Segmentation fault (core dumped) in flow.IntTensor

[deepliao]

Summary

When attempting to create or fill IntTensor with unsupported data types (e.g., strings or floats), OneFlow does not raise a Python-level error, but instead leads to a segmentation fault (core dumped). This differs from frameworks like PyTorch, which correctly raise type validation exceptions. This is a critical robustness flaw in OneFlow's tensor constructor and input validation, potentially causing silent data corruption or unsafe crashes.

Code to reproduce bug

import oneflow as flow

# Set fixed seed for reproducible behavior
flow.manual_seed(42)

# Example 1: Filling IntTensor with float values
int_tensor = flow.IntTensor([[1, 2], [3, 4]])
print(int_tensor)  # Expected: safe output

# This line would raise an error in PyTorch, but may not in OneFlow
# Potential silent downcast or crash
# int_tensor.fill_(3.14)

# Example 2: Initialize IntTensor with non-integer (invalid) type
invalid_type = 'invalid'
try:
    int_tensor = flow.IntTensor(invalid_type)
except Exception as e:
    print(f"Error during invalid type initialization: {e}")

output:

tensor([[1, 2],
        [3, 4]], dtype=oneflow.int32)
Segmentation fault (core dumped)

System Information

OneFlow installation: pip OS: Ubuntu 20.04 OneFlow version (run python3 -m oneflow --doctor):

path: ['/root/miniconda3/envs/myconda/lib/python3.8/site-packages/oneflow']
version: 0.9.0
git_commit: 381b12c
cmake_build_type: Release
rdma: True
mlir: True

Python version: 3.8 CUDA driver version: 11.6 GPU models: NVIDIA A16