Install Tensorflow/Keras in WSL2 for Windows with NVIDIA GPU
ฝัง
- เผยแพร่เมื่อ 8 ก.ย. 2024
- Welcome to this tutorial on how to install TensorFlow/Keras for use with a GPU on Windows! In this video, we will guide you through the process of setting up TensorFlow/Keras to utilize the power of your GPU, specifically an NVIDIA RTX 6000 (Ada).
Before we get started, it's important to note that the current versions of TensorFlow can only be installed using Windows Subsystem for Linux 2 (WSL2). So, if you haven't already installed WSL2, we will show you how to do that as well.
Once you have WSL2 set up, we will guide you through the steps of installing the necessary drivers, setting up your environment variables, and installing TensorFlow/Keras with GPU support. We'll also show you how to verify that everything is working correctly.
As mentioned earlier, we'll be using an NVIDIA RTX 6000 (Ada) for this tutorial, which was kindly provided by NVIDIA. However, the steps we cover should be applicable to other NVIDIA GPUs as well.
By the end of this tutorial, you'll have a fully functioning installation of TensorFlow/Keras with GPU support on your Windows machine, ready to take on your machine learning tasks with lightning-fast performance.
So, grab a cup of coffee and join us as we dive into the world of TensorFlow/Keras with GPU support on Windows!
TensorFlow Install Docs
www.tensorflow...
NVIDIA RTX 6000 Ada
www.nvidia.com...
NVIDIA Driver
www.nvidia.com...
0:19 State of Windows TensorFlow in 2023
1:34 NVIDIA Driver
3:00 Install WSL2
3:31 Install Miniconda (Python)
4:42 Create Virtual Environment
5:27 Check GPU
5:37 Install CUDA
6:46 System Path
7:32 Install Tensorflow
8:19 Verify Install
![[TH] 2024 PMSL SEA Finals D2 | Fall | ต้องรักษาฟอร์ม อย่ายอมให้ใครแซง](http://i.ytimg.com/vi/x38QsGn7A3M/mqdefault.jpg)
2024, and this video is still relevant!
The installation instructions at tensorflow website have changed, and I couldn't get any GPU support to work with them. The "old" command flow listed in this video still works just fine!
One note: after following your instructions, I had to downgrade Tensorflow to version 2.10.
With 2.11 it would install and could be loaded correctly, but then it would crash when trying to train or run a model.
You sir are a true hero. Thank you so much!
PERFECT!
JUST DONE!
This is the best procedure that i've found on internet.
Remarks: the TF web page now has new instruction for the GPU set-up section. I used these new ones and all goes well (10/04/2023).
Thank you Jeff.
Readers should realise, but 10/4/2023 :)
You got this done without the old method with MSVS, Cuda, CuDNN installation? Thanks.
Bro I freaking love you!! Thank you!
As a tip for everyone trying to install this with gpu support, stick to the versions he's using in the video. I followed the steps but used the version in the TF instructions and it failed.
You got this done without the old method with MSVS, Cuda, CuDNN installation? Thanks.
Jeff, I cannot thank you enough. You've not only shown us the entire installation process but also explained it. Thanks a lot!
Your work on this subject and this simplified tutorial for installing CUDA and Tensorflow is highly appreciated Professor. Thank you
Thanks for the new tutorial! Been struggling for the last 3 weeks to have tensorflow detect my laptop's dedicated gpu. It only sees my integrated gpu. I've gone through many forums, web searches etcc.. and I've tried at least 5 different tutorials including your old ones without success. I'll try this one tonight. God, I hope it works because this is driving me nuts and it's become an obsession of mine to get it to work 😅
Isn't this a BIOS problem? I think i had something similar with my laptop... 🤔
Anyway, you could force a CUDA-GPU in nVidia Control Panel -> Manage 3D Settings -> Program Settings.
BUT i'm not sure if it works with WSL.
