ResearchBenchmarksOptimizer State Memory Measurement on DeepSeek-V2-Lite (16B) (train)Follow55.3Average Optimizer State Memory (MB)PowerStep40.352141.251242.15343.049May 11, 2026Evaluation ResultsMethodMethodLinksAverage Optimizer State Memory (MB)PowerStepPrecision=int8, Number...Precision=int8, Number of NPUs=2562026.0555.3AdamWPrecision=fp32, Number...Precision=fp32, Number of NPUs=2562026.05429