Jeff you have been a lifesaver this semester. I'm taking my first Computer Vision course, and for quite literally every single question I had or roadblock I faced, you had a video that explained it perfectly. This is just the latest of many. I had issues with my Windows/Ubuntu dual boot, and it was a hassle to switch between the two when working (deliverables required code and typed documents); this, while imperfect for now until it gets updated more, is the perfect solution that both allows me to work in the comfort of Windows and run tf code that uses my GPU. Thanks a million!
Professor Heaton, Thank you for the great video! I I believe in minute 23 of the video where you run the jupyter notebook the error can be solved using the following command: jupyter --no-browser instead of just using jupyter.
Jeff, thank you for going through all of the pain for us. Due to some constraints with my application stack, I am stuck with a Windows base OS but needed a setup like this to drive the deep learning hardware with Tensorflow 2.2. Right on the money! Thanks again!
Thank you very much. For work I have to bounce between Windows for corp apps, and Linux for network management. This gives me the ability to do both from the same laptop.
Excellent video, Jeff! I finally had time to get around to setting up WSL 2 and miniconda. I can't wait to play around with the workflow and compare it to the full Windows setup I had been using. Thanks again!
Great video. Also 0:52 they don't put Nvidia on Macs. This is why I've never seriously consider a Mac as an ML engineer for training models locally. Unless you do all your model training in the cloud (and there's a few barriers to going that route vs just running locally). Thanks Jeff. Also, the beard is looking good on you!
Wow!! I don't even know WSL existed. U the best. I used to dual boot back and forth between Linux and Windows. I will definitely try this. Great video.
Finally! I was trying since two days by going through the process of installing cuda toolkit on the nvidia page but your method of doing it through conda worked like a charm!.
@@abhinand5 yeah that problem solved but now I have a new problem that occurs when I try to run any model on the gpu using Cuda where it says no Cuda driver available even when my gpu is not running.
Thanks Jeff, really helpful, will check out rest of the stuff in the repo. Not sure if you fixed this later but your memory in the BIOS was only clocked at 2133mhz. The XMP profile for RAM kits in a high-end workstation probably clocks to at least 3600 mhz which probably increases speeds by a double-digit percentage. You might have to update the BIOS and or do set the RAM timings.
Thanks for the video, though I've read in an article which concluded with few benchmarks that using gpus (cuda cores) in WSL2 is almost 80% slower than using them in the normal windows... so it would be great if you could do some benchmarks for that by running some basic scripts and comparing the model's performance in wsl2 and outside of it. Thanks for your time.
Yes, it is slower. I plan to do some benchmarking soon, once I setup for dual boot. I also find WSL a little bit brittle, about a month into setting it up. Updates tend to break it and force me to reinstall.
@@HeatonResearch Yeah same scenario in my case too. One positive side is that wsl2 is very faster as compared to wsl1 (at least about 10x in most of my comparisons), but at the same time is not even close to linux installed locally... :{
Great video! Getting this setup still seems kinda crazy. I'm gonna wait for Windows and WSL2 to settle down before trying this. Looking forward to your videos. Thanks for all the hard work!
I have GeForce RTX 2080 GPU which has Tensor Cores. Tensorflow doesn't utilise TensorRT library which takes advantage of Tensor Cores on Windows, but it does utilise them on Linux. It is interesting to see if the new WSL would allow Ubuntu to use TensorRT through Windows.
Would I create a docker container any differently if I was using WSL2 (I dont need to install docker within the wsl2 environment right?) or native linux?
Don't get me wrong, the WSL is capable. But it's really just a powergrab from Microsoft. They just want to extinguish competition. And that means Linux in this case. DirectX on WSL in proof for that. They just try to get a foothold in the traditionally Linux-dominated market. They don't like Linux, or Open-source. They try to copy the Unix ecosystem of compilers, userspace, etc. while intentionally making it incompatible with Linux and not contributing back. But TBH, I never really got the Windows-GUI is so awesome. I just set up a new Windows 10 install, and the first thing I noticed was the start menu doesn't calculate positions for the tiles correctly. WTF. Yes, it had all the updates installed. Besides, most of what I do with the GUI is moving browser, text editors/IDE's and command-line windows around. You've used the GUI in this video only for a browser and a terminal. You can do that with and GUI, Windows, OS X and Linux. And to be honest, windows lacks behind. I don't see a proper tiling window manager for windows, especially one that supports keyboard navigation. But to each his own. What kind of hypervisor *needs* SVM to be *disabled*? I've never heard of that. Is that a VMware thing? Anyway, still a great guide. I'm sure it could come in handy.
