Segmentation fault (core dumped) in flow.nn.functional.log_softmax/flow.nn.functional.gumbel_softmax

[x0w3n]

Summary

A crash is triggered on the specified input, possibly due to a precision problem.

Code to reproduce bug

import oneflow as flow

eps = flow.tensor(1e-6, dtype=flow.float32)
target_sum = eps + eps 
x = flow.log(target_sum) - flow.log(target_sum - eps)
print(x)
log_softmax = flow.nn.functional.log_softmax(x)

import oneflow as flow

eps = flow.tensor(1e-6, dtype=flow.float32)
target_sum = eps + eps 
x = flow.log(target_sum) - flow.log(target_sum - eps)
print(x)
log_softmax = flow.nn.functional.gumbel_softmax(x)

output:

tensor(0.6931, dtype=oneflow.float32)
Segmentation fault (core dumped)

System Information

• What is your OneFlow installation (pip, source, dockerhub): pip
• OS: Ubuntu 22.04.3 LTS
• OneFlow version (run python3 -m oneflow --doctor):

path: ['/home/miniconda3/envs/oneflow/lib/python3.9/site-packages/oneflow']
version: 0.9.0
git_commit: 381b12c
cmake_build_type: Release
rdma: True
mlir: True

• Python version: 3.9.13
• CUDA driver version: 12.2
• GPU models: NVIDIA GeForce RTX 4090
• Other info: None