Support for new NVIDIA RTX 50xx-Series GPUs

[sage quest]

I upgraded my graphics card to the newest NVIDIA RTX 5080 GPU. Invoke (or to be more specific PyTorch) seems to have some issues with it. Even GPU-Z does not seem to properly recognize the GPU...

[fathom sorrel]

hm, for this, i'd check the pytorch repo and see if they have a version (maybe a nightly?) that supports the 50x0 series. if so, you should be able to install it into your Invoke install (in dev console). hopefully, then it'll work

[sage quest]

I don't think that this is an option for me now, I don't feel confident and competent enough to do that...

[fathom sorrel]

fair enough ... to be honest, i'm not even sure it's possible right now. i just looked around and people are only getting it to work with some nvidia-released docker container or something... very much a band-aid solution right now

[surreal moss]

It might require 12.8? Or at least 12.6 - which would need 2.6.x for windows, and I don't think Invoke supports above 2.4.x on windows? Odds are various things need to be updated.

[fathom sorrel]

yep, it requires a whole new cuda version, but i don't even think 2.6.x supports it right now. there might be a nightly build of pytorch that does (3 days ago, there wasn't)

not sure what nvidia's docker image is doing differently, but that seems to work, but it's incomplete

[surreal moss]

There's an 'early access' nvidia pytorch wheel for blackwell.

[fathom sorrel]

ahh okay

[surreal moss]

But yeah at the end of the day... things aren't settled just yet. Needs to get into actual pytorch builds, and applications need support for it.

Looks like A1111 and it's forks (forge, reforge) has some kind of thing working. But for testing purposes, so who knows if it simply 'works'.
https://huggingface.co/w-e-w/torch-2.6.0-cu128.nv

[fathom sorrel]

in theory that should work for any project that uses pytorch ... comfy, invoke, etc

but i'm sure some things would not work

[surreal moss]

With the uv python though, is there even a way to install different pytorch? I'm going to assume the usual pip commands are not it...

[fathom sorrel]

yep, uv has a pip interface, and in general it's still working with the pypi ecosystem - installing prebuilt wheels should be similar

[surreal moss]

ok!

[sage quest]

I found this, which could help: https://forums.developer.nvidia.com/t/software-migration-guide-for-nvidia-blackwell-rtx-gpus-a-guide-to-cuda-12-8-pytorch-tensorrt-and-llama-cpp/321330

[sage quest]

https://github.com/pytorch/pytorch/issues/145570

[hot pawn]

I am also waiting for the support of my RTX 5080. please make a (beta) version of invoke for that.

[surreal moss]

It's not something about Invoke. It's torch that needs to be updated for RTX 50xx

[sage quest]

It seems like they are making progress though, so all in all a good sign

The windows builds should be ready in some time I think

[surreal moss]

Yeah they've been working on it for several weeks. Not moving as quickly as blackwell owners would like, I think 🙂

[sage quest]

Yeah haha, I thought there were some nightly versions of PyTorch already a few days after the release of the Blackwell generation, but after some research I found out that they are only for Linux

ComfyUI already seems to have some kind of workaround already

They are moving forward!

Only ONE thing is missing now!

It seems like they get done just now (in a few days maybe?)

Committed 18 hours ago

[solemn flame]

hello! I didn't want to start a whole new thread just for this, looks like the solution's here ... mind if I ask how to apply this fix? I just installed invoke today and I'm getting the same error; it doesn't like my 5080 lol (community edition if that matters)

[sage quest]

There is no fix for that at the moment, we have to wait until PyTorch is fully updated to CUDA Version 12.8 and until it is implemented into InvokeAI

[solemn flame]

ohhh i see okay

[marble pumice]

what about that?
https://forums.developer.nvidia.com/t/software-migration-guide-for-nvidia-blackwell-rtx-gpus-a-guide-to-cuda-12-8-pytorch-tensorrt-and-llama-cpp/321330

[sage quest]

Oh that's a good article, I think this could help the developers to migrate the software to CUDA 12.8 for the new Blackwell architecture

[sage quest]

Support for new NVIDIA RTX 50xx-Series GPUs

[sage quest]

I have found a solution! Some guy on Reddit (props to him!) answered me and it just works with my RTX 5080

[sage quest]

#1335289732854190170 message

[marble pumice]

i wait till this is in the installer. I installed the newest version v5.7.2rc2 and it is still not working with the 5080 rtx because of pytorch:

[obtuse mango]

https://www.reddit.com/r/invokeai/comments/1ix3s77/comment/mexqtyt/?context=3
I found this comment but I'm confused on step 4, how do I install the two files?