@@eMgotcha77Bios problem, could you be more specific? Also, how do you force a CUDA-GPU in Nvidia control panel. I've tried various settings permutations in the nvidia control panel but it hadn't solved my issue in the past. Also, not sure if it's related but running the nvidia-smi command brings up my dedicated gpu but the "volatile uncorr. ECC" value shows "N/A" which is not what I see on other nvidia-smi output screenshots in various forums.
Hey @Malleyka I am having the exact same problem and it's driving me crazy. I run tf in WSL on my laptop and when I monitor the nvidia GPU there is 0% activity. If you have found a solution I would greatly appreciate any tips you have to share!
@@spencerduncan7006 Bro, I kept trying other things afterwards and the solution I ended up going for was drastic. Drastic as in getting a new machine and listing the old one for sale. On the new machine, I followed the same instructions and got it to work without any issues in 30mn or so. Not what I was hoping for but it came to that in the end.
You got this done without the old method with MSVS, Cuda, CuDNN installation? Thanks.
Hi Dr. Heaton. I was watching your videos last year and messed around with TensorFlow on my own. I just finished up an MSCS bridge program and am now in an AI program. Thanks for getting me on this path.
As others have said, perfect video. Interesting thing of note - I found your video by giving GPT4 with VoxScript the following prompt:
I want to learn more about creating a lstm in pycharm inside wsl2 to have tensorflow use my gpu. Please find a list of no more than 10 youtube videos that will expand my knowledge to an expert's level on these topics.
Your video was the number one result. This wasn't the entire chat history but I found it interesting regardless!
Jeff! Thank you so much for the information. It's so much better than rifling through the docs and GitHub issues. Subscribed!
Great, thanks for this. May 21, current versions give Python 3.10 and TF 2.12, and a 4090 shows up fine after install; reboot wasn't required and followed the TF notes verbatim. Whether running models will work that I built and have been using for a couple of years with early TF2 and a 1070 will be another matter! Also created a .wslconfig in the Windows home dir to increase RAM from the default of 50% physical with:
[wsl2]
memory=56GB
processors=8
swap=16GB
Oh gooosh. It has just solved the crazy problem I was dealing with. Thank you 👍👍👍
Thank you for producing this video!
Thanks man this was a good video. Nothing as frustrating as installing Tensorflow on windows.
10 months later this is slightly dated compared to the current instructions on the TensorFlow site, but it's still close enough that I was able to figure it out. Thanks!
Hi! Can you explain what you did that made it work because I am getting stuck at the pip install tensorflow==2.11 step, it wont let me install tensorflow 2.11, but the process preceding it has gone just as explained in the video
It’s hard to say for sure from this information, but one thing that I encountered a lot of issues with is that both the WSL2 install and any virtual environments within WSL2 are completely clean slates that don’t have any of the files, folders, or programs of your main windows machine. So for example, if you want to pip install something in WSL2 or a virtual environment within WSL2, you would first need to install pip in that environment.
ChatGPT is also really good at explaining computer error messages, especially if you supplement them with the context of what you’re trying to do
This is a brillant demonstration. Thank you.
Thank you so much, man! It worked with a few tweaks. I used the newest versions of cuda and tf, it worked fine!
The best of the videos I've seen! Thank you!
Hi Jeff, thank you very much for this video. You too all the guess work out of the process which, for me is the path of least resistance 🙂
Fantastic video. clearly explains everything that you need to know.
So Py Torch 2.0 is a much better fit with Windows? Thank you Path of Least Resistance!
PyTorch install is a copy/paste from their homepage.... TensorFlow is about 5 pages of instructions that usually works (except when it doesn't). Plus TensorFlow is bailing from just about every aspect of Windowsd.... (but I am not frustrated with their direction, at all 😆😠)
This really helped me. Thanks a lot Dr. Heaton
Thank you very much !!
I love your detailed explanations.
Eran
It's very helpful guide!
