Closed Redmept1on closed 3 months ago
oneflow.log10 has accuracy differences on the cpu and cuda, but this issue does not seem to exist on other log apis
import oneflow as flow import numpy as np x = flow.tensor(np.array([[float('inf'), 0, -1, float('nan'), 5]], dtype=np.float32)) x = x.cuda() y = flow.log10(x) print(y.device,y) x=x.cpu() y1 = flow.log10(x) print(y1.device,y1)
Summary
oneflow.log10 has accuracy differences on the cpu and cuda, but this issue does not seem to exist on other log apis
Code to reproduce bug
System Information