[tame bison]

Yeah - trying to patch this support together has been frustrating, is there any indication that official support is incoming?

[sage quest]

In the developer console, write "pip install (path to the first file)"
And "pip install (path to the other file)"

[sage quest]

It's not that hard, just do all the steps that are explained in this Reddit comment and it should work

[sage quest]

You need to get the latest nightly release of PyTorch, carefully read the image above your message

[obtuse mango]

It worked, thank you!!
My 5080 is so fast!

[sage quest]

Mine is also very fast!

[golden cobalt]

I have a 5080, i'm having the issue installing these to fix that No Kernal Image error as some other people are getting.

[fathom sorrel]

yeah, that's due to a permissions update microsoft did - they don't let you run powershell scripts.

#new-release-discussion message this should fix it

[sage quest]

Why are you using power shell? You can just simply do the steps in the developer console #1335289732854190170 message

[fathom sorrel]

the developer console uses powershell in the new launcher

[sage quest]

Oh okay, I'm still using the legacy install method for this

Without powershell

[tame bison]

Hey! I finally got this to work for me. Thank you very much!

[sage quest]

I'm glad to help! 🙂

[tame bison]

Is there a specific way we'll need to update the app to keep things working?

[fathom sorrel]

you'll have to update torch again if any invokeai updates replace it

[tame bison]

Thanks! Do you know if there's a stated plan for updating this support into the main Invoke app/install?

[jagged lynx]

any one got it working yet or is it a waiting game now? ty

[tame bison]

After many attempts here's what worked for me.

Installing Invoke on Blackwell Architecture GPUs (circa mid-March 2025)
Note: Since I was on a new system, I deleted the Invoke folders from previous installation attempts and started anew.

1. Update NVIDIA Drivers

2. Review Invoke AI Requirements:
   Follow the Legacy Script Install requirements:
   Invoke AI Requirements

3. Install Invoke AI via Legacy Script (download the latest .zip):
   Follow the instructions here:
   Legacy Script Installation

4. Download PyTorch & Torchvision Wheels for Python 3.11:
   Download the following files from:
   Hugging Face - Torch 2.6.0 cu128.nv

   • Torch: torch-2.6.0+cu128.nv-cp311-cp311-win_amd64.whl
   • Torchvision: torchvision-0.20.0a0+cu128.nv-cp311-cp311-win_amd64.whl

5. Ensure CUDA Toolkit 12.8 Is Installed:
   Verify that CUDA Toolkit 12.8 is installed by checking your Windows Environment Variables for a variable like CUDA_PATH_v12.8. If it’s not installed, download and install it from:
   CUDA Toolkit 12.8 Download Archive

6. Install PyTorch & Torchvision via the Developer Console:
   Open the Invoke AI executable, enter the developer console, and run:

   pip install "C:\path\to\torch-2.6.0+cu128.nv-cp311-cp311-win_amd64.whl"
   pip install "C:\path\to\torchvision-0.20.0a0+cu128.nv-cp311-cp311-win_amd64.whl"

   (Replace C:\path\to\... with the actual paths to your wheel files.)

7. Launch Invoke AI:
   Once the packages are installed, start Invoke AI. It should now work correctly with your GPU.

[sage quest]

This! @jagged lynx

[jagged lynx]

ill give it a try thank you 🙂

[jagged lynx]

pip install "C:\Users\mrgav\Downloads\torch-2.6.0%2Bcu128.nv-cp311-cp311-win_amd64.whl"
WARNING: Requirement 'C:\Users\mrgav\Downloads\torch-2.6.0+cu128.nv-cp311-cp311-win_amd64.whl' looks like a filename, but the file does not exist
ERROR: torch-2.6.0+cu128.nv-cp311-cp311-win_amd64.whl is not a supported wheel on this platform.
ERROR: torch-2.6.0+cu128.nv-cp311-cp311-win_amd64.whl is not a supported wheel on this platform.

[sage quest]

You have to do this inside of the developer console

And by the way, which version of Python are you using?

[jagged lynx]

Python 10 the tried 11 no luck ty

[steel slate]

With my RTX 5070 i got: The installed version of bitsandbytes was compiled without GPU support. 8-bit optimizers, 8-bit...

