Nvidia 50 Series (Blackwell) support thread: How to get ComfyUI running on your new 50 series GPU.

[comfyanonymous]

I will try keeping this post up to date as much as possible with the latest developments.

To get your nvidia 50 series GPU working with ComfyUI you need a pytorch that has been built against cuda 12.8

In the next few months there will likely be a lot of performance improvements landing in pytorch for these GPUs so I recommend coming back to this page and updating frequently.

Windows

The recommended download is the standalone package with nightly pytorch 2.7 cu128 that you can download from here

Old package with torch 2.6

Windows users can download this standalone ComfyUI package with a cuda 12.8 torch build

Manual Install

At this moment Blackwell is not yet supported by stable pytorch.

pytorch nightly cu128 is available for Windows and Linux:

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

You can also use the Nvidia Pytorch Docker Container as an alternative which might give more performance.

Link: https://catalog.ngc.nvidia.com/orgs/nvidia/containers/pytorch

Here's how to use it:

docker run -p 8188:8188 --gpus all -it --rm nvcr.io/nvidia/pytorch:25.01-py3

Inside the docker container:

git clone https://github.com/comfyanonymous/ComfyUI
cd ComfyUI
grep -v 'torchaudio\|torchvision' requirements.txt > temp_requirements.txt
pip install -r temp_requirements.txt
python main.py --listen

[mcmonkey4eva]

There are also Windows prebuilt wheels of pytorch on cuda 128 that nvidia gave w-e-w to publish https://huggingface.co/w-e-w/torch-2.6.0-cu128.nv

[mmartial]

FYSA, the nightly for 12.8 appears available https://download.pytorch.org/whl/nightly/cu128

[comfyanonymous]

Yeah I see the package but it's still missing the cu128 torchvision, torchaudio and windows packages.

[mmartial]

It is. Thanks for clarifying the components we are waiting for.

I was trying to follow the guide by removing torchvision and torchaudio
and doing a pip3 install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu128

spandrel too uses torchvision, which forces a downgrade on a Linux installation within a virtualenv.
https://github.com/chaiNNer-org/spandrel/blob/main/libs/spandrel/pyproject.toml#L36

EDIT: I will postpone testing until this is ready. Thanks for the details

[comfyanonymous]

I just noticed there are now nightly torchvision and torchaudio cu128 wheels for Linux. No windows packages yet though.

I tried them and they seem to work well so I have updated the post.

[mmartial]

Thank you 👍

With this news, I have made an updated release of https://github.com/mmartial/ComfyUI-Nvidia-Docker

For CUDA 12.8, I pre-install torch using pip3 install --pre torch torchaudio torchvision --index-url https://download.pytorch.org/whl/nightly/cu128 before installing ComfyUI from git and installing its requirements. I have tested it and can generate the bottle :)

No 50xx yet for me, so this test was done on a 3090 with driver 570 installed.

[edisonchan]

The Windows version said:
ModuleNotFoundError: No module named 'torchaudio'

[comfyanonymous]

This is because there was no torchaudio cu128 package provided.

The only thing this affects is the audio related nodes so it only affects you if you are using stable audio.

[edisonchan]

nf4 node does not work.

[ltdrdata]

Anything related to NF4 support should be resolved in the node pack repo that provides NF4 feature.

[edisonchan]

ImportError: tokenizers>=0.21,<0.22 is required for a normal functioning of this module, but found tokenizers==0.20.3.
Try: pip install transformers -U or pip install -e '.[dev]' if you're working with git main

How can I do to fix problem?

[adamreading]

