The same input runs mindspore.ops.reciprocal on both CPU and GPU outputting one as INF and one as 0

[PhyllisJi]

import mindspore
import numpy as np

mindspore.context.set_context(device_target='GPU')

inp = np.load('/mnt/test_inp.npz', allow_pickle=True)['inp']
print(inp)

ms_inp = mindspore.Tensor(inp.astype(np.float32))

mindspore_layer = mindspore.ops.reciprocal(ms_inp)

print(mindspore_layer)

import mindspore
import numpy as np

mindspore.context.set_context(device_target='CPU')

inp = np.load('/mnt/test_inp.npz', allow_pickle=True)['inp']

ms_inp = mindspore.Tensor(inp.astype(np.float32))

mindspore_layer = mindspore.ops.reciprocal(ms_inp)

print(mindspore_layer)

The specific input data is in our GIthub repository https://github.com/PhyllisJi/MoCoDiff_Bug/blob/ms-issue-%23293/README.md

git clone -b ms-issue#293 https://github.com/PhyllisJi/MoCoDiff_Bug.git python test_cpu.py python test_gpu.py

[singularity6033]

Could you please tell us the version and commit id of the MindSpore used in this test case? I have tried master branch with commit id 2694987 (2.3.0), the result seems to be identical in both CPU and GPU.

[PhyllisJi]

Could you please tell us the version and commit id of the MindSpore used in this test case? I have tried master branch with commit id 2694987 (2.3.0), the result seems to be identical in both CPU and GPU.

Mindspore 2.2.14, I installed it from your official website.

[singularity6033]

I just tried MindSpore 2.2.14 in our official website, the result of GPU is a maximum of FP32 but not 0, can't reproduce your result, I checked the code, logic is right when input of the reciprocal is 0.

[PhyllisJi]

I just tried MindSpore 2.2.14 in our official website, the result of GPU is a maximum of FP32 but not 0, can't reproduce your result, I checked the code, logic is right when input of the reciprocal is 0.

Presumably this is hardware related? I was using an Nvidia 3090.