Flux won´t work, to get it to work only go to Dev-Console and do:

pip install -U bitsandbytes

[steel slate]

you used the executable installer...

it worked with the text-based (legacy-)installer, available at release-page (5.6.0, no worry it installs 5.9.0 or newer)

Use Python 3.11 if you want copy&paste

Don´t use powershell, use commandline!

[jagged lynx]

trying that now

[jagged lynx]

got it working but really slow :/ Fetching 17 files: 100%|███████████████████████████████████████████████████████████████████████| 17/17 [00:00<?, ?it/s]
Loading pipeline components...: 100%|████████████████████████████████████████████████████| 7/7 [00:02<00:00, 2.76it/s]
100%|██████████████████████████████████████████████████████████████████████████████████| 30/30 [08:45<00:00, 17.53s/it]
[2025-03-30 10:34:22,567]::[InvokeAI]::INFO --> Graph stats: 89fe9ae1-bb5e-4186-9951-7281086529cb
Node Calls Seconds VRAM Used
sdxl_model_loader 1 0.008s 0.000G
sdxl_compel_prompt 2 3.857s 0.000G
collect 2 0.001s 0.000G
noise 1 0.001s 0.000G
denoise_latents 1 525.882s 0.000G
l2i 1 12.650s 0.000G
core_metadata 1 0.001s 0.000G
img_resize 1 0.026s 0.000G
TOTAL GRAPH EXECUTION TIME: 542.425s
TOTAL GRAPH WALL TIME: 542.429s
RAM used by InvokeAI process: 14.79G (+14.244G)
RAM used to load models: 12.92G
RAM cache statistics:
Model cache hits: 9
Model cache misses: 1
Models cached: 7
Models cleared from cache: 0
Cache high water mark: 12.92/0.00G

[sage quest]

That seems to work

[jagged lynx]

works but took 8 mins to do a 512 by 512 I think its using the cpu ?

[sage quest]

Well, there should be an indication if it is using the CPU or the GPU, maybe you are out-of-memory?

[jagged lynx]

maybe how do i check that :/ i got a 5090 and 64 gig ram

[sage quest]

Yeah it should only take a few seconds, not 8 Minutes

[broken elk]

I didn't see anyone have this in one place so I thought i'd add it.

Blackwell RTX 50* cards
Open the dev console in bottom left of launcher

Set-ExecutionPolicy Unrestricted -Scope process
.\.venv\Scripts\activate.ps1
uv pip install --upgrade --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128
uv pip install -U bitsandbytes

[jagged lynx]

still no luck for me

[late harbor]

thanks! I've managed to do it using this tutorial. Very fast speed indeed, using 5080. Haven't tried Flux yet.

Also, since I can't use Legacy Script, I've installed the PyTorch & Torchvision first, after that installed InvokeAi.

[ripe sluice]

same

Anyone who can't get it working (like me), use Stability Matrix and install it that way, much simpler, after installing just click the 3 dots on the top right, click the + sign, then paste

--pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128

Then off to the races, can use everything, including Flux Quanitized versions

[broken elk]

I would get rid of the install and start a new with what I had above.I can tell you that you didn't run the pip update I have because of the warning in your image, but that's the least of your issues currently

that or your GPU isn't recognized, which means the install wasn't upgraded to nightly torch

[ripe sluice]

No worries, I've spent hours over the last few months trying different things, I'm content for now with stability matrix lol

[kind laurel]

ugh maybe i should have stuck with automatic1111 cause i actually got it working

just trying to get it to work with my rtx 5080 and i did some of the steps I found here but am tired...

[crimson jetty]

try this - after opening the invoke app (reset your invoke install) - click on manage, click next, select the latest stable version (V5.10.0 at the moment), click next, select the nvidia 30x and above option, click next, click on the repair mode checkbox and click install.

after this completes, click on the Dev console button in the lower left hand corner of the window and click on "start dev console". run the following command

uv pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128 --upgrade --force-reinstall

try running invoke after that

[kind laurel]

I can't seem to run the virtual environment in the dev console but can do it in a regular cmd

I ran the launcher in admin

[kind laurel]

Ugh this is so annoying...