Sorry - I have been banging my head against this for 6 hours - trying to run a cosmos workflow with 5090 - without SageAttention or TorchCompile - a 14 minute job on 4090 is taking 25 minutes. To turn them on I installed Triton and Sageattention - then I get this error afterwards whether I bypass the patch and compile or not:- KSampler
Command '['C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Tools\MSVC\14.42.34433\bin\Hostx64\x64\cl.EXE', 'C:\Users\JOANNA1\AppData\Local\Temp\tmpy8t3476k\cuda_utils.c', '/nologo', '/O2', '/LD', '/wd4819', '/IC:\Users\JoannaThompson\ComfyUI\ComfyUI_cu128_50XX\python_embeded\Lib\site-packages\triton\backends\nvidia\include', '/IC:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.8\include', '/IC:\Users\JOANNA1\AppData\Local\Temp\tmpy8t3476k', '/IC:\Users\JoannaThompson\ComfyUI\ComfyUI_cu128_50XX\python_embeded\Include', '/IC:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Tools\MSVC\14.42.34433\include', '/IC:\Program Files (x86)\Windows Kits\10\Include\10.0.26100.0\shared', '/IC:\Program Files (x86)\Windows Kits\10\Include\10.0.26100.0\ucrt', '/IC:\Program Files (x86)\Windows Kits\10\Include\10.0.26100.0\um', '/FoC:\Users\JOANNA1\AppData\Local\Temp\tmpy8t3476k\cuda_utils.cp312-win_amd64.obj', '/link', '/LIBPATH:C:\Users\JoannaThompson\ComfyUI\ComfyUI_cu128_50XX\python_embeded\Lib\site-packages\triton\backends\nvidia\lib', '/LIBPATH:C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.8\lib\x64', '/LIBPATH:C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Tools\MSVC\14.42.34433\lib\x64', '/LIBPATH:C:\Program Files (x86)\Windows Kits\10\Lib\10.0.26100.0\ucrt\x64', '/LIBPATH:C:\Program Files (x86)\Windows Kits\10\Lib\10.0.26100.0\um\x64', '/LIBPATH:C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Tools\MSVC\14.42.34433\lib\x64', '/LIBPATH:C:\Program Files (x86)\Windows Kits\10\Lib\10.0.26100.0\ucrt\x64', '/LIBPATH:C:\Program Files (x86)\Windows Kits\10\Lib\10.0.26100.0\um\x64', '/LIBPATH:C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Tools\MSVC\14.42.34433\lib\x64', '/LIBPATH:C:\Program Files (x86)\Windows Kits\10\Lib\10.0.26100.0\ucrt\x64', '/LIBPATH:C:\Program Files (x86)\Windows Kits\10\Lib\10.0.26100.0\um\x64', '/LIBPATH:C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Tools\MSVC\14.42.34433\lib\x64', '/LIBPATH:C:\Program Files (x86)\Windows Kits\10\Lib\10.0.26100.0\ucrt\x64', '/LIBPATH:C:\Program Files (x86)\Windows Kits\10\Lib\10.0.26100.0\um\x64', '/LIBPATH:C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Tools\MSVC\14.42.34433\lib\x64', '/LIBPATH:C:\Program Files (x86)\Windows Kits\10\Lib\10.0.26100.0\ucrt\x64', '/LIBPATH:C:\Program Files (x86)\Windows Kits\10\Lib\10.0.26100.0\um\x64', '/LIBPATH:C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Tools\MSVC\14.42.34433\lib\x64', '/LIBPATH:C:\Program Files (x86)\Windows Kits\10\Lib\10.0.26100.0\ucrt\x64', '/LIBPATH:C:\Program Files (x86)\Windows Kits\10\Lib\10.0.26100.0\um\x64', 'cuda.lib', '/OUT:C:\Users\JOANNA1\AppData\Local\Temp\tmpy8t3476k\cuda_utils.cp312-win_amd64.pyd', '/IMPLIB:C:\Users\JOANNA1\AppData\Local\Temp\tmpy8t3476k\cuda_utils.cp312-win_amd64.lib', '/PDB:C:\Users\JOANNA1\AppData\Local\Temp\tmpy8t3476k\cuda_utils.cp312-win_amd64.pdb']' returned non-zero exit status 2.

I have run this through GPT and followed every instruction. Step 1: Verify CUDA Toolkit Installation
Run the following command in ComfyUI's embedded Python:

python
Copy
Edit
import torch
print(torch.version.cuda)
Expected Output: 12.8

If it does not match 12.8, you may need to reinstall PyTorch with the correct CUDA version.

✅ Step 2: Check NVIDIA Toolkit and Drivers
Ensure you have the latest NVIDIA GPU driver and CUDA toolkit installed.

Update your NVIDIA driver (Download latest)
Confirm CUDA installation with:
powershell
Copy
Edit
nvcc --version
Expected Output: release 12.8
If not, reinstall CUDA 12.8 from NVIDIA.

✅ Step 3: Fix Microsoft Visual Studio Build Tools
Open Developer Command Prompt for VS 2022

Run:

sh
Copy
Edit
cl.exe
If it says command not found, reinstall Microsoft C++ Build Tools:
Download: Visual Studio Build Tools 2022
Install:
C++ CMake tools for Windows
MSVC v143 - VS 2022 C++ x64/x86 build tools
Windows 10 SDK (latest version)
Restart the PC.

[rubentorresbonet]

anyone knows if other pytorch's cuda versions like 12.6 will work with blackwell?

[mpfcs162]

No, Blackwell GPUs require 12.7 or newer IIRC.
Edit: So 12.8 since that is the current version. It's a new architecture so I guess it makes sense that 12.6 and older versions are not supported. CUDA is not backwards compatible?

[fasfafefggwgrw]

When I use [ComfyUI package with a cuda 12.8 torch build], many custom_nodes are "IMPORT FAILED" including Manager, InstantID, ReActor ...

[Shad0w23333]

same, is is possible to build a PyTorch that works for 50s?

[mpfcs162]

Check Python version, reinstall dependencies, and look for incompatibilities in the logs - some nodes might need updates for CUDA 12.8. Fresh install or upgrade?

[Shad0w23333]

I will try keeping this post up to date as much as possible with the latest developments.

To get your nvidia 50 series GPU working with ComfyUI you need a pytorch that has been built against cuda 12.8

In the next few months there will likely be a lot of performance improvements landing in pytorch for these GPUs so I recommend coming back to this page and updating frequently.

Windows

Windows users can download this standalone ComfyUI package with a cuda 12.8 torch build

Manual Install

At this moment Blackwell is not yet supported by stable pytorch.

pytorch nightly cu128 is available for Linux:

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

You can also use the Nvidia Pytorch Docker Container as an alternative which might give more performance.

Is is possible to build a pytorch that works for 5090? If so how to do it?

