single_shot: CUDA_LAUNCH_BLOCKING for debugging

single_shot_inference.py

@@ -17,6 +17,7 @@ This is the ground truth for vLLM / SGLang integration.
 """
 import os, sys, time, json, math, argparse, logging
 import torch
+os.environ['CUDA_LAUNCH_BLOCKING'] = '1'  # Catch CUDA errors synchronously
 import torch.nn.functional as F
 from pathlib import Path