One more note, I was wondering why we dont have to install CUDA tool kit and CudaNn within the ubuntu - I am not well versed in Linux so I am not sure if this is reduntant. Thanks again
Jeff, you have some of the best ML videos on RUclips. Would you have installed windows on your new box if this was box was dedicated to machine learning? It would be interesting to see metrics on this box from Linux only compared to Windows WSL
Hi Jeff, a timely series. So now I’m wondering after running WSL 2 with PyTorch for a bit have you hit any gotchas that make it either unusable for certain tasks or even have quirks that you have to work around? I ask as I’m about to setup a GPU enabled Linux compatible laptop as dual boot but if I can get away with WSL 2 it will simplify life a bunch. I mean I’m not complaining about having to switch in and out of operating systems since I will live mostly in Linux but there just certain things Linux still barfs on that would be mitigated if the base OS was windows with WSL 2. Any thoughts would be most appreciated and thank you for the ongoing videos. :)
Excellent video Jeff. A question on performance, do you see worse performance for your GPU runs in WSL? I have an Nvidia 2070 Super but switching to GPU I see epochs going from 3-5 seconds to 17-19 secs. just curious if you saw the same performance loss
thx very cool. do you plan to make a followup video when this is out of beta? I would like to run it this way but I can't just install some windows on my production system.. ;)
Well, everything is cool, you configured everything and it's running but as far as I know GPU performance using WSL is poor for now. It would make me crazy to buy Titan RTX and train models slowly... Will see when it gets out of beta. Are you planning to make some tests for your PC running he same code on native Linux and Windows + WSL 2? That would be interesting. Also did you consider installing Ubuntu Server 20 + VSCode? (for more info google for Remote - SSH plugin). This way you can connect to your PC from anywhere using your MacBook to combine polished UI of macOS and computational power of your new PC running native Linux. Also in this case you could save money by not buying monitor)))
Thank you for the guidance, Mr. Heaton! But unfortunately, I am facing an issue with "wsl cat /proc/version" and "wsl.exe - - update" commands as they show two different kernel versions which are 4.4.0 and 5.4.72 respectively. I am not sure about the actual kernel version. If you could clarify for me here? Thanks in advance!!
Mate, have been using Ubuntu since 16.04 ( current 20.04) as a desktop without any problems, emails, office, spreadsheets, zoom, and of course deep with Nvidia, never has a problem. Never had a problem, while WSL might be there, Linux desktops are heaps better, as they say, the original is original.
I don't doubt you. I am considering Ubuntu on my next laptop, I mostly exist in both Windows/Linux. If I had to pick a platform I am thinking less of lately, it would be Mac. (as I type this on a Mac, lol)
@@HeatonResearch Hello Jeff I have known about your work since Encog, as a matter of fact, the first neural network I wrote was in Encog C#. 8 Years later ( Data Scientist). Don't you think Keras layers are very much inspired by Encog?
Thank you for this informative video- I actually installed wsl2 on my new windows 10 workstation, and all going well with one exception though- I was running the 04-02 multiclass notebook file - and encountered some issues that may be CUDA related. I have cuda running fine on windows 10 (version 10.1 now , with tensorflow 2.1.0) . THe section for building the NN model produced an internal error: InternalError: Blas GEMM launch failed : a.shape=(32, 30), b.shape=(30, 100), m=32, n=100, k=30 [[node sequential/dense/MatMul (defined at :25) ]] [Op:__inference_train_function_824]. From the ubunto screen see the cuda dll were successfully loaded- but had this failure : ...failed to create cublas handle: CUBLAS_STATUS_ALLOC_FAILED I was wondering if you had something similar to this case and what was the remedy for it. Thank you very much for your work here
Question Jeff: When cloning the repos from your jeffheaton repo into the WSL Ubuntu environment, which directory should we be cloning your repo into? Something like /home/user_name/repos/jeffheaton ? Trying to maintain some consistency for yours and future github repos
I typically put them into a folder called "projects" right off of my ~ home directory. However, for videos I sometimes clone right to my home directory for fewer steps.
