https://blog.ans20xx.com/tags/amp/Recent content in AMP on 安橙的博客Hugo -- 0.163.3zhTue, 26 May 2026 00:00:00 +0000https://blog.ans20xx.com/posts/ai/day12/Tue, 26 May 2026 00:00:00 +0000https://blog.ans20xx.com/posts/ai/day12/拆开混合精度训练:FP32 / TF32 / FP16 / BF16 / FP8 的数值范围、精度与溢出风险;理解 torch.amp.autocast 的算子选择、GradScaler 的动态 loss scaling,并跑通一个可对比的 AMP benchmark。