[mpfcs162]

What do you mean? There is already one for CUDA 12.8: https://download.pytorch.org/whl/nightly/cu128/torch/ For Windows you can try the EA version found here: https://huggingface.co/w-e-w/torch-2.6.0-cu128.nv Or just use the standalone package linked above with a CUDA 12.8 torch build.

We spent 6 hours + yesterday trying to install comfyui portable around the special version at the top of the screen. It wont let me run Sageattention or Torchcompile in workflows - and because it’s missing TorchAudio - things I use every day like GripTape won’t load into custom nodes any more. Running the cosmos workflow without sage and compile - is twice as long as on a 4090 system (25 mins for 8 sec gen compared to 14 minutes on 4090) Although I was able to install SageAttention and Triton I don’t think they are working with this version of PyTorch

Have you checked the logs for errors or tried a clean environment to rule out conflicts?

[adamreading]

What do you mean? There is already one for CUDA 12.8: https://download.pytorch.org/whl/nightly/cu128/torch/ For Windows you can try the EA version found here: https://huggingface.co/w-e-w/torch-2.6.0-cu128.nv Or just use the standalone package linked above with a CUDA 12.8 torch build.

We spent 6 hours + yesterday trying to install comfyui portable around the special version at the top of the screen. It wont let me run Sageattention or Torchcompile in workflows - and because it’s missing TorchAudio - things I use every day like GripTape won’t load into custom nodes any more. Running the cosmos workflow without sage and compile - is twice as long as on a 4090 system (25 mins for 8 sec gen compared to 14 minutes on 4090) Although I was able to install SageAttention and Triton I don’t think they are working with this version of PyTorch

Have you checked the logs for errors or tried a clean environment to rule out conflicts?

Yes we went through the logs - they are in the earlier post above - we tried for 6 hours downloading different versions - Visual build tools, various Nvida toolkits - did everything GPT suggested - its a brand new out of the box machine - nothing previously installed except office.

[adamreading]

5090 Crash 10.txt
One of the complete logs

[mpfcs162]

Tried running Linux? https://download.pytorch.org/whl/nightly/cu128/torch/
To my knowledge there is no official Windows build of PyTorch that supports 12.8 yet.
Sorry I can't offer more help...

[adamreading]

Tried running Linux? https://download.pytorch.org/whl/nightly/cu128/torch/ To my knowledge there is no official Windows build of PyTorch that supports 12.8 yet. Sorry I can't offer more help...

Thanks for looking. The app I want to run (a comfyui-discord bot) requires portable and lm
Studio to work apparently. Plus I use the machine for office365 at the same time lol. Looks like I just have to be patient and work out what will run and what won’t. It ran an ltxv video generator in 36 seconds so some stuff is ok. Couldn’t load anything Pulid.

[adamreading]

Having given up on Portable for now - too many errors lol - I am using WSL - Ubuntu - everything else is set up and working in Comfyui - I have the latest pytorch nightly from pip install --pre torch torchvision torchaudio --index-url[ https://download.pytorch.org/whl/nightly/cu128] - but when I use Sageattention I get this:
loaded completely 29493.675 3667.902587890625 True
0%| | 0/80 [00:00<?, ?it/s]'sm_120' is not a recognized processor for this target (ignoring processor)
'sm_120' is not a recognized processor for this target (ignoring processor)
'sm_120' is not a recognized processor for this target (ignoring processor)
'sm_120' is not a recognized processor for this target (ignoring processor)
'sm_120' is not a recognized processor for this target (ignoring processor)
'sm_120' is not a recognized processor for this target (ignoring processor)
'sm_120' is not a recognized processor for this target (ignoring processor)
'sm_120' is not a recognized processor for this target (ignoring processor)
'sm_120' is not a recognized processor for this target (ignoring processor)
'sm_120' is not a recognized processor for this target (ignoring processor)
'sm_120' is not a recognized processor for this target (ignoring processor)
'sm_120' is not a recognized processor for this target (ignoring processor)
LLVM ERROR: Cannot select: intrinsic %llvm.nvvm.shfl.sync.bfly.i32
Aborted (core dumped)

[adamreading]

I managed to get it working in Windows yesterday. Had to manually install the torch-2.6.0+cu128.nv-cp312-cp312-win_amd64.whl and torchvision-0.20.0a0+cu128.nv-cp312-cp312-win_amd64.whl.

Originally had the wrong Python version installed. Easy fix but took me a long time (longer than I want to admit) to figure out...

There are some errors though (torchaudio missing), but I'm running SD 3.5 L to test on a 5090 and I get output. And fast too. But 32GB VRAM is not as much as I'd hoped and the impact compared to a 4090 is not that big - about ~30% faster which is expected I guess...

But I'd honestly recommend just running it on Linux atm.
Yeah - I spent the last 6 hours trying to set all the stuff up - had to compile sageattention and apex - xformers I don't think is compatible yet - running it after lunch to test

[Shad0w23333]

I managed to get it working in Windows yesterday. Had to manually install the torch-2.6.0+cu128.nv-cp312-cp312-win_amd64.whl and torchvision-0.20.0a0+cu128.nv-cp312-cp312-win_amd64.whl.

