Removal of logs from Phase 2 runs

I’m currently having an issue where my PyTorch based models fail to inference despite reproducing the environment and the local evaluation script working correctly. Without the logs, how do you suggest debugging the problem?