Hey Jeff, trying to understand what cudatoolkit 11 provides for WSL. Since we're using `conda install pytorch cudatoolkit -c pytorch`, does that mean that our WSL 2 system's cudatoolkit 11 isn't used? Instead whatever version conda installs is used?
Hi. Do someone know how to stop the window insider ? It is bringing me some problems and I want to leave the insider program, but I don't know how to proceed. Thanks
Hello, do you have to install the Nvidia driver before install WSL2 to make it work? I did the opposite and now in WSL, It states that it can't find the driver.
Did everything in the video. But when Im trying to declare aa model using .to(device) it gives me the error RuntimeError: CUDA error: all CUDA-capable devices are busy or unavailable
I'm confused on why you didn't have to install cuDNN SDK? I have my setup working but I'm just curious why you didn't have to I had my WSL2 setup with CUDA support (and could run the CUDA Samples) but tensorFlow would not work until I installed libcuDNN. I haven't followed your tutorial but it is identical up until the mini conda part (I like to use python3-venv instead).
Basically I am letting CONDA handle the drivers for me. It works the same way outside of WSL-2 as well. You can have drivers loaded at the system level, but then it becomes difficult to manage two different projects that need different CUDA/CUDNN versions.
@@HeatonResearch Did not know conda can do that too! That's a genius way to handle that and I'll have to try it out now Thanks for the reply Professor :)
I'm trying this on WSL 1. Is there any compelling reason for preferring WSL 2 other than the ones already stated in docs.microsoft.com/en-us/windows/wsl/compare-versions?
no dislikes... you just "saw the light"... as much as I dislike M$... and I've been through all other GUI(s) in Ux/Lx/Mac/... the Windows Shell is the more consistent/coherent. and now with Powershell Core... things are becoming more controllable... As for the MacOS CLI and GUI... it's almost the worst I ever found... (literally 'Shitty') I really don't understand some opinions... almost every Linux GUI is better then MacOS GUI
hello jeff thanks for the video i followed your instructions and got a ubuntu terminal but when i try to install miniconda i get this error ak@LAPTOP-U5UBCOF5:~$ wget repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh --2020-09-15 21:08:13-- repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh Resolving repo.anaconda.com (repo.anaconda.com)... failed: Temporary failure in name resolution. wget: unable to resolve host address ‘repo.anaconda.com’ i have waited for a day and still getting the same error. can you provide the solution
He did it correctly. I think he made a mistake when he edit the video. He typed "wsl.exe --update" and the results should have been like this "Checking for updates... No updates are available. Kernel version: 4.19.121.1" It's hard to edit a video which takes more time and energy than just recording the video.
Partly I agree with you, partially I don't. I really think Ubuntu and Manjaro are trying hard to make that experience intuitive to new users. While windows tho its committed towards linux (which i highly doubt) are making users comfortable to use clear instead of cls. 😂
Good point, really, for various reasons I need to be in Win, Mac, & Linux. Linux desktop so far has not impressed me, I am mostly command line. But some are looking pretty good, such as Mint Linux.
day are gone when microsoft said linux is cancer...now microsoft is embracing linux and they will bring all features of linux on window..including gui of linux
Hello Professor Heaton, Thank you for the great video! I followed your steps but at the end, it shows GPU is not available. I do have a GeForce GPU and it did work when I followed your other video on installing tensorflow on Windows. What do you think might happened? Thanks!
It turns out even I type "wsl --set-default-version 2", my wsl version is version 1. I found a command "wsl --set-version Ubuntu-20.04 2 " from David Bombal to manually set wsl to version 2. Now, it shows GPU is available!
@@yuchenxiao1163 Hi Yuchen, I have exactly the same problem as you. I confirmed running WLS 2 and followed each and every step in the video as explained to set up Tensorflow with GPU, but in the end my GPU is evaluated to be not available. It was quite a hassle to install the the Windows insider updates and my laptop crashed several times while updating, but, indeed, after 2 days work, Tensorflow 2.2 is installed successfully while showing GPU is not available. Exactly like you, when installing it in Windows itself, GPU is classified to be available. I suspect it has something to do with installing/setting up CUDA on Linux using WSL2. Does anyone has an idea? My Linux kernel is: version 5.4.72-microsoft-standard-WSL2 Edit: Something is wrong the communication between the communication of my graphics card and Ubuntu 20.04. When installing sudo apt install nvidia-utils-440 and running "nvidia-smi" in the terminal to obtain information about my graphics card, I yield: Failed to initialize NVML: GPU access blocked by the operating system Failed to properly shut down NVML: GPU access blocked by the operating system
@@hbbexxter4666 Please have you been able to fix this problem? Having too much trouble with wsl2 set-up these days. After the latest update from the insider program, I couldn't access GPU from wsl2 anymore. It's so disturbing.
