medium10 pts · first answer countsgpu-computingmixed-precision

TF32 on Ampere and newer

On an A100, what does torch.set_float32_matmul_precision("high") do for FP32 models?

A.Converts all weights to FP64 for accuracyB.Stores activations in FP8C.Lets FP32 matmuls use TF32 tensor coresD.Turns off tensor cores completely

Discussion 0

Join free to post in the discussion.