Fetching only the most critical high-precision data from slower storage during inference to recover accuracy without stalling generation.