Hey again Jeff i know you have 64Gb of 3200Mhz Ram in your system but the ram is only running at 2133Mhz by default!! you need to set the speed to 3200Mhz in the bios. AMD in a nutshell.
Hi, windows Developer OK ,WSL2 OK, nvidia insider driver install OK, docker ok ... and no cuda available in my ubuntu ... I use a regular windows 10 pro version 20h2 19042.630 i think i am screew because i must have a version >= to 2000 only available with windows insider version ? edit : PS1: Ok I have found my PB, insider program is only available to administrator acount not to normal user under w10 this is why i am stuck in 19042.630 and not be able to deploy windows with a build >20145 as require by nvidia wsl user guide docs.nvidia.com/cuda/wsl-user-guide/index.html PS2 : As an administrator my "insider program" show only a blank page ... i have use this trick to be able to activate and join a dev ring : www.howto-connect.com/fix-insider-program-settings-blank-empty-missing-windows-10/ I hope this will help others !
The only thing i am waiting for is AMD to release their version of cuda. And make it open source. AMD has been underdog for long now its time for them to show who the bigdawg.
Thank you so much. I have learned a lot from you over the past couple of months. I have a question. Whenever i run anything that has to do with conda, such as " conda install -y jupyter" or "conda update conda" I get this error : Collecting package metadata (current_repodata.json): failed CondaHTTPError: HTTP 000 CONNECTION FAILED for url Elapsed: - An HTTP error occurred when trying to retrieve this URL. HTTP errors are often intermittent, and a simple retry will get you on your way. If your current network has www.anaconda.com blocked, please file a support request with your network engineering team. 'repo.anaconda.com/pkgs/main/linux-64' I am using my work computer. Do you have any idea what is causing the problem?
Might be an issue with the SSL verification. More info here: stackoverflow.com/questions/50125472/issues-with-installing-python-libraries-on-windows-condahttperror-http-000-co
Jeff you have been a lifesaver this semester. I'm taking my first Computer Vision course, and for quite literally every single question I had or roadblock I faced, you had a video that explained it perfectly. This is just the latest of many. I had issues with my Windows/Ubuntu dual boot, and it was a hassle to switch between the two when working (deliverables required code and typed documents); this, while imperfect for now until it gets updated more, is the perfect solution that both allows me to work in the comfort of Windows and run tf code that uses my GPU. Thanks a million!
Professor Heaton, Thank you for the great video! I
I believe in minute 23 of the video where you run the jupyter notebook the error can be solved using the following command:
jupyter --no-browser instead of just using jupyter.
It took everything in me I could to trust you after seeing you were using Chrome and Edge at the same time
Jeff, thank you for going through all of the pain for us. Due to some constraints with my application stack, I am stuck with a Windows base OS but needed a setup like this to drive the deep learning hardware with Tensorflow 2.2. Right on the money! Thanks again!
Thank you very much. For work I have to bounce between Windows for corp apps, and Linux for network management. This gives me the ability to do both from the same laptop.
Yes I agree, it gives me a similar advantage.
Very helpful. Probably a better approach than traditional VM. Thanks.
Excellent video, Jeff! I finally had time to get around to setting up WSL 2 and miniconda. I can't wait to play around with the workflow and compare it to the full Windows setup I had been using. Thanks again!
Glad to help! I am liking this technology to use the latest TensorFlow 2.3 on Windows currently. Can't believe CoLab already updated.
Great video. Also 0:52 they don't put Nvidia on Macs. This is why I've never seriously consider a Mac as an ML engineer for training models locally. Unless you do all your model training in the cloud (and there's a few barriers to going that route vs just running locally).
Thanks Jeff. Also, the beard is looking good on you!
I just feel like lately Mac is making it more and more difficult for me to like them.
Nice, I wasn't aware that WSL had been updated to allow GPU usage. Definitely installing this when I get the new 3090 PC. Thanks Jeff.
Wow!! I don't even know WSL existed. U the best. I used to dual boot back and forth between Linux and Windows. I will definitely try this. Great video.
