Câu 52: NCP-AII: NCP-AI Infrastructure

During a 48-hour NeMo question-answering model burn-in test, GPU memory errors occur when processing large datasets. Which configuration strategy prevents Out-of-Memory (OOM) errors while maintaining processing efficiency?

1. A. Increase files_per_partition to 1000 for larger batch processing.
2. B. Switch from FP16 to FP32 precision for numerical stability.
3. C. Disable add_filename for Parquet files to reduce metadata.
4. D. Set blocksize= “1GB” for data loading and enable RMM asynchronous allocation. — đáp án hiện tại