[kind laurel]

If I can't figure it out I'll be reverting back to a1111

[sage quest]

You can do it in the regular CMD if you use the legacy install method as described (not the installer)

[minor shadow]

yes recently I have the same issue not sure what is up with the dev console

plus the dev console was fine a week ago 🥲

[minor shadow]

I think the issue is with the last release.... can't make it work anymore 5.9.1 was fine I think

[minor shadow]

I wish it Invoke was as easy to do as getting CComfyUI desktop to work on a 50 series card

just change the UV torch location in the setting, delete torch in the venv habe the app repair/reinstall torch 😉

[minor shadow]

also if it helps I go my Invoke working again but had to delete the python UV folders

[kind laurel]

I got automatic1111 to work but haven't gotten invoke to generate an image yet, but I had to change the parameters for invoke... I can't easily go back 😭

I'm also not entirely sure what those are you're talking about but I shall research

[kind laurel]

omg python why

i need the virtual enviroment to only use those specific versions of torch, which i am trying to put on the D drive, but I think since i previously installed some stuff for A1111 on the C drive it's using python globally and picking up on those, despite me trying to delete everything that says "torch" on the C drive....

https://tenor.com/view/this-cannot-continue-stinkbugs11number1smeck-fan-cat-cat-in-water-water-cat-gif-25921319

idk how coder people do this stuff it's maddening

[steel slate]

Invokai installed with Installer.exe worked for me with, i have a rtx 5070:

Run installer update/install to 5.10.0 (repair mode)

download torch-2.6.0+cu128.nv-cp312-cp312-win_amd64.whl from huggingface link in https://discordapp.com/channels/1020123559063990373/1335289732854190170/1353582518691037257

download torchvision-0.20.0a0+cu128.nv-cp312-cp312-win_amd64.whl from huggingface link in https://discordapp.com/channels/1020123559063990373/1335289732854190170/1353582518691037257

open in installer dev-console run:
uv pip install "D:\torch-2.6.0+cu128.nv-cp312-cp312-win_amd64.whl"
uv pip install "D:\torchvision-0.20.0a0+cu128.nv-cp312-cp312-win_amd64.whl"
(change D:\ to your path)

run invokeai

[minor shadow]

yes sounds a bit like the issue I had, was a mess

but I think the culprit at lease in my case was the python cache not python itself

@green rivet @nimble rampart sorry if you are not the right ones to ask, have you ever thought of adding a 50 series install option in the launcher? This issue probably will keep popping up up till the next big PyTorch release

[nimble rampart]

Certainly. But:

• We don't have any 50xx series cards to do testing.
• The requisite versions of torch and related dependencies are not yet stable, as far as I know.

That said, the launcher is fully open-source. I'm happy to review a PR that adds support for 50xx series GPUs.

[green rivet]

Supporting a non-stable torch as a “stable” option seems like a bad idea.

Once the stable torch is released, it should “just work”

[minor shadow]

ok thanks for the feedback

[kind laurel]

Idk if this was for me but I've done all that and invoke won't use that torch version because it's reading something from a different drive

Like yall said the virtual environment should be independent yet before I force commanded it, it wouldn't fully integrate the newer version because an older version already existed

[kind laurel]

How long it gonna take them to make all the right stable versions

[green rivet]

im shocked they havent released it yet

[kind laurel]

I'm giving up till then... I've been at this for like 5 days..

[nimble rampart]

It really sounds like your venv was borked. There’s no reason the invoke app venv, created by the launcher, will use any other packages than those installed to it.

FWIW, The launcher does not do anything fancy at all. It just activates and uv pip installs. So I’m skeptical that the launcher is causing problems here.

[latent coral]

After trying a few methods, this worked for me. Thanks!

[kind laurel]

I am having ChatGPT help me with the majority of this, so I'm obviously no genius, but what else is going on here?

Like most of those are the right torch version but why is it pulling from there... idk code is confusing

I'll try to give a short breakdown of where I'm at but I'm burnt out.. might try again soon tho

[fathom sorrel]

you should be using uv pip instead of pip. by default in a terminal, uv won't be in the path, so you can either add it, or just use the launcher's dev console

[kind laurel]

is this helpful?

ill uninstall torch and torchvision using uv pip lol

then re-install

omg so confused, do i want 126 to show for torch or 128

(I feel like it should be 128)
IDK where i got these specific whl's from(chatgpt) but the uv is saying they aren't compatible

then chatgpt says to force it... why not XD

[kind laurel]

Welp. Something broke 😅

[jagged lynx]

😦 ow long dose it take for the update for the 5090 cards

[sage quest]

You need 128 for the new RTX 50XX series

[sage quest]

With the tutorial that's in here (#1335289732854190170 message) it already works for the new RTX 50XX cards

[sage quest]

Yes, there can be error messages that the newest version is not compatible, but you can ignore those and it will still work

[nimble rampart]

It looks like pytorch is preparing for their 2.7.0 release which supports the new cards. Once that's stable Invoke will work on support

