search

Oneflow-Inc / oneflow

OneFlow is a deep learning framework designed to be user-friendly, scalable and efficient.
http://www.oneflow.org
Apache License 2.0
5.79k stars 658 forks source link

oneflow.max/min perform differently between cpu and cuda. #10475

Open Redmept1on opened 3 months ago

Redmept1on commented 3 months ago

Summary

oneflow.max/min perform differently between cpu and cuda.

Code to reproduce bug

oneflow.max

import oneflow as flow
import numpy as np

x1 = flow.tensor(np.array([[float('inf'), 0, -1, float('nan'), 5]], dtype=np.float32))
x1 = x1.cuda()
y1 = flow.max(x1,dim=1)
print(y1)

x1 = flow.tensor(np.array([[float('inf'), 0, -1, float('nan'), 5]], dtype=np.float32))
x1 = x1.cpu()
y2 = flow.max(x1,dim=1)
print(y2)

image

oneflow.min

import oneflow as flow
import numpy as np

x1 = flow.tensor(np.array([[float('inf'), 0, -1, float('nan'), 5]], dtype=np.float32))
x1 = x1.cuda()
y1 = flow.min(x1,dim=1)
print(y1)

x1 = flow.tensor(np.array([[float('inf'), 0, -1, float('nan'), 5]], dtype=np.float32))
x1 = x1.cpu()
y2 = flow.min(x1,dim=1)
print(y2)

image

System Information

Redmept1on commented 3 months ago

oneflow.median has the same bug

import oneflow as flow
import numpy as np

x1 = flow.tensor(np.array([[float('inf'), 0, -1, float('nan'), 5]], dtype=np.float32))
x1 = x1.cuda()
y1 = flow.median(x1,dim=1)
print(y1)

x1 = flow.tensor(np.array([[float('inf'), 0, -1, float('nan'), 5]], dtype=np.float32))
x1 = x1.cpu()
y2 = flow.median(x1,dim=1)
print(y2)

image