Its something I've been following, but until I saw the announcement about the NVIDIA GPU support, I did not bother.
Finally! I was trying since two days by going through the process of installing cuda toolkit on the nvidia page but your method of doing it through conda worked like a charm!.
It stopped working after windows got updates ;-;
@@phsyco202 yeah and for I had to reset my PC
@@abhinand5 yeah that problem solved but now I have a new problem that occurs when I try to run any model on the gpu using Cuda where it says no Cuda driver available even when my gpu is not running.
@@phsyco202 Better you remove WSL altogether and follow Jeff's DL setup in Windows. I've been using it without any problems for more than a month now.
@@abhinand5 Can you send me the link
Thanks Jeff, really helpful, will check out rest of the stuff in the repo. Not sure if you fixed this later but your memory in the BIOS was only clocked at 2133mhz. The XMP profile for RAM kits in a high-end workstation probably clocks to at least 3600 mhz which probably increases speeds by a double-digit percentage. You might have to update the BIOS and or do set the RAM timings.
Great video Jeff, keep the great vid's coming man. Love it
Thanks for the video, though I've read in an article which concluded with few benchmarks that using gpus (cuda cores) in WSL2 is almost 80% slower than using them in the normal windows... so it would be great if you could do some benchmarks for that by running some basic scripts and comparing the model's performance in wsl2 and outside of it.
Thanks for your time.
Yes, it is slower. I plan to do some benchmarking soon, once I setup for dual boot. I also find WSL a little bit brittle, about a month into setting it up. Updates tend to break it and force me to reinstall.
@@HeatonResearch Yeah same scenario in my case too. One positive side is that wsl2 is very faster as compared to wsl1 (at least about 10x in most of my comparisons), but at the same time is not even close to linux installed locally... :{
Great video! Getting this setup still seems kinda crazy. I'm gonna wait for Windows and WSL2 to settle down before trying this. Looking forward to your videos. Thanks for all the hard work!
I know, once WSL2 settles down I may quit the Windows insider program. Normally I do not run early release operating systems.
Hi Jeff, great video and instructions. I will try. I have a GPU Nvidia GeForce GTX 1660 Ti.
Really appreciate this tutorial Jeff. Subbed!
I have GeForce RTX 2080 GPU which has Tensor Cores. Tensorflow doesn't utilise TensorRT library which takes advantage of Tensor Cores on Windows, but it does utilise them on Linux. It is interesting to see if the new WSL would allow Ubuntu to use TensorRT through Windows.
Thank you for the great detail. Indeed, you are amazing.
You are so welcome
Thank you so much but for me Linux forever !!!
Would I create a docker container any differently if I was using WSL2 (I dont need to install docker within the wsl2 environment right?) or native linux?
Don't get me wrong, the WSL is capable. But it's really just a powergrab from Microsoft. They just want to extinguish competition. And that means Linux in this case. DirectX on WSL in proof for that. They just try to get a foothold in the traditionally Linux-dominated market. They don't like Linux, or Open-source. They try to copy the Unix ecosystem of compilers, userspace, etc. while intentionally making it incompatible with Linux and not contributing back.
But TBH, I never really got the Windows-GUI is so awesome. I just set up a new Windows 10 install, and the first thing I noticed was the start menu doesn't calculate positions for the tiles correctly. WTF. Yes, it had all the updates installed.
Besides, most of what I do with the GUI is moving browser, text editors/IDE's and command-line windows around.
You've used the GUI in this video only for a browser and a terminal.
You can do that with and GUI, Windows, OS X and Linux. And to be honest, windows lacks behind. I don't see a proper tiling window manager for windows, especially one that supports keyboard navigation.
But to each his own.
What kind of hypervisor *needs* SVM to be *disabled*? I've never heard of that. Is that a VMware thing?
Anyway, still a great guide. I'm sure it could come in handy.
This was so so cool! You saved me
One more note, I was wondering why we dont have to install CUDA tool kit and CudaNn within the ubuntu - I am not well versed in Linux so I am not sure if this is reduntant. Thanks again
I am also curious of this
Jeff, you have some of the best ML videos on RUclips. Would you have installed windows on your new box if this was box was dedicated to machine learning? It would be interesting to see metrics on this box from Linux only compared to Windows WSL
Getting ready to do just that... going to be adding a new M.2 for dual boot soon.
You are amazing! Tried other methods and nothing worked. Thank you!
Glad it helped!
