[Fix Issue] fix issue 73692

[enkilee]

PR Category

Execute Infrastructure

PR Types

Bug fixes

Description

#73692

In pytorch:

import torch

x_cpu = torch.tensor([1, 1, float('nan'), float('nan')], dtype=torch.float32)
unique_cpu, counts_cpu = torch.unique(x_cpu, return_counts=True)
print("CPU Result:", unique_cpu, counts_cpu)

if torch.cuda.is_available():
    x_gpu = x_cpu.cuda()
    unique_gpu, counts_gpu = torch.unique(x_gpu, return_counts=True)
    print("GPU Result:", unique_gpu.cpu(), counts_gpu.cpu())
else:
    print("GPU not available")

The result is below:

CPU Result: tensor([1., nan, nan]) tensor([2, 1, 1])
GPU Result: tensor([1., nan, nan]) tensor([2, 1, 1])

So the value on pytorch when use cpu is same as gpu.

In paddle, maybe in the unique_functor.h:

std::set<InT> unique(in_data, in_data + in.numel());

The comparison behavior of std::set is undefined for NaN, which may lead to inconsistent results.

So try to fix it.

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[luotao1]

请 @ddchenhao66 审核下

[paddle-ci-bot]

Sorry to inform you that 9c51642's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

[enkilee]

/re-run all-failed

[enkilee]

/re-run all-failed