tensorflow gpu about tensorflow

[alexanderbeatson]

It is horrendous how hard it is to get this working on Windows.

Hello @Retalak
Please note that Tensorflow 2.10 was the last version that support GPU on Windows. If you want to use Tensorflow-gpu on Windows, you need to install Tensorflow in WSL2.

[Retalak]

It is horrendous how hard it is to get this working on Windows.

[sushreebarsa]

@soheil-asgari Could you verify if you have a compatible Nvidia GPU by running nvidia-smi in your terminal. If there's no output, you might not have an Nvidia GPU or the drivers might not be installed?
Thank you!

[alexanderbeatson]

I had the similar issue with Tensorflow GPU, but the wheel install from this wheel solved the issue.

Make sure your nvidia-smi running properly and your cuda version is compatible with Tensorflow 2.15 (Default 2.16 version is yet to be supported).

You can also search compatible wheel file here!

[zzj0402]

Python 3.9.19 (main, Mar 21 2024, 17:11:28) 
[GCC 11.2.0] :: Anaconda, Inc. on linux
NVIDIA-SMI 550.54.10              Driver Version: 551.61         CUDA Version: 12.4
Skipping registering GPU devices...
[name: "/device:CPU:0"
device_type: "CPU"
memory_limit: 268435456
locality {
}
incarnation: 6427527073389244107
xla_global_id: -1
]
TF Version
2.16.1

On WSL2 Windows 11 23H2 4090 laptop version. Followed pip install tensorflow[and-cuda] no luck

[Retalak]

I eventually got it working. Part of the problem with the documentation IMO is that it does not clearly state what is required on the host vs the guest machine for WSL2. I have the CUDA Toolkit 12.4 installed on my Windows host, not sure if that is necessary but I'm not messing with it now that I got it working. I did not install anything on the guest before running the above install command, it installed the needed versions of the toolkit and CUDNN (I also did a fresh install of my WSL2 - Ubuntu LTS 22.04.3 - to clear anything I had done previously). Another issue is there are a lot of error messages that are considered "normal" and may throw off new users. Here is the test command result on my WSL2 guest with it working (I think):

retalak@**-********:~$ python3 -c "import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))"
2024-04-08 16:07:56.717457: 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: AVX2 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
2024-04-08 16:07:57.579692: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
2024-04-08 16:07:58.688620: I external/local_xla/xla/stream_executor/cuda/cuda_executor.cc:984] could not open file to read NUMA node: /sys/bus/pci/devices/0000:01:00.0/numa_node
Your kernel may have been built without NUMA support.
2024-04-08 16:07:58.852701: I external/local_xla/xla/stream_executor/cuda/cuda_executor.cc:984] could not open file to read NUMA node: /sys/bus/pci/devices/0000:01:00.0/numa_node
Your kernel may have been built without NUMA support.
2024-04-08 16:07:58.852827: I external/local_xla/xla/stream_executor/cuda/cuda_executor.cc:984] could not open file to read NUMA node: /sys/bus/pci/devices/0000:01:00.0/numa_node
Your kernel may have been built without NUMA support.
[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

I also had to add this to the end of ~/.bashrc on my WSL2 guest (replace USER_NAME and make sure to exit WSL2 terminal and start a new one after saving and before running the test command):

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/USER_NAME/.local/bin;
export NVIDIA_DIR=$(dirname $(dirname $(python3 -c "import nvidia.cudnn;print(nvidia.cudnn.__file__)")))
export LD_LIBRARY_PATH=$(echo ${NVIDIA_DIR}/*/lib/ | sed -r 's/\s+/:/g')${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}

Note I am not using miniconda, so this may be different if you are trying to do this in a miniconda environment.

[sushreebarsa]

@soheil-asgari Could you please let us know any update on this issue?
Thank you!

[github-actions]

This issue is stale because it has been open for 7 days with no activity. It will be closed if no further activity occurs. Thank you.

[svkgn4DL]

This is still an issue. Documentation not states to set environment variable required for GPU like mentioned by @Retalak. If i remember correctly tensorflow used to have this documented earlier versions.