Hi Jeff, a timely series. So now I’m wondering after running WSL 2 with PyTorch for a bit have you hit any gotchas that make it either unusable for certain tasks or even have quirks that you have to work around? I ask as I’m about to setup a GPU enabled Linux compatible laptop as dual boot but if I can get away with WSL 2 it will simplify life a bunch. I mean I’m not complaining about having to switch in and out of operating systems since I will live mostly in Linux but there just certain things Linux still barfs on that would be mitigated if the base OS was windows with WSL 2.
Any thoughts would be most appreciated and thank you for the ongoing videos. :)
Excellent video Jeff. A question on performance, do you see worse performance for your GPU runs in WSL? I have an Nvidia 2070 Super but switching to GPU I see epochs going from 3-5 seconds to 17-19 secs. just curious if you saw the same performance loss
I've seen similar. I've since switched to dual boot Linux. I am going to do a video just about that.
thx very cool. do you plan to make a followup video when this is out of beta? I would like to run it this way but I can't just install some windows on my production system.. ;)
Yes, that's the plan!
Latest windows public update 2004 (Windows Home -- release channel) allows WSL and WSL2 without being in the Insider Program. Released July 2020
True, but NVIDIA does still require insider program for their driver to work.
Thanks so much for lettting us know easy installation process, I request you to let us know OpenCV installation in WSL for both C++ and Python
pip install opencv-python
In terminal for python
Great topic.
Thanks!
Thanks a lot Jeff!!
Jeff... You're a blessing !!!
Between, i love the new look !!!
Hah, thanks
Thanks for sharing
I am with OpenBSD🎉
Well, everything is cool, you configured everything and it's running but as far as I know GPU performance using WSL is poor for now. It would make me crazy to buy Titan RTX and train models slowly... Will see when it gets out of beta. Are you planning to make some tests for your PC running he same code on native Linux and Windows + WSL 2? That would be interesting. Also did you consider installing Ubuntu Server 20 + VSCode? (for more info google for Remote - SSH plugin). This way you can connect to your PC from anywhere using your MacBook to combine polished UI of macOS and computational power of your new PC running native Linux. Also in this case you could save money by not buying monitor)))
I am going to benchmark it further. For smaller jobs WSL is not that good as ramp up/down takes longer than the actual epoch.
your videos are allways great
thank you for your video, I can't see the Dev channel, I have Windows education ? is that a problem?
great introduction
Very useful. Thank you.
Thank you for the guidance, Mr. Heaton!
But unfortunately, I am facing an issue with "wsl cat /proc/version" and "wsl.exe - - update" commands as they show two different kernel versions which are 4.4.0 and 5.4.72 respectively. I am not sure about the actual kernel version. If you could clarify for me here? Thanks in advance!!
After running the jupyter notebook I get a blank page on selecting the kernel while creating a new ipynb file
Hi Jeff - Big fan of the channel! Would love to do a collaboration video at some point. Would you be interested?
Can you make a video on using libtorch with gpu support?
Mate, have been using Ubuntu since 16.04 ( current 20.04) as a desktop without any problems, emails, office, spreadsheets, zoom, and of course deep with Nvidia, never has a problem. Never had a problem, while WSL might be there, Linux desktops are heaps better, as they say, the original is original.
I don't doubt you. I am considering Ubuntu on my next laptop, I mostly exist in both Windows/Linux. If I had to pick a platform I am thinking less of lately, it would be Mac. (as I type this on a Mac, lol)
@@HeatonResearch Hello Jeff I have known about your work since Encog, as a matter of fact, the first neural network I wrote was in Encog C#. 8 Years later ( Data Scientist). Don't you think Keras layers are very much inspired by Encog?
all good with me till minute @15:46, you lost me there, i want to install nvidia drivers in wsl2 the normal way.
Thanks Jeff!!!
Thank you for this informative video- I actually installed wsl2 on my new windows 10 workstation, and all going well with one exception though- I was running the 04-02 multiclass notebook file - and encountered some issues that may be CUDA related. I have cuda running fine on windows 10 (version 10.1 now , with tensorflow 2.1.0) . THe section for building the NN model produced an internal error:
InternalError: Blas GEMM launch failed : a.shape=(32, 30), b.shape=(30, 100), m=32, n=100, k=30
[[node sequential/dense/MatMul (defined at :25) ]] [Op:__inference_train_function_824].