[west silo]

I (foolishly) just got an RTX 5070ti, thinking it would be an improvement, never imagining InvokeAI would fail to run completely. I am NOT a techie (which is why I love Invoke so much, it's clear and easy to use). I'm trying to solve it—clearly others have made it work—but, for instance, when I read things like "install on Blackwell Architecture GPU" I have no idea what that means: should I be installing something called Blackwell Architecture GPU, or is it a solution that only works with people who have this marvellous architecture already? When I start Invoke it tells me up front that it won't work as is, and I need to go to pytorch and follow the instructions there, but those instructions don't really exist. Is there any resource yet for the dummies like me who just spend $1500 and now can't use their computer?

[fathom sorrel]

pytorch 2.7 was just released, and it should support the new series of cards. that said, it was just released, and invoke hasn't had a new release since then. your options are to be patient for maybe a few weeks, or you can try clicking the 'gear' icon in the lower-left of the launcher, click 'open dev console', then run these 3 commands (might need to close and re-open the launcher after the first one) if on windows:

Set-ExecutionPolicy Unrestricted -Scope process
.\.venv\Scripts\activate.ps1
uv pip install -U torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
uv pip install -U bitsandbytes

[west silo]

Thanks so much. I've tried that (just now, thanks to you), and I got it to run those commands which in itself is a thrill. But when Invoke starts up, it's still the same warnings: D:\Invoke AI Five.Ten.venv\Lib\site-packages\torch\cuda_init_.py:235: UserWarning:
NVIDIA GeForce RTX 5070 Ti with CUDA capability sm_120 is not compatible with the current PyTorch installation.
The current PyTorch install supports CUDA capabilities sm_50 sm_60 sm_61 sm_70 sm_75 sm_80 sm_86 sm_90.
If you want to use the NVIDIA GeForce RTX 5070 Ti GPU with PyTorch, please check the instructions at https://pytorch.org/get-started/locally/

warnings.warn(
[2025-04-25 20:57:47,468]::[InvokeAI]::INFO --> Using torch device: NVIDIA GeForce RTX 5070 Ti
[2025-04-25 20:57:50,578]::[InvokeAI]::INFO --> cuDNN version: 90501
[2025-04-25 20:57:52,400]::[InvokeAI]::INFO --> Patchmatch initialized
[2025-04-25 20:57:52,806]::[InvokeAI]::INFO --> InvokeAI version 5.10.1

I'm happy to pay extra money if I need like a Geek Squad type person to sort this out for me remotely ... is this the sort of issue reasonable competent technical types can solve? I know it's beyond me, but I'd love to be able to get back to business as usual as soon as possible, rather than waiting and hoping ...

[fathom sorrel]

looks like the commands didn't take

anyhow, be patient and wait for an official release. in the future, check on compatibility and the state of things before big purchases 😉

[minor shadow]

you might have the same issue I had, make sure you are on D: in the dev console and not C:

the commands gogurt enjoyer posted will not work if you are on the wrong drive

not sure why the dev console is not in the right path maybe it is a bug or because I used to have Invoke on the C: drive at first 🤷‍♂️

[lilac charm]

When I clicked the dev console button, it simply told me the directory where Invoke had been installed. I opened up Windows Terminal in PowerShell mode, navigated to that directory, and ran the commands. Invoke loaded without error, but then crashed on my first attempt to generate an image, but that was due to me trying to load a local model I'd downloaded directly from HuggingFace. When I downloaded Flux Dev through the app (required adding a HF token), the GUI crashed after installation. After a restart, it works. I generated this image in 42.4 seconds on a 5080 with Flux Dev. That said, I'll probably wipe and reinstall when the official release with PyTorch 2.7 support drops.

[west silo]

Thanks! I tried following the steps switching to D: first before running them, but ultimately the same errors and error messages persist. Looks like I'll just have to be patient.

[minor shadow]

also not sure if anyone is using them but the latest studio driver are behaving weird for me I want back to the ones from march

[kind laurel]

oh. my. god

i got it 🥺

it's beautiful

[wild forge]

Thanks man! It works!

[plush rune]

For anyone still having issues I got it to work using these commands but switching it to /whl/nightly/cu128

[marble pumice]

Thanks! This is finaly working on my RTX 5080 🙂

[kind laurel]

Where is all this at right now? Have there been new updates?

[dreamy schooner]

it works in the new versions

[kind laurel]

im a bit confused on how to update to the latest version lol