tf.raw_ops.WriteScalarSummary can segfault on GPU instead of raising a Python exception

Issue type

Bug

Have you reproduced the bug with TensorFlow Nightly?

Yes

Source

source

TensorFlow version

2.21.0

Custom code

Yes

OS platform and distribution

Ubuntu 24.04.3 LTS x86_64

Mobile device

No response

Python version

3.10.20

Bazel version

No response

GCC/compiler version

No response

CUDA/cuDNN version

13.0

GPU model and memory

No response

Current behavior?

System information

• TensorFlow version: 2.21.0
• Python version: 3.10.20
• OS: Linux 6.17.0-19-generic x86_64
• GPU: 4 x NVIDIA GeForce RTX 2080 Ti
• Reproduced in: conda env tf_test

Description

I found a case where tf.raw_ops.WriteScalarSummary crashes the Python process on GPU instead of raising a normal Python exception.

In batch validation this showed up as a generic signal_kill, but when I reran the same script standalone, it consistently reproduced as:

• Segmentation fault (core dumped)

This looks like a hard crash rather than expected input validation behavior.

Reproducer

import tensorflow as tf

def test_write_scalar_summary():
    logdir = "logs/"
    writer = tf.summary.create_file_writer(logdir)

    step = tf.constant(1, dtype=tf.int64)
    tag = tf.constant("test_tag", dtype=tf.string)

    values = [
        tf.constant(0.0, dtype=tf.float32),
        tf.constant(0.0, dtype=tf.float64),
        tf.constant(0, dtype=tf.int32),
        tf.constant(0, dtype=tf.int64),
        tf.constant(0, dtype=tf.float16),
        tf.constant([], dtype=tf.float32),
        tf.constant([[0]], dtype=tf.float32),
        tf.constant([[0, 0]], dtype=tf.float32),
    ]

    with writer.as_default():
        for value in values:
            try:
                tf.raw_ops.WriteScalarSummary(
                    writer=writer._resource,
                    step=step,
                    tag=tag,
                    value=value,
                )
                print(f"Successfully wrote summary for value: {value.numpy()}")
            except Exception as e:
                print(f"Failed to write summary for value: {value.numpy()} with error: {e}")

    writer.close()

with tf.device('/GPU:0'):
    test_write_scalar_summary()


### Standalone code to reproduce the issue

```shell
conda run --no-capture-output -n tf_test python reproduce.py
Actual behavior
The process terminates with a segfault:

Segmentation fault (core dumped)
Before the crash, TensorFlow initializes GPU devices normally.

Expected behavior
tf.raw_ops.WriteScalarSummary should reject unsupported inputs by raising a normal Python / TensorFlow exception such as InvalidArgumentError or ValueError, rather than crashing the whole process.

pip list
Package                  Version
------------------------ ----------
absl-py                  2.4.0
aiohappyeyeballs         2.6.1
aiohttp                  3.13.3
aiosignal                1.4.0
annotated-types          0.7.0
anyio                    4.12.1
astunparse               1.6.3
async-timeout            5.0.1
attrs                    25.4.0
cachetools               7.0.5
certifi                  2026.2.25
charset-normalizer       3.4.5
colorlog                 6.10.1
distro                   1.9.0
exceptiongroup           1.3.1
flatbuffers              25.12.19
frozenlist               1.8.0
gast                     0.7.0
google-pasta             0.2.0
grpcio                   1.78.0
h11                      0.16.0
h5py                     3.14.0
httpcore                 1.0.9
httpx                    0.28.1
idna                     3.11
jiter                    0.13.0
keras                    3.12.1
libclang                 18.1.1
markdown-it-py           4.0.0
mdurl                    0.1.2
ml_dtypes                0.5.4
multidict                6.7.1
namex                    0.1.0
networkx                 3.4.2
numpy                    2.2.6
nvidia-cublas-cu12       12.9.1.4
nvidia-cuda-cupti-cu12   12.9.79
nvidia-cuda-nvcc-cu12    12.9.86
nvidia-cuda-nvrtc-cu12   12.9.86
nvidia-cuda-runtime-cu12 12.9.79
nvidia-cudnn-cu12        9.20.0.48
nvidia-cufft-cu12        11.4.1.4
nvidia-curand-cu12       10.3.10.19
nvidia-cusolver-cu12     11.7.5.82
nvidia-cusparse-cu12     12.5.10.65
nvidia-nccl-cu12         2.29.7
nvidia-nvjitlink-cu12    12.9.86
openai                   0.28.0
opt_einsum               3.4.0
optree                   0.19.0
packaging                26.0
pip                      26.0.1
propcache                0.4.1
protobuf                 7.34.0
pydantic                 2.12.5
pydantic_core            2.41.5
Pygments                 2.19.2
requests                 2.32.5
rich                     14.3.3
setuptools               82.0.1
six                      1.17.0
sniffio                  1.3.1
tensorflow               2.21.0
termcolor                3.3.0
tqdm                     4.67.3
typing_extensions        4.15.0
typing-inspection        0.4.2
urllib3                  2.6.3
wheel                    0.46.3
wrapt                    2.1.2
yarl                     1.23.0

Relevant log output

(tf_test) user@user-Standard-PC-Q35-ICH9-2009:/home/lhj$ python callChainBuild/src_tensorflow/output_tensorflow/targeted_mutation/validation_sharded_full_20260322/potential_bugs/tf_raw_ops_WriteScalarSummary/0.py
WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1774265664.422049 3732488 port.cc:153] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
I0000 00:00:1774265664.492746 3732488 cpu_feature_guard.cc:227] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: AVX2 AVX512F AVX512_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1774265665.815049 3732488 port.cc:153] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
I0000 00:00:1774265667.333059 3732488 gpu_device.cc:2043] Created device /job:localhost/replica:0/task:0/device:GPU:0 with 9606 MB memory:  -> device: 0, name: NVIDIA GeForce RTX 2080 Ti, pci bus id: 0000:01:00.0, compute capability: 7.5
I0000 00:00:1774265667.335750 3732488 gpu_device.cc:2043] Created device /job:localhost/replica:0/task:0/device:GPU:1 with 9606 MB memory:  -> device: 1, name: NVIDIA GeForce RTX 2080 Ti, pci bus id: 0000:02:00.0, compute capability: 7.5
I0000 00:00:1774265667.338406 3732488 gpu_device.cc:2043] Created device /job:localhost/replica:0/task:0/device:GPU:2 with 9606 MB memory:  -> device: 2, name: NVIDIA GeForce RTX 2080 Ti, pci bus id: 0000:03:00.0, compute capability: 7.5
I0000 00:00:1774265667.341101 3732488 gpu_device.cc:2043] Created device /job:localhost/replica:0/task:0/device:GPU:3 with 9606 MB memory:  -> device: 3, name: NVIDIA GeForce RTX 2080 Ti, pci bus id: 0000:04:00.0, compute capability: 7.5
E0000 00:00:1774265667.602828 3732488 util.cc:131] oneDNN supports DT_HALF only on platforms with AVX-512. Falling back to the default Eigen-based implementation if present.
Successfully wrote summary for value: 0.0
Successfully wrote summary for value: 0.0
Successfully wrote summary for value: 0
Successfully wrote summary for value: 0
Successfully wrote summary for value: 0.0
Segmentation fault (core dumped)