From the ubunto screen see the cuda dll were successfully loaded- but had this failure :
...failed to create cublas handle: CUBLAS_STATUS_ALLOC_FAILED
I was wondering if you had something similar to this case and what was the remedy for it.
Thank you very much for your work here
how to change install path to a more large disk instead of C disk
Question Jeff: When cloning the repos from your jeffheaton repo into the WSL Ubuntu environment, which directory should we be cloning your repo into?
Something like /home/user_name/repos/jeffheaton ?
Trying to maintain some consistency for yours and future github repos
I typically put them into a folder called "projects" right off of my ~ home directory. However, for videos I sometimes clone right to my home directory for fewer steps.
Can I run notebooks which are not in Linux directory but are in d or e drive?
What should I check if i got "gpu is not available"?
Hey Jeff, trying to understand what cudatoolkit 11 provides for WSL. Since we're using `conda install pytorch cudatoolkit -c pytorch`, does that mean that our WSL 2 system's cudatoolkit 11 isn't used? Instead whatever version conda installs is used?
Yes, the cudatoolkit is installed as part of the environment, this allows you to use different versions in different environments.
Hi. Do someone know how to stop the window insider ? It is bringing me some problems and I want to leave the insider program, but I don't know how to proceed. Thanks
so currently on ur github document it says google colab has tensflow 2.5, windows 2.3 and linux 2.2 is that correct?
Hello, do you have to install the Nvidia driver before install WSL2 to make it work? I did the opposite and now in WSL, It states that it can't find the driver.
Great video. Have you been able to set nvidia-docker on wsl2?
That is actually next on my list! Regular docker worked just fine so far.
You can check this link for docker + kubernetes. towardsdatascience.com/dual-boot-is-dead-windows-and-linux-are-now-one-27555902a128
Did everything in the video. But when Im trying to declare aa model using .to(device) it gives me the error RuntimeError: CUDA error: all CUDA-capable devices are busy or unavailable
did you install cuda on windows?
This might be the issue
@@rayg5573 I installed it the way he installed it opin the video. using conda
Partition and install Deepin and you will have a great experience.
I'm confused on why you didn't have to install cuDNN SDK? I have my setup working but I'm just curious why you didn't have to
I had my WSL2 setup with CUDA support (and could run the CUDA Samples) but tensorFlow would not work until I installed libcuDNN.
I haven't followed your tutorial but it is identical up until the mini conda part (I like to use python3-venv instead).
Basically I am letting CONDA handle the drivers for me. It works the same way outside of WSL-2 as well. You can have drivers loaded at the system level, but then it becomes difficult to manage two different projects that need different CUDA/CUDNN versions.
@@HeatonResearch Did not know conda can do that too! That's a genius way to handle that and I'll have to try it out now
Thanks for the reply Professor :)
I'm trying this on WSL 1. Is there any compelling reason for preferring WSL 2 other than the ones already stated in docs.microsoft.com/en-us/windows/wsl/compare-versions?
no dislikes... you just "saw the light"...
as much as I dislike M$...
and I've been through all other GUI(s) in Ux/Lx/Mac/...
the Windows Shell is the more consistent/coherent.
and now with Powershell Core... things are becoming more controllable...
As for the MacOS CLI and GUI...
it's almost the worst I ever found... (literally 'Shitty')
I really don't understand some opinions...
almost every Linux GUI is better then MacOS GUI
hello jeff thanks for the video i followed your instructions and got a ubuntu terminal but when i try to install miniconda i get this error
ak@LAPTOP-U5UBCOF5:~$ wget repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
--2020-09-15 21:08:13-- repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
Resolving repo.anaconda.com (repo.anaconda.com)... failed: Temporary failure in name resolution.
wget: unable to resolve host address ‘repo.anaconda.com’
i have waited for a day and still getting the same error. can you provide the solution
you look great in the beard.
Thanks!
14:57 when you re-run the command it said that you still have the old kernel version?
He did it correctly. I think he made a mistake when he edit the video.
He typed "wsl.exe --update" and the results should have been like this
"Checking for updates...
No updates are available.
Kernel version: 4.19.121.1"
It's hard to edit a video which takes more time and energy than just recording the video.
Partly I agree with you, partially I don't.
I really think Ubuntu and Manjaro are trying hard to make that experience intuitive to new users.