Originally had the wrong Python version installed. Easy fix but took me a long time (longer than I want to admit) to figure out...

There are some errors though (torchaudio missing), but I'm running SD 3.5 L to test on a 5090 and I get output. And fast too. But 32GB VRAM is not as much as I'd hoped and the impact compared to a 4090 is not that big - about ~30% faster which is expected I guess...

But I'd honestly recommend just running it on Linux atm.

Wait, can u explain how to do that? Thank u

[mpfcs162]

Wait, can u explain how to do that? Thank u

Do what?

[adamreading]

I managed to get it working in Windows yesterday. Had to manually install the torch-2.6.0+cu128.nv-cp312-cp312-win_amd64.whl and torchvision-0.20.0a0+cu128.nv-cp312-cp312-win_amd64.whl.

Originally had the wrong Python version installed. Easy fix but took me a long time (longer than I want to admit) to figure out...

There are some errors though (torchaudio missing), but I'm running SD 3.5 L to test on a 5090 and I get output. And fast too. But 32GB VRAM is not as much as I'd hoped and the impact compared to a 4090 is not that big - about ~30% faster which is expected I guess...

But I'd honestly recommend just running it on Linux atm.

Have you managed to get SageAttention working yet? (With Linux or Windows)

[mpfcs162]

No. I assume it won't be fully stable until we see an official release of PyTorch with CUDA 12.8 support. Or req some additional manual config. I just can't be bothered with that at this point. Been stuck in integration hell all day trying to troubleshoot some code for a client so haven't had much time to dig further into this topic. Sorry.

[Teuflor]

im using comfyui over pinokio, is there any way to update my comfy to work with my new 5080?

i deleted the old files and changed in the torch.js on the patch above, but it wont start

[gachapengames]

Will using a docker allow for a working torchvision in windows?

[Six6pounder]

@gachapengames Yes, use nvcr.io/nvidia/pytorch:25.01-py3 as image and it works

[gachapengames]

Can someone write a little guide for getting a docker running torchvision, etc on window with the portable Blackwell comfyui release? Please write it for a normal user who has no programming knowledge.

There are basic instructions in the OP, but what even is a docker?

"docker run -p 8188:8188 --gpus all -it --rm nvcr.io/nvidia/pytorch:25.01-py3"

Where is this command supposed to be run? Is a docker something to be installed to system python or standalone folder?

How would this be installed on a fresh portable comfyui install?

With more Blackwell cards trickling out, there will most likely be more users needing help setting this up. Thank you.

[mmartial]

Please check the "Windows: WSL2 using podman" from https://github.com/mmartial/ComfyUI-Nvidia-Docker

With the caveat that you will need to adapt the podman run line if you want to use the proposed containers to be something like (adapted to attempt to reflect the above order):

podman run --userns=keep-id -p 8188:8188 --device nvidia.com/gpu=all --rm -it nvcr.io/nvidia/pytorch:25.01-py3

with the caveat that everything will be stored within the container since you have no mount (the -v from my original writeup)

so maybe:

podman run --userns=keep-id -p 8188:8188 --device nvidia.com/gpu=all --rm -it -v `pwd`/comfy:/comfy nvcr.io/nvidia/pytorch:25.01-py3

and then do the other steps listed above (pip install...) after cd /comfy so that the files persist when you exit the container

[gachapengames]

To anyone having trouble with the portable version for Blackwell gpus, do not update it! I noticed during updating that it uninstalled the cu128 and installed another version for example. Torchvision works with a fresh install without updating!

[mpfcs162]

So I've used both Comfy UI and Stability Matrix with SD Forge WebUI these past days since getting my 5090.

Each instance needs to have the correct version of Torch, and updating/installing will not ensure that (edit: if anyone reads this in the future there will eventually be an official torch release that supports CUDA 12.8 and in that case what is written here might no longer valid). Best way is to install Torch from the files above, making sure you use the correct version for your system and Python install. For SM, you need to install it to the package directly for example.

And please, do not run commands you find online unless you understand what they do. This applies to the below commands as well.

Navigate to the correct dir and type this to uninstall torch, torchvision and torchaudio:
python_embeded\python.exe -m pip uninstall -y torch torchvision torchaudio

Then, for ComfyUI at least, you can run this. Adjust the command as needed:
python_embeded\python.exe -m pip install H:\torch\torch-2.6.0+cu128.nv-cp310-cp310-win_amd64.whl H:\torch\torchvision-0.20.0a0+cu128.nv-cp310-cp310-win_amd64.whl

You can run this in your dir to check what is installed:
python_embeded\python.exe -m pip list

[mpfcs162]

I just want to add that asking chatGPT (and similar LLMs) about this might not be as helpful as you think.

A very common issue is coherence with these models, and they can easily spew out answers that look and seem correct, while in reality are not what you actually need. For example, if I ask it why I'm getting kernel error when attempting to run stock ComfyUI it will tell me to update using a pip command, but it will give me a link to a build that does not exists (cu127) since it believes that CUDA 12.7 is the version I need for my 5090 since that is the min. version of CUDA that Blackwell supports.

