tf.pow returns inconsistent value on CPU vs GPU #97125
Description
Issue type
Bug
Have you reproduced the bug with TensorFlow Nightly?
Yes
Source
source
TensorFlow version
2.20.0-dev20250716
Custom code
Yes
OS platform and distribution
Linux Ubuntu 24.04
Mobile device
No response
Python version
3.12
Bazel version
No response
GCC/compiler version
No response
CUDA/cuDNN version
No response
GPU model and memory
No response
Current behavior?
tf.pow returns inconsistent values when using int64 tensors on CPU vs CUDA. The CPU value matches the value calculated by numpy.
Also reproducible with 2.19.0 (colab)
Standalone code to reproduce the issue
import tensorflow as tf
import numpy as np
print(tf.__version__) # 2.20.0-dev20250716
x = tf.constant([-48], dtype=tf.int64)
y = tf.constant([66], dtype=tf.int64)
with tf.device('/cpu:0'):
output_cpu = tf.pow(x=x, y=y)
print("Output on CPU:", output_cpu) # 0
with tf.device('/gpu:0'):
output_gpu = tf.pow(x=x, y=y)
print("Output on GPU:", output_gpu) # 2304
output_np = np.power(x.numpy(), y.numpy())
print("Output with NumPy:", output_np) # 0Relevant log output
2.20.0-dev20250716
Output on CPU: tf.Tensor([0], shape=(1,), dtype=int64)
Output on GPU: tf.Tensor([2304], shape=(1,), dtype=int64)
Output with NumPy: [0]