While windows tho its committed towards linux (which i highly doubt) are making users comfortable to use clear instead of cls. 😂
Good point, really, for various reasons I need to be in Win, Mac, & Linux. Linux desktop so far has not impressed me, I am mostly command line. But some are looking pretty good, such as Mint Linux.
day are gone when microsoft said linux is cancer...now microsoft is embracing linux and they will bring all features of linux on window..including gui of linux
Hello Professor Heaton, Thank you for the great video! I followed your steps but at the end, it shows GPU is not available. I do have a GeForce GPU and it did work when I followed your other video on installing tensorflow on Windows. What do you think might happened? Thanks!
It turns out even I type "wsl --set-default-version 2", my wsl version is version 1. I found a command "wsl --set-version Ubuntu-20.04 2 " from David Bombal to manually set wsl to version 2. Now, it shows GPU is available!
Cool, yes that is an important step! And unfortunately there are few error messages other than the GPU just not showing up.
@@yuchenxiao1163 Hi Yuchen,
I have exactly the same problem as you. I confirmed running WLS 2 and followed each and every step in the video as explained to set up Tensorflow with GPU, but in the end my GPU is evaluated to be not available.
It was quite a hassle to install the the Windows insider updates and my laptop crashed several times while updating, but, indeed, after 2 days work, Tensorflow 2.2 is installed successfully while showing GPU is not available. Exactly like you, when installing it in Windows itself, GPU is classified to be available. I suspect it has something to do with installing/setting up CUDA on Linux using WSL2.
Does anyone has an idea?
My Linux kernel is: version 5.4.72-microsoft-standard-WSL2
Edit: Something is wrong the communication between the communication of my graphics card and Ubuntu 20.04. When installing
sudo apt install nvidia-utils-440
and running "nvidia-smi" in the terminal to obtain information about my graphics card, I yield:
Failed to initialize NVML: GPU access blocked by the operating system
Failed to properly shut down NVML: GPU access blocked by the operating system
@@hbbexxter4666 Please have you been able to fix this problem? Having too much trouble with wsl2 set-up these days. After the latest update from the insider program, I couldn't access GPU from wsl2 anymore. It's so disturbing.
@@hbbexxter4666 i am facing the same problem
Hey again Jeff i know you have 64Gb of 3200Mhz Ram in your system but the ram is only running at 2133Mhz by default!! you need to set the speed to 3200Mhz in the bios. AMD in a nutshell.
Thanks!! Good eye on the BIOS screen. I thought I had set that previously, but I guess not or it reset somehow.
Hi, windows Developer OK ,WSL2 OK, nvidia insider driver install OK, docker ok ... and no cuda available in my ubuntu ...
I use a regular windows 10 pro version 20h2 19042.630 i think i am screew because i must have a version >= to 2000 only available with windows insider version ?
edit :
PS1: Ok I have found my PB, insider program is only available to administrator acount not to normal user under w10 this is why i am stuck in 19042.630 and not be able to deploy windows with a build >20145 as require by nvidia wsl user guide docs.nvidia.com/cuda/wsl-user-guide/index.html
PS2 : As an administrator my "insider program" show only a blank page ... i have use this trick to be able to activate and join a dev ring : www.howto-connect.com/fix-insider-program-settings-blank-empty-missing-windows-10/
I hope this will help others !
I don't see those 30 dislikes you mentioned
Ah, yes that is good!
The only thing i am waiting for is AMD to release their version of cuda. And make it open source.
AMD has been underdog for long now its time for them to show who the bigdawg.
There is potential there. I am a fan of AMD on the CPU side. But the fact that they are not on AWS also causes me to ignore their GPUs.
Thank you so much. I have learned a lot from you over the past couple of months. I have a question. Whenever i run anything that has to do with conda, such as " conda install -y jupyter" or "conda update conda" I get this error :
Collecting package metadata (current_repodata.json): failed
CondaHTTPError: HTTP 000 CONNECTION FAILED for url
Elapsed: -
An HTTP error occurred when trying to retrieve this URL.
HTTP errors are often intermittent, and a simple retry will get you on your way.
If your current network has www.anaconda.com blocked, please file
a support request with your network engineering team.
'repo.anaconda.com/pkgs/main/linux-64'
I am using my work computer. Do you have any idea what is causing the problem?
Might be an issue with the SSL verification. More info here: stackoverflow.com/questions/50125472/issues-with-installing-python-libraries-on-windows-condahttperror-http-000-co