tf.keras.ops.numpy.rot90 returns incorrect values for two rotations of a rectangular plane

Issue type

Bug

Have you reproduced the bug with TensorFlow Nightly?

Yes

Source

source

TensorFlow version

tf 2.21.0, also reproduced on tf 2.22.0-dev20260904

Custom code

Yes

OS platform and distribution

Linux Ubuntu 22.04

Mobile device

No response

Python version

Python 3.13.5

Bazel version

No response

GCC/compiler version

No response

CUDA/cuDNN version

No response

GPU model and memory

No response

Current behavior?

tf.keras.ops.numpy.rot90 returns incorrect forward values for k=2 when rotating a rectangular plane.

A rotation with k=2 should be exactly equivalent to reversing along both selected axes. In the reproducer below, the input has shape (2, 2, 3) and rot90 is applied with axes=(1, 2), so the selected plane has unequal dimensions. The expected result is computed from the corresponding index permutation.

For the weighted selected output, the expected value is -577.0, but TensorFlow returns -232.0.

Expected behavior?

tf.keras.ops.numpy.rot90(x, k=2, axes=(1, 2)) should reverse the selected rectangular plane along both selected axes. For the reproducer below, the weighted selected output should be -577.0, not -232.0.

Standalone code to reproduce the issue

import os
os.environ["CUDA_VISIBLE_DEVICES"] = ""
os.environ["OMP_NUM_THREADS"] = "1"
os.environ["TF_NUM_INTRAOP_THREADS"] = "1"
os.environ["TF_NUM_INTEROP_THREADS"] = "1"

import tensorflow as tf

def target(t):
    x = tf.reshape(
        tf.stack([(i + 1) * t + i for i in range(12)]),
        (2, 2, 3),
    )
    y = tf.keras.ops.numpy.rot90(
        x,
        k=2,
        axes=(1, 2),
    )
    return y[0, 0, 0] - 2 * y[1, 1, 2] + 3 * y[1, 0, 1]

x = tf.constant(-24.0, dtype=tf.float64)

actual = target(x)
expected = -577.0

print("actual:", actual.numpy())
print("expected:", expected)

if abs(float(actual) - expected) > 1e-9:
    print("BUG REPRODUCED: tf.keras.ops.numpy.rot90 returns incorrect values for two rotations of a rectangular plane")
else:
    print("not reproduced")

Relevant log output

actual: -232.0
expected: -577.0
BUG REPRODUCED: tf.keras.ops.numpy.rot90 returns incorrect values for two rotations of a rectangular plane