tf.keras.ops.numpy.sinc returns NaN gradient at zero where derivative should be zero

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.sinc returns nan for the first derivative at a point where the normalized sinc function has a smooth extension and the derivative should be finite.

In the reproducer below, the function evaluates tf.keras.ops.numpy.sinc(2 * t + 1) at t = -0.5, so the argument to sinc is exactly zero. The forward value is correctly returned as 1.0. However, TensorFlow autodiff returns nan for the derivative with respect to t.

The expected derivative is 0.0, including the chain rule through 2 * t + 1.

Expected behavior?

The first derivative of tf.keras.ops.numpy.sinc(2 * t + 1) at t = -0.5 should be 0.0, not nan.

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):
    return tf.keras.ops.numpy.sinc(2 * t + 1)

x = tf.constant(-0.5, dtype=tf.float64)

with tf.GradientTape() as tape:
    tape.watch(x)
    y = target(x)

actual = tape.gradient(y, x)
expected = 0.0

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

if tf.math.is_nan(actual):
    print("BUG REPRODUCED: tf.keras.ops.numpy.sinc returns NaN gradient at zero")
else:
    print("not reproduced")

Relevant log output

forward: 1.0
actual: nan
expected: 0.0
BUG REPRODUCED: tf.keras.ops.numpy.sinc returns NaN gradient at zero