I can follow that trail together with chatGPT for hours and end up trying to compile my own version, uninstall/reinstall various extensions and packages and get absolutely nowhere. And chatGPT will always act and respond with 100% confidence no matter what, and I will never figure out the culprit unless I go to the repo or follow the trail myself to discover that it (cu127) does not exists.

Of course, LLMs are great for a number of reasons - but beware when trying to do stuff like this that the output might be complete nonsense.

[adamreading]

I just want to add that asking chatGPT (and similar LLMs) about this might not be as helpful as you think.

A very common issue is coherence with these models, and they can easily spew out answers that look and seem correct, while in reality are not what you actually need. For example, if I ask it why I'm getting kernel error when attempting to run stock ComfyUI it will tell me to update using a pip command, but it will give me a link to a build that does not exists (cu127) since it believes that CUDA 12.7 is the version I need for my 5090 since that is the min. version of CUDA that Blackwell supports.

I can follow that trail together with chatGPT for hours and end up trying to compile my own version, uninstall/reinstall various extensions and packages and get absolutely nowhere. And chatGPT will always act and respond with 100% confidence no matter what, and I will never figure out the culprit unless I go to the repo or follow the trail myself to discover that it (cu127) does not exists.

Of course, LLMs are great for a number of reasons - but beware when trying to do stuff like this that the output might be complete nonsense.

I concur from personal experience - however as someone who doesn't understand this stuff - and having had an incredible machine arrive a week ago, that I just couldn't get to work - and all my techie friends on discord were at a loss as well - ChatGPT was my only option. Since talking to LLMS is what i do a lot for work etc, I just kept everything inside one thread - so it remembered everything we had been doing - I used its commands to trouble shoot - but used my brain to realise when we were repeating or when it had forgotten about having to use 12.8 sm_120 etc and kept reminding it of these things - and with a little prompting it helped me through a maze of tests and trials I never would have got through on my own. In the end - I went back to the original Comfy Windows stand alone install thats at the top of the page. I learned to install custom nodes in new workflows - 1 at a time, restarting and checking the logs after every node - and trouble shooting anything that threw up one step at a time. And it's now pretty much running. It is slow however - without SageAttention or FlashAttention etc I think it struggles. And I am sure there will be more performance increases as more tools get compiled to use its abilities. I am old and have been around a while - so I realise that life at the bleeding edge is not always easy. I am still confused about what the difference is between Comfyui Stand Alone and Comfyui Portable - half my friends are saying go one way the other the other - but I think they are just two different names for the same install? As far as I can tell (having installed all of these this week - for a Windows machine the options are 1) Desktop Comfy that installs from an exe file - doesn't open in a browser - has a built in terminal in the bottom pull up in a tab next to the logs, but seemed restrictive to use (though I can't tell you why). 2) The Windows Stand Alone I am currently using - which, once you figure out where to run the pip commands etc is pretty straight forward, easy to back up directories - or reinstall by dragging the zip file over the top again 3) WSL - allowed me to use TorchAudio right away - feels like gets quicker support from the community because its in Linux - but got quite complicated very quickly - and installing an unrelated program in a docker in the same WSL killed the whole thing - and I lost everything and no way to recover my files... @mpfcs162 a case in point for your statement - ChatGPT told me to remove the WSL to clear the issue with YuE and gave me the command to run - which I did - and then it wiped everything. (My own fault)

[edisonchan]

So I've used both Comfy UI and Stability Matrix with SD Forge WebUI these past days since getting my 5090.

Then, for ComfyUI at least, you can run this. Adjust the command as needed: python_embeded\python.exe -m pip install H:\torch\torch-2.6.0+cu128.nv-cp310-cp310-win_amd64.whl H:\torch\torchvision-0.20.0a0+cu128.nv-cp310-cp310-win_amd64.whl

You can run this in your dir to check what is installed: python_embeded\python.exe -m pip list

python_embeded\python.exe --version
Python 3.12.8
PS E:\ComfyUI_cu128_50XX> python_embeded\python.exe -m pip install ".\update\torch-2.6.0+cu128.nv-cp310-cp310-win_amd64.whl" ".\update\torchvision-0.20.0a0+cu128.nv-cp310-cp310-win_amd64.whl"
ERROR: torch-2.6.0+cu128.nv-cp310-cp310-win_amd64.whl is not a supported wheel on this platform.

not work?

[adamreading]

So I've used both Comfy UI and Stability Matrix with SD Forge WebUI these past days since getting my 5090.
Then, for ComfyUI at least, you can run this. Adjust the command as needed: python_embeded\python.exe -m pip install H:\torch\torch-2.6.0+cu128.nv-cp310-cp310-win_amd64.whl H:\torch\torchvision-0.20.0a0+cu128.nv-cp310-cp310-win_amd64.whl
You can run this in your dir to check what is installed: python_embeded\python.exe -m pip list

python_embeded\python.exe --version Python 3.12.8 PS E:\ComfyUI_cu128_50XX> python_embeded\python.exe -m pip install ".\update\torch-2.6.0+cu128.nv-cp310-cp310-win_amd64.whl" ".\update\torchvision-0.20.0a0+cu128.nv-cp310-cp310-win_amd64.whl" ERROR: torch-2.6.0+cu128.nv-cp310-cp310-win_amd64.whl is not a supported wheel on this platform.

not work?

