Issue type

Bug

Have you reproduced the bug with TensorFlow Nightly?

No

Source

binary from pip

TensorFlow version

2.16.1

Custom code

Yes

OS platform and distribution

Linux Ubuntu 24.04

Mobile device

No response

Python version

3.12.3

Bazel version

No response

GCC/compiler version

No response

CUDA/cuDNN version

No response

GPU model and memory

No response

Current behavior?

I believe I tracked down a memory leak in my code to a tf.constant creation in a short lived object of a class. I can reproduce it in an even simpler way by just creating the constant in a loop. No @tf.function decoration or model training necessary to cause it. It seems to happen if I replace the tf.constant with a tf.random.uniform(). I tried suggestions I've seen elsewhere, like trying to use del on the variable followed by gc.collect().

Some similar looking bug reports:

• Ram memory leak when using tf.function at tf 2.8,2.9 and 2.10, doesn't happen with tf 2.5 #57982
• Memory leak in custom training loop + tf.function  #50765
• https://discuss.tensorflow.org/t/tensorflow-memory-leak-during-inference-in-loop/23044

It looks like upgrading to 2.15 solved it for some people, but it appears to be back in 2.16?

is this expected and I'm doing something wrong?

Standalone code to reproduce the issue

import time
import psutil
import tensorflow as tf


def loop_copy():
    for _ in range(2000):
        y = tf.constant(5.0)  # still leaks with this


start_mem = psutil.Process().memory_info().rss

for _ in range(500):
    time.sleep(0.1)
    print((psutil.Process().memory_info().rss - start_mem) / 10**6)
    loop_copy()

### Relevant log output

```shell
0.0
6.029312
6.815744
7.471104
8.126464
8.781824
9.437184
10.092544
10.747904
11.534336
12.189696
12.845056
13.500416
14.155776
14.811136
15.597568
16.252928
...