Thanks very much 😉
Thank you Jeff.
That worked flawlessly thank you
on wsl2 windows 11, after installing, i get the error after importing tensorlow
ModuleNotFoundError: No module named 'tensorflow'
Can someone help? It's urgent
i'm also having the same issue , did you solved it?
Make sure that you have activated the "tf" environment before running the testing script, otherwise it will run from the base environment which is not the one which we installed tensorflow in.
To enter tf environment simply run; conda activate tf
Then run the testing script to see if it worked
I think they already update their docs. I don't found the miniconda part.
hi
i get this error when i type "nvidia-smi"
NVIDIA-SMI has failed because it couldn't communicate with the NVIDIA driver. Make sure that the latest NVIDIA driver is installed and running.
I had same problem. To Fix, I needed to set my container from WSL1 to WSL2. This is done like this: wsl --set-version 2
Thank you so much!!! I have been struggling on how to setup tf GPU for wsl
It shows an empty list at the end :(
How do i start fresh by uninstalling everything and reinstalling.. i have tried today multiple times but each time same results.. i do have a compatible gpu… please help
It is absurdly difficult to get this working properly.
THANKS A LOT JEFF !!!!
7:45 I don't get it. You said "do not install 2.11" as instructed, but then you install 2.11 as instructed?
How do I active this newly created environment in Pycharm?
Was thinking the same thing.
Good job, its worked
thank you good dude.
I installed everything the same way as it's said in official Tensorflow site, however, the system path are not automatically configured, so I have to do it manually every time. Please, help
Follow the section instructions after setting environment paths where it instructs to create activation file and more two commands given for automatic configuration.
I have got an error when running the final command to check whether the gpu is activated or not. The error is “AttributeError: module ‘tensorflow.python.platform.self_check’ has no attribute ‘preload_check’. How to solve this??
not works. I installed all components as video says. Tensorflow not seen GPU. I lose many times to install necessary libs, as shown at video end. I got: could not open file to read NUMA node: /sys/bus/pci/devices/0000:03:00.0/numa_node. I hope i can solve this over metod.
I am bit confused, so do I have to install jypter inside wsl and write code there? but all my file present in my local windows system
dank je wel!
what happens when I use cmd instead of power shell?
Link for miniconda installation site pls
Ok but how do I use the conda environment that we just created in wsl if I'm writing code and using notebooks in e.g. VSC in Windows?
Activate the conda environment in vscode. Use Jupyter
@@jagadeeshk6652 yeah thanks but what I meant is how do you refer to the venv you created in WSL while using VSC on Windows. After a bit of research I figured out you need to use the WSL extension in VSC
@@xntumrfo9ivrnwf It was so critical point man, and thank you for your reply here!! Everything is done!
@@xntumrfo9ivrnwf Does it work for you? Whenever I want to fit my model, it says "libdevice.10.bc" not found
@@ichbin1984 hey yes it does, but usually not without having to figure out random errors etc. Just today I had to change my torch version… see if there are similar issues on google
Do any of those errors on your screen mean anything important? I have the same ones and I am not sure whether I need to fix it or whether tensorflow should be able to run using my gpu just fine
Hi, In powershell it shows gpu devices, but in Jupiter nootbook it is not showing me gpu device, please help
I'm running Windows 11, directML, tensorflow2.12, python 3.10 just fine. the only funky buisness is tensorflow databases, I can only get 3.2.1 (latest v3) to work, tfds 4 is just buggy. So even though Windows native DML claims not to work natively, it's defnitely worth a try. The nice part is it works for my AMD RX6800 and RTX 3080! the 6800 has 16 gigs of VRAM - sweet!!!
Thanks!
guess what
my 'wsl --install' command having errors
Is there a video relevant to install all these in 2024. Some websites are different now.
I have an integrated GPU and a GTX 1650 in my laptop. At the end, was I supposed to get 2 gpu's as a return?