You are running python 3.12. You are trying to use the wheels for python 3.10… first go here and download the two 3.12 files to the same directory you are running the command in. https://huggingface.co/w-e-w/torch-2.6.0-cu128.nv Them adapt the command to use your version of the wheel you just downloaded.

[Tathiane123]

Excelente

[Tayyab-H]

Are there no windows pytorch options available for cuda 12.8 still? Any help would be greatly appreciated.

[adamreading]

I couldn't find the 3b but I am downloading this to test memory lmstudio-community/Qwen2.5-7B-Instruct-1M-GGUF/Qwen2.5-7B-Instruct-1M-Q4_K_M.gguf

[adamreading]

Looks like you can import any model though - https://lmstudio.ai/docs/basics/import-model

[adamreading]

looks like 2gb of my gpu VRAM is there used all the time with nothing obviously running - then adding a 5gb Quen took it up to 7GB

[adamreading]

llava 1.5 used 12gb VRAM ran super fast

[Tayyab-H]

thanks for the help! Ill give LM studio a try! Its a shame it isnt working through code, id love to build stuff with it.

[alisson-anjos]

For those who want to use Sage Attention + Hunyuan + RTX 50xx with a simple tutorial, it is interesting to keep the files inside WSL instead of using the files in Windows due to the reading time (you can test it and you will understand) that's why I recommend copying your comfyui to WSL.

https://github.com/alisson-anjos/ComfyUI_Tutoriais/blob/main/WSL/install.md

Notes:

For those who want to compile SageAttention due to some lack of compatibility with the OS, it is necessary to compile Triton and then Sage Attention. To compile Sage Attention it is necessary to change setup.py with the code from this issue: [Sage Attention Issue] (thu-ml/SageAttention#107 (comment))

Thanks to Kijai for precompiling triton and sage attention.

[adamreading]

Thanks so much for making this, you really saved me from brain destruction this afternoon, and I’m glad it will help others as well. I can’t tell you how appreciated this was!! (AJO FittyNunty)

[maxious]

Flash Attention now supports and recommends CUDA 12.8 Dao-AILab/flash-attention@454ce31

[cyber827]

I will try keeping this post up to date as much as possible with the latest developments.

To get your nvidia 50 series GPU working with ComfyUI you need a pytorch that has been built against cuda 12.8

In the next few months there will likely be a lot of performance improvements landing in pytorch for these GPUs so I recommend coming back to this page and updating frequently.

Windows

Windows users can download this standalone ComfyUI package with a cuda 12.8 torch build

How do you install the ComfyUI-Manager on the standalone ComfyUI package with a cuda 12.8 torch build?

The script messes up the installation.

[adamreading]

check this: 6980

I’m already running comfyui in WSL2 with all Kijai’s wheels etc - all i am missing is BitsAndBytes - is the compiled wheel anywhere to be found? I have tried compiling it for hours and don’t get anywhere. I am loath to redo the whole setup for the sake of compiling in a different venv

[maxious]

@adamreading the version 0.45.3 of bitsandbytes you get from pypi now has windows and linux binaries for cuda 12.8 (libbitsandbytes_cuda128.dll and libbitsandbytes_cuda128.so) via bitsandbytes-foundation/bitsandbytes#1491

[adamreading]

@adamreading the version 0.45.3 of bitsandbytes you get from pypi now has windows and linux binaries for cuda 12.8 (libbitsandbytes_cuda128.dll and libbitsandbytes_cuda128.so) via bitsandbytes-foundation/bitsandbytes#1491

Amazing! Thanks. I’ll try that when I get home.

[adamreading]

@adamreading the version 0.45.3 of bitsandbytes you get from pypi now has windows and linux binaries for cuda 12.8 (libbitsandbytes_cuda128.dll and libbitsandbytes_cuda128.so) via bitsandbytes-foundation/bitsandbytes#1491

I just tried and it just says requirements already fulfilled - and still get the same errors:
"Could not find the bitsandbytes CUDA binary at PosixPath('/home/ajo6268/envs/comfyui/lib/python3.12/site-packages/bitsandbytes/libbitsandbytes_cuda128.so')
The installed version of bitsandbytes was compiled without GPU support. 8-bit optimizers, 8-bit multiplication, and GPU quantization are unavailable."

[adamreading]

libbitsandbytes_cuda128.so
This happened 2 hours ago... now all working with that upgrade command...
We released v0.45.3 with CUDA 12.8 builds. You can upgrade with pip install -U bitsandbytes.

[shidarin]

Would it be possible to get an updated comfyui build with WAN support? Downloading the above link still pins to 3.12.8 and downloading the latest release doesn't have built in blackwell support.

Run update/update_comfyui.bat to update after downloading, duh.

[comfyanonymous]

you can run the update/update_comfyui.bat script.

[shidarin]

Please excuse my comment- you are correct. I had run this and thought it was stuck at 3.12.8 because of the "Python version" in the log. A quick git log tells me I'm on the bleeding edge and something I should have done first. Thank you for taking the time.

[a1chera]

I thought people were saying NOT to update because it breaks the standalone package for blackwell gpus?

