvmlDeviceGetMemoryInfo) that fail on MIG GPU instances running as
non-root, causing: NVML_SUCCESS == r INTERNAL ASSERT FAILED at
CUDACachingAllocator.cpp:1165. The cudaMallocAsync backend avoids
these NVML queries entirely.

Set via PYTORCH_CUDA_ALLOC_CONF at module level (before any CUDA init)
so it is inherited by ZeroGPU's subprocess.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app.py

@@ -4,6 +4,9 @@ import time
 import threading
 import traceback
 
+# cudaMallocAsync bypasses NVML memory queries that fail on MIG GPU instances
+os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "backend:cudaMallocAsync")
+
 import gradio as gr
 import numpy as np
 import spaces