How to faff around with pytorch + cuda on a Precision M6700?
-
Hello @all,
.plan: I'm trying to set up a Dell Precision M6700 / GPU: Nvidia Quadro K4000M in a way which allows me to give pytorch (easyDiffusion, the odd LM from huggingface etc) a whirl.
but.. I find I'm having quite a hard time to find the ..sweet spot.. re OS <version> + Nvidia-Driver <version> + nvdia-toolkit <version> (=nvcc & whateverthefuck) + pytorch <version> [+ dependencies].
Either pytorch complains 'cos <driver> is too old or <something> complains because <driver> doesn't match nvcc or python is too old or (worst case) the recommended driver causes X to crash.
Windows <?Version?> I haven't tried yet but I'm not that optimistic.
What does
say about this endeavour?[ ] scrap the Dell, it's too old for this
[ ] try a -> b -> c
[ ] buy a Jetson, it's the same shite/hassle but one can spend some bucks
[ ] ..?..Edit#1: since the Linux-install is b0rked anyways and installing Windows afterwardws will fuck up UEFI/Boot for sure I decided to pave the box and restarted with W10 Pro.
Base Install (+ powershell: Get-AppxPackage | Remove-Appxpackage to get rid of the cruft, nothing else)
-> Windows reports "Nvidia Driver 382" (Add/Remove Programs)
-> compmgmt.msc -> Devices -> ... -> Nvidia K4000M shows: 22.21.14.8216
How does one find the "correct" CUDA-Package for this?
.. is it 10.2, 10.1, <latest> (which I know that it doesn't work)..
hmm
-
@iKnowItsLame said in How to faff around with pytorch + cuda on a Precision M6700?:
How does one find the "correct" CUDA-Package for this?
I've never directly attempted to use CUDA directly. But on my machine I did the Easy Diffusion thing and it has mostly been user error after that part.
Trying to hand-craft the playbook thingies they were talking about never got me close to being off the ground, but this (mostly) Just Worked.
-
@Tsaukpaetra
yeah... me2.Difficulty: this Dell machine is quite old and ED can't do much about <fucking Nvidia barfs around all over the place>
NB: had a HP Workstation Z or something with a "not that old GPU" a few days ago to play around with it: "just worked" but sadly this box got reclaimed for some reasons (need the memory-sticks mumblemumble somesuch) and ... :-(
Edit: oh.. OH. ED on Windows just might do something more than git <whatever> + python & pip whatever...
The 10.2 Cuda-Toolkit mentioned on <top> definitely doesn't work with <OOB-Window-Drivers>; since some googling found me the odd hint about "use 10.1 'cos ..." I removed that shite, removed <Windows-Drivers> and dared meself to run cuda_10.1.105_win10_network.exe
/will refrain to post URLs for now 'cos the Interweb is already full of "this & that & ..nothing.. & 2 years later: 'solved it, kthxbye!'" and I always wanted to do this too
-
CUDA 10.2: Installed successfully but doesn't work (the .exe provided all complain about ..no device found.. or somesuch)
CUDA 10.1: Install cancels with "..some error" but of course doesn't show which component caused this.
Can we please have a survey about "Is NV even worse than Oracle 8?"
-
Update 2023-12-22: "Windows 10 with cuda_10.1.105_win10_network.exe" worked
-> used the "quick install"-OptionSamples seem to work,
ex: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.1\extras\demo_suite
dir *.exe
-> all exe except "busgrind.exe" report "pass or show some output"
--> busgrind.exe complains re VCOMP110.DLL not foundlooks good so far, ED (2023-12-22) next.
-
I've never directly attempted to use CUDA directly. But on my machine I did the Easy Diffusion thing and it has mostly been user error after that part.
Having downloaded <link> this raises the question:
do you do actually click "of course, install this shite!" on every popup Windows decides to pop?/just asking 'cos the warnings my fresh installed W10 somewhat scared me (Edge -> do you really?... -> explorer: do you really ...)