[Hapseleg]

I thought people were saying NOT to update because it breaks the standalone package for blackwell gpus?

it broke mine when i tried a couple of days ago

[Yuriuori]

thank you so much it worked !! hope they fix it asap
Also does anyone know how to make Kohya work with rtx 50 ? seems like doesn't work cuz of cuda

[bigZos]

I will try keeping this post up to date as much as possible with the latest developments.

To get your nvidia 50 series GPU working with ComfyUI you need a pytorch that has been built against cuda 12.8

In the next few months there will likely be a lot of performance improvements landing in pytorch for these GPUs so I recommend coming back to this page and updating frequently.

Windows

Windows users can download this standalone ComfyUI package with a cuda 12.8 torch build

Manual Install

At this moment Blackwell is not yet supported by stable pytorch.

pytorch nightly cu128 is available for Linux:

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

You can also use the Nvidia Pytorch Docker Container as an alternative which might give more performance.

Awesome! That Windows package got me rolling with image gen (FLUX and all) - works like a charm. But Wan's video generation isnt working yet. Any clue if that'll be sorted with those upcoming pytorch improvements? Hoping so!

[ndrco]

1. install last ComfyUI portable version with nvidia support fron here: https://github.com/comfyanonymous/ComfyUI/releases.
2. uninstal torchaudio: python_embeded\python.exe -m pip uninstall torchaudio
3. update torch and torchvision to actual nightly version from here: https://download.pytorch.org/whl/nightly/cu128: python_embeded\python.exe -m pip install ....
4. profit
   note that compatible builds of torch and torchvision may not have the same numbers. for example torchvision-0.22.0.dev20250301+cu128-cp312-cp312-win_amd64.whl, but torch-2.7.0.dev20250228+cu128-cp312-cp312-win_amd64.whl worked successfully on my computer. Good luck!

[rosso32]

Is there really no proper way to run ComfyUI at full speed on the RTX 50 series? The portable version doesn’t seem to perform well, and since I’m not a developer, I end up endlessly following links from NVIDIA, GitHub, and other sites, installing unnecessary things in a messy loop.

I just want to run it on Windows—not WSL2, not Docker—while installing CUDA 12.8, the corresponding nightly versions of Torch and Torchvision, and TensorRT to use the latest version of ComfyUI. But I can’t find any clear instructions on how to do this anywhere.

It’s been over a month since the 50 series was released…
There are just tons of reference links, but no actual guide on how to set it up properly to fully utilize the Blackwell architecture and Tensor cores on Windows.

Can someone please provide a proper setup guide and step-by-step instructions?

Related links:
🔗 #6643
🔗 https://github.com/comfyanonymous/ComfyUI#manual-install-windows-linux
🔗 https://github.com/comfyanonymous/ComfyUI
🔗 https://catalog.ngc.nvidia.com/orgs/nvidia/containers/pytorch
🔗 https://download.pytorch.org/whl/nightly/torch
🔗 https://download.pytorch.org/whl/nightly/torchvision

Every time I try to follow these links, I end up going in the wrong direction… I’m not familiar with Linux, so I’m really struggling. Please, help me… I never expected getting ComfyUI to work on the 50 series would be this difficult.

[fireYtail]

As I've mentioned, there were always several options. I've used both (portable installation and docker) and both worked just fine. Also, you are wrong. In this specific case, using Linux/WSL2 will provide improved performance, independent of running within docker. If you have enough resources, the fact that the code is running withing a virtual machine has near zero impact, at all. I know, I've tested both ways. Also, the Docker "developers" mentioned is an Nvidia backed image (pytorch:25.02-py3), not just some random developer made image. And if you understand docker and container security, having an image with almost nothing except the only thing that matters, with container security, non root access, read only filesystem, will provide as much security as possible: No ability to perform privilege escalation, no ability to perform living of the land attacks as there are no tools at all in the container, no ability to escape the virtual environment, and so on. But hey, suit yourself, each one does what it thinks its best for their own usage cases.

So, let's make an analogy. When you use Google Chrome and browse using a private tab, you are STILL entrusting Google with spying all your web activity. Using a private tab does not save you from the developer (Google). Running a virtual machine in isolation does not save you from the developer, the same analogy applies. Docker devs can just spy all your virtual machines. But you do you. You feel super safe because everything runs "isolated" in the same box. You just don't know (or care) if the box has backdoors or is an illusion of security.

[Azkrath]

Are you serious? I could really explain it to you as it is my day job but I am not going to bother. Comparing using a private tab in a browser with container security against malware threats is just bonkers.

[HDANILO]

It's like saying, oh, cheese is made of mold so we might as well jump into a poop pit and swim. 🤣🤣🤣

[fireYtail]

Are you serious? I could really explain it to you as it is my day job but I am not going to bother. Comparing using a private tab in a browser with container security against malware threats is just bonkers.

I will make a different analogy then, since you still don't understand it.

The founder of Telegram can give you all the security and encryption technologies in the world, but if they are enforced by law to hand over your "private" data, there is nothing stopping them from breaking their promises of privacy. They made the software, so they can always break the security, including obtaining all your encryption keys.

The same goes for virtual machines.

The only real safety is to permanently disconnect your device from the internet. Beyond that, all is nice promises, but words are just words.