Thank you for providing an amazing video but how to use jupyter notebook after all the installation is done so that I can use the gpu?
Very Very Great 🚀🚀
Do you have a tutorial for working this with Rstudio on a windows system with GPU?
I've been looking into Rstudio-server on WSL with Tensorflow and Keras but having no luck installing it.
I followed all instructions in the video but no gpu listed
actually he has no idea about the situation of viewer. we are literally struggling but at the end he are begging for subscribe. but did not solve our problem. rubbish....
Hi Jeff, I have native windows 11, AMD Ryzen CPU, nvidia 3060 GPU. I tried all of your videos, but I am still stuck. which version or Pytorch and Tensorflow should I install, so they both can run smoothly, which CUDA and CUDAnns should I install ?
THANKS
well explained, thanks a lot.
by the way do you have another tutorial about the same thing but for ubuntu 22.04.
Can I get your keyboard model? I like the sound it makes! would like to buy one
sir, after installing run a small Deep Learning Model on GPU like as the part of tutorial so that in case we can check for runtime problems
how to use jupyter notebook by the same, after all the installation is done?
Is there any way to also install jupyter notebook on wsl2 and run deep learning on GPU kernel ?
hey so do i have to repeat this all steps for a new project or venv
This didn't work for me and I followed it step by step (on windows 11). I went for Windows native with tensorflow 2.10 instead for the moment.
At 7:11 you (and TF) describe how to set the enviroment variable and make it more convienient so you dont have to do so every reboot... / restart of wsl...
Well... this works absolutly not for me... and I am going crazy at this point trying to fix it...
I have the file in the path... and have "export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CONDA_PREFIX/lib/:$CUDNN_PATH/lib" written inside... (If I check with nano)
Does anyone have insight into that?
I got the same issue!! I solved by editing the *env_vars.sh* (as you did) and added the line:
CUDNN_PATH=$(dirname $(python -c "import nvidia.cudnn;print(nvidia.cudnn.__file__)"))
before export LD... worked perfect at the end. This creates the CUDNN_PATH variable everytime it reboots.
Oh hey are you able to make an updated review I think stuff has changed
Thanks but i am not sure how can i integrate it to vs code now
Me too I am trying to use the new environment with Pycharm but I can't, If you found a solution to this problem please tell me
@@hamzaazeem4602 me too, did you find any solution?
@@hamzaazeem4602 anything new?
@@mirrorslav5803 I found a way to use and install cuda environment inside a linux environment, But I could’t use it outside of jupyter notebook and I couldn’t access any folders or files stored in windows
After Following that , how to use this to run pycharm ?
I am SO FRUSTRATED that there is no native support. I do not think that I am ever willing to move into Linux world. Found it frustrating, overly difficult and taking long hours during the night in the university, so I basically gave up on Linux years and years ago. Maybe that was a mistake I now find, but I am just too damn old to move over anymore.
Hi thanks for video, do we have github link for this?
No, sorry, but I do have links to the source articles in the description.
Excellent! With this, I do not need to do the MSVS, CUDA, CuDNN installation? If so, this is so much simpler. Thanks.
P/S Did the steps and finally got device:GPU:0 !!! I was about to give up. Many thanks!
Appreciate help on:
A. I did it the old way installing MSVS, Cuda, CuDNN. Not successful. Can I now uninstall MSVS? I assume with WSL2, no need for C compiler now in MSVS and thus can uninstall MSVS? MSVS takes up 29 GB!
B. Can I remove the installed Cuda toolkit and CuDNN directories from my old method attempt?
C. Tried to run this " import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))" in Pycharm but got error - AttributeError: module 'tensorflow' has no attribute 'config'. How can I enable tensorflow GPU Cuda access in Pycharm?
Many thanks! Finally.
PP/S Found out that only Pycharm Pro can use WSL interpreter. I am using Pycharm community. Pity.