//of course I pushed "YES, hop along!"
-
ofcourse ED@Windows does this before even attempting to check if ...
/3.1MB/s -> Powerline 'cos the cave here ain't permitted to get more because of
:"Very good, Sir! Is there anything else I can fuck you up with?"
-
@iKnowItsLame
... doesn't recognize <whatever_is_needed_to_do_the_needful>.
-> running on CPI./today me cannae be arsed to investigate this further; i'm gonna to postpone this after <xmas-festivities>
-
@iKnowItsLame said in How to faff around with pytorch + cuda on a Precision M6700?:
do you do actually click "of course, install this shite!" on every popup Windows decides to pop?
What, you haven't been trained well enough by the Yes Man yet?
@iKnowItsLame said in How to faff around with pytorch + cuda on a Precision M6700?:
does this before even attempting to check
It is annoying you have to download it but there aren't many better starting models that are smaller. Luckily only need to do it once, unless you lose it somehow.
@iKnowItsLame said in How to faff around with pytorch + cuda on a Precision M6700?:
... doesn't recognize <whatever_is_needed_to_do_the_needful>.
-> running on CPI.Ah... Yeah I think you're going to be up the creak based on the collected posts you've made so far.
According to my research, you should be using version 474.64 of the Nvidia drivers and CUDA Toolkit 11.4
It appears that driver support for that card has basically ended, so that should be effectively it (no "They fixed it in later versions" or "The newer versions don't work"), the compatibility matrix is locked, so if this don't work it probably won't ever.
If you wanted to try Linux again, you'll need version 470.223.03 of the drivers and 11.4 of the toolkit (same as before).
-
@Tsaukpaetra
sorry for the delay, the aforementioned "festivities" (haha but this is a completely different matter, let's just say that dementia -FatherInLaw, not me ..yet..- ain't fun...)Anyhoo: thankyou for your research, i absolutely will give this a whirl or five, Win & Lin both.
All the best for 2024 & THX again
PS: of course I do trust people from
more than some rando popups from Edge, why do you ask?
-
Quick update:
Windows:
474.64-quadro-rtx-desktop-notebook-win10-win11-64bit-international-dch-whql.exe from @Tsaukpaetra 's link above refuses with...
Basically it says:
-) This NVIDIA driver isn't compatible with this version of Windows.
-) The graphics-driver could not find compatible hardware
(Q: which one is it now? but let's not digress...)Next: Reinstall Linux (e.g. download Ubuntu 20.04 since this is the latest one supported from @Tsaukpaetra 's links above. Let's not fuck this up/overcomplicate with Linux Mint <blech> is Ubuntu <blah> since NV does quite the job here already...)
.. this will take a minute or ten .., schaumermal, dann sehnwirschon.
-
@iKnowItsLame said in How to faff around with pytorch + cuda on a Precision M6700?:
let's just say that dementia
Been there, got the T-Shirt, probably going to have a note about that on my conditional DNR...
-
@iKnowItsLame said in How to faff around with pytorch + cuda on a Precision M6700?:
This NVIDIA driver isn't compatible with this version of Windows.
Yeah, that's usually Nvidia fucking something up (unless you grabbed 32-bit Windows on accident?). Usually I'd nuke-and-pave with DDU before trying to switch to different drivers.
Let's hope your attempt on Linux goes better. Probably going to have a slightly better chance with it getting the Legacy tag anyways...
-
Quick update, just to close this one.
Summary: I give up.
Windows: see above postings
Linux: same, just more difficultiesI'm quite sure one can overcome this shite with the "correct" incantation of dpkg | apt | whateverthefuck to get a working combo of [driver + tools + cuda-toolkit] but then you're still stuck with an outdated version of torch.
@Tsaukpaetra thankyou!
-
@iKnowItsLame said in How to faff around with pytorch + cuda on a Precision M6700?:
Summary: I give up.
It was a valiant attempt! 🌷