You are free to live in a bubble of wonders where everyone has good intentions and everything is safe, but the real world doesn't quite work that way.

You are the kind of person that has tenths of passwords stored in some online password manager and you think you are super safe because the provider claims to encrypt your data, but you never think what would happen if the password manager got hacked and all your data leaked.

Or that free VPNs exist for charity and not because they sell all your internet activity data. You are safe because the connection to the VPN uses some basic encryption, but the VPN provider will NEVER dare sell your data to anyone. Right?

You just think that because someone claims something is safe, that automatically makes it immune to any danger. It's the postmodernist era we live in, everyone thinks magically, something typical of small children.

Until reality slaps your face, that is...

You need any more analogies? Because I already wasted enough time explaining super basic things and I'm not here to talk to toddlers. If you just want to live in a fantasy world, you do you. Just don't cry when you eventually find out the world wasn't what you thought it was. Sooner or later reality imposes itself.

[Azkrath]

Lol. You know nothing about cybersecurity. End of story. You can make all the shit you want. Also, you know nothing about me so all your “assumptions” are wrong. I’ve wasted enough time on this. You will not get another reply.

[HDANILO]

I know it has been mentioned here and there, but if you guys want a docker container that is already working in 5090, I've put this together:

https://github.com/HDANILO/comfyui-docker-blackwell

I got a 5090 4 days ago and has been struggling myself so I guess if I would spend the time making it work, better make it public for everyone to access...

I'm using it on windows, WSL 2, Ubuntu, Docker in Ubuntu, no performance impact this way. Download the models directly into the container and avoid volumes if you don't want slowdown.

[HDANILO]

Looks like you've got a conflict rising from numpy version, check if your numpy version is within <1.29.0

[mexyn]

I docker exec'ed into the container and checked numpy version. It's 2.1.2. But that's what i don't understand. I thought the purpose of this image is, that everything (python, packages, dependencies, etc.) are all set up to work with comfyui and RTX 50xx series or did i get that wrong?

[HDANILO]

You got it wrong, you still have to manage things like you would manage on your own machine, and error like this that would happen on you machine will still happens in docker.

This is happening likely because you have 2 conflicting custom nodes. One requires numpy to be a lower version and the other requires it to be higher, there's no way for the docker to know which numpy version is correctly, its just sequentially installing custom nodes requirements and ended up breaking one custom node instead of the other.

Btw this error has little to do with 50 series or blackwell.

[mexyn]

Gotcha. Thank you for clarifying

[adamreading]

Gotcha. Thank you for clarifying

Unfortunately these conflicts happen almost every time you load a big workflow with a bunch of new nodes. You get used to finding the best compromise. Pip list will be your best friend lol. Also, there’s a new blacklist file in comfyui where you can tell it to ignore any requested to update a list of modules you specify (mine has numpy, torch, torch vision etc ) so they don’t get downgraded or upgraded by installing nodes inside comfy manager itself.

[comfyanonymous]

I updated the post with a link to a standalone package with pytorch nightly 2.7 cu128

[bigZos]

Thank you!!!!

[alloystorm]

Just got my 5070ti.

I had to disable xformers using --disable-xformers, otherwise ksampler wouldn't run. The error message is saying flash attention not supporting the gpu yet. I'm not sure if disabling xformers is the best solution but since that error is from xformers code I gave it a try and it worked. I don't know if that's going to have a performance impact.

[pawnzhang]

I will try keeping this post up to date as much as possible with the latest developments.

To get your nvidia 50 series GPU working with ComfyUI you need a pytorch that has been built against cuda 12.8

In the next few months there will likely be a lot of performance improvements landing in pytorch for these GPUs so I recommend coming back to this page and updating frequently.

Windows

The recommended download is the standalone package with nightly pytorch 2.7 cu128 that you can download from here

Old package with torch 2.6

Manual Install

At this moment Blackwell is not yet supported by stable pytorch.

pytorch nightly cu128 is available for Windows and Linux:

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

You can also use the Nvidia Pytorch Docker Container as an alternative which might give more performance.

Yes.
Today I successfully ran ComfyUI on my NVIDIA RTX 5080 GPU with the following configuration:

Torch: 2.7.0.dev20250306+cu128 (file: torch-2.7.0.dev20250306+cu128-cp312-cp312-win_amd64.whl)
TorchVision: 0.22.0.dev20250307+cu128 (file: torchvision-0.22.0.dev20250307+cu128-cp312-cp312-win_amd64.whl)
TorchAudio: 2.6.0.dev20250306+cu128 (file: torchaudio-2.6.0.dev20250306+cu128-cp312-cp312-win_amd64.whl)
This resolved the previous issue where comfyui-manager failed to load.

Critical Notes:

Environment isolation is mandatory - always use ComfyUI's embedded Python (python_embeded) rather than system-wide Python installations.
Path priority must be explicitly set in startup scripts to avoid dependency conflicts.
CUDA 12.8 compatibility was verified through custom-built PyTorch binaries.

[HexColors60]

My version, it uses force-reinstall. (Skip the remove steps.)
update\update_5090.bat
pause
..\python_embeded\python.exe -s -m pip install --