I generally followed these instructions (2/27/24), but had to use conda to install tensorflow to get the test print() command to run. Mixing pip and conda to install into a conda environment is problematic. Pip did the install for me, but it went into a different location than conda installs to, and that location wasn't on the search path. (In fact, pip installed tensorflow into ~/.local/lib/python3.10/site-packages, outside of any conda environment.
Try to open the Python Env from the Linux WSL in a Kernel on VS Code. When you do this, just run the "%pip install library" (with the % in front) on the Code cell, it should install the libs on the path of the Python Enviroment.
after all this how can I run the code from VS code
How can I address this python evironment to vs code?
One key step is: reboot before test the tf installation.
thank you thank you thank you
It is not working anymore even though it worked earlier. (28 March 2024)
and now how to connect the environment into the code editor like jupyter notebook or pycharm? like i little bit confuse about is it integrated with windows application?
I am using Pycharm community. Just learned that only Pycharm Pro can use WSL interpreter. Pity!
hello when i verify install, this message appears "Could not find cuda drivers on your machine, GPU will not be used." i install the tensorflow 2.13, cudnn 8.6 and cuda 11.8 and i have rtx3060ti with the NVIDIA STUDIO, what is the possible problem?
Hi facing same problem. Could you solve it?
Dear man! please help to use opencv in the environment you teached, I think there is some crash with it but I cant solve!!
GODLIKE
I have tried this setup in two machines and i keep getting issues about “no numa support”. Have looked around for solutions, nothing worked so far. Anybody else having this issue?
same error kept showing me also
@@krishnagoyal7351try adding this environment var.
export TF_ENABLE_ONEDNN_OPTS=0
Yes ,I'm also getting the same error and jupyter notebooks kernel is dying
@@saikiranreddy01123 u have to run all new commands which is updated at tensorflow site while runing all commands step by step u will face no error 😊
@@johncraven3560 I tried it , the kernel still keeo dying.
Does WSL2 method made the TF processing slower compared to Native Installation method?
Often slow... no pinned memory, however, TF no longer supports native, so you would be restricted to an older version of TF.
@@HeatonResearch ok, thanks Professor. ❤️ I tried TF v 2.10 since it is said it's the last version with native support, but every method to install that version failed. 😭
Thank you for this tutorial, I was able to install tensorflow!!
But now I have problems with TensorRT, can you explain how to install it? I can't run my scripts in WSL it give me a lot of errores that I haven't on native windows
I solved my problems copying the "nvvm" folder from cuda to the floder of the script, not the cleanest solution, but it works
i also got the same issue, and got irritated after spending hours on it and I uninstalled everything..heck
@@bharathkonatham6043 I abandoned the idea of TensorRT, I found that TensorFlow needs to be builded with TensorRT to be able to use TensorRT, but TensorRT it is for real time processing as video analysis. Since I do not need it, I abandoned the idea.
But I was able to run my scripts as I descrived in my other comment, copying the "nvvm" folder to the folder of the script
I ask to ChatGPT4 and it helpme a lot
Which is more preferrable to run TF/PyTorch, Windows 11 or WSL2 on Windows11? Thanks.
I would stay in native Windows if I can. WSL2 does not support memory pinning, which is used in many models for a performance boost.
@@HeatonResearch Thank you.
hey! Great video! I was able to install tensorflow successfully but I get these warnings when I run the tests :
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.
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.
what would need to be done to remove these warnings? Thanks!
Do i need to install a specific version of anaconda for this to work?
Not necessarily, unless it is something more than a few years old
Thank you for this useful video.
please help me to rebuild TensorFlow with the appropriate compiler flags in Windows 10 WSL2.
because of this information message:
I tensorflow/core/platform/cpu_feature_guard.cc:182] 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.
Anyone with Nvidia MX450?
Most people are on win10
I am going to curse a lot because ive been at it for a week and still do not have the fucking thing working.
