No compatibility with python 3.13.5 version for tensorflow-rocm wheel package

[Potomac]

Issue type

Build/Install

Have you reproduced the bug with TensorFlow Nightly?

Yes

Source

binary

TensorFlow version

2.19.0

Custom code

Yes

OS platform and distribution

Arch Linux 64 bits

Mobile device

No response

Python version

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?

I use Arch Linux 64 bits,
the python version provided with Arch Linux is 3.13.5,

I can not install tensorflow-rocm with pip, as the current most recent tensorflow-rocm package has no compatibility with python 3.13.5 :

tf_nightly_rocm-2.19.0.dev20250408-cp312-cp312-manylinux_2_28_x86_64.whl

The repo I use is :
https://repo.radeon.com/rocm/manylinux/rocm-rel-6.4/

The error message I have when I try to install tensorflow-rocm with pip :

ERROR: tf_nightly_rocm-2.19.0.dev20250408-cp312-cp312-manylinux_2_28_x86_64.whl is not a supported wheel on this platform.

Can you please create a tensorflow-rocm package with compatibility for python 3.13.5 ?

Thanks

Standalone code to reproduce the issue

pip install tensorflow-rocm

or 

pip3 install wheel /tmp/tf_nightly_rocm-2.19.0.dev20250408-cp312-cp312-manylinux_2_28_x86_64.whl

Relevant log output

ERROR: tf_nightly_rocm-2.19.0.dev20250408-cp312-cp312-manylinux_2_28_x86_64.whl is not a supported wheel on this platform.

[harkgill-amd]

Hi @Potomac, Arch Linux seems to be one of the only distros to have adopted Python 3.13 by default. As the release matures and as upstream Tensorflow begins support https://www.tensorflow.org/install/pip#software_requirements, I'm sure we'll have Python 3.13 packages start to roll out. For now, can you try using a virtual environment or an environment management system such as conda to downgrade your Python version?

[Potomac]

Hello harkgill-amd,

thanks for your answer, I will try conda for using a virtual environment with an older version of python.

[harkgill-amd]

@Potomac, any luck with using conda to downgrade your python version?

[Potomac]

I made a virtualenv with conda, and installed python 3.12 :

source /opt/anaconda/bin/activate 
conda create -n tf-rocm-env python=3.12
conda activate tf-rocm-env
pip install  tensorflow-rocm==2.18.1 -f https://repo.radeon.com/rocm/manylinux/rocm-rel-6.4.3/ --upgrade

export ROCM_PATH=/opt/rocm
export  HSA_OVERRIDE_GFX_VERSION=11.0.0 
export KERAS_BACKEND=tensorflow

but when I run a test python source code for tensorflow-rocm I get an error :

double free or corruption (!prev)
I tried also tf-nightly-rocm 2.19.0.dev20250807 --> same problem.

I tried also "export HSA_OVERRIDE_GFX_VERSION=11.0.2" (instead 11.0.0) : no change, still the same error.

My rocm version is 6.4.3 on Arch Linux, my graphic card is AMD Radeon RX 7600.
Here is the output of rocminfo in attached file.

rocminfo.txt

[Potomac]

The test python tensorflow-rocm file I use in attached file.

test_tensorflow.py

[Potomac]

It seems that this instruction "print(tf.config.list_physical_devices())" triggers the error " double free or corruption (!prev)" :

$ python
import tensorflow as tf
2025-08-29 02:05:14.644483: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: SSE3 SSE4.1 SSE4.2 AVX AVX2 AVX512F AVX512_VNNI AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.

print(tf.config.list_physical_devices())

double free or corruption (!prev)
Abandon python

My kernel linux version is : 6.16.2.