WHY ARE THE NEW INSTRUCTIONS ON THE WEBSITE NERFED. There are no step by step instructions anymore, there is no instruction to install miniconda etc, everything is out of sequence please make a new tutorial on the latest version of tenrosflo cuda etc i cant run anything I am about to nuke the ubuntu i installed and start from the beginning fuck this
yes did it get work for u? i also need solution
@@hareram4233 yeah I got it to work eventually. You need to follow the table of compatibility that is on tensorflow website
@@hareram4233 yes just match the versions correctly
Worked for me by using 2.15 version...there is some issue with 2.16 right now...hope for others this comment helps... In video he used 2.11 u can use 2.15... and also follow miniconda steps...even tho official website anymore doesn't have miniconda steps
I fuckinghate conda. Can you do it w/o it?
Jeff, thank you for your videos. I've been using TF on windows for about 2yrs now. I followed your directions for setting up TF11 using WSL2, and everything seems to be working correctly on my system. I get the same output as you get at 9:11 in your video. My system can see the GPU and my models seem to be running with the same times as under Windows. However even from within python (on WSL2 console), if I type: os.system('nvidia-smi') it's still showing the CUDA 12.1 and not the target 11.2 . Is this expected? How can I 'know' that in fact CUDA 11.2 is really being used? Output below:
Your kernel may have been built without NUMA support.
Num GPUs Available: 1
>>> os.system('nvidia-smi')
Sun Mar 12 13:19:32 2023
+---------------------------------------------------------------------------------------+
| NVIDIA-SMI 530.30.02 Driver Version: 531.18 CUDA Version: 12.1 |
|-----------------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+======================+======================|
| 0 NVIDIA GeForce RTX 3090 On | 00000000:01:00.0 On | N/A |
| 0% 45C P8 39W / 390W| 1480MiB / 24576MiB | 2% Default |
| | | N/A |
+-----------------------------------------+----------------------+----------------------+
I may be wrong but from what I've seen, the CUDA version it displays when you run "nvidia-smi" will almost always be what you have installed on your host machine(windows). Your models all use the CUDA version that you installed inside your virtual environment but your linux distro itself is still just using the version that you have on windows(which is how its supposed to be). Honestly tho the fact that it works for you at all means you're all set due to how incredible sensitive tensorflow is with CUDA versions.
check cuda version with : nvcc --version
hello ,I have done exactly like in video and it worked but when used tf.reshape the kernel is dying . It happpens every single type , the terminal log is
Your kernel may have been built without NUMA support.
2023-05-21 15:28:23.030317: I tensorflow/compiler/xla/stream_executor/cuda/cuda_gpu_executor.cc:982] 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.
2023-05-21 15:28:23.030387: I tensorflow/compiler/xla/stream_executor/cuda/cuda_gpu_executor.cc:982] 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.
2023-05-21 15:28:23.032833: I tensorflow/compiler/xla/stream_executor/cuda/cuda_gpu_executor.cc:982] 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.
2023-05-21 15:28:23.032884: I tensorflow/compiler/xla/stream_executor/cuda/cuda_gpu_executor.cc:982] 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.
2023-05-21 15:28:23.032913: I tensorflow/compiler/xla/stream_executor/cuda/cuda_gpu_executor.cc:982] 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.
2023-05-21 15:28:25.942089: I tensorflow/compiler/xla/stream_executor/cuda/cuda_gpu_executor.cc:982] 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.
2023-05-21 15:28:25.947470: I tensorflow/compiler/xla/stream_executor/cuda/cuda_gpu_executor.cc:982] 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.
2023-05-21 15:28:25.947486: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1722] Could not identify NUMA node of platform GPU id 0, defaulting to 0. Your kernel may not have been built with NUMA support.
2023-05-21 15:28:25.948739: I tensorflow/compiler/xla/stream_executor/cuda/cuda_gpu_executor.cc:982] 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
I'm a student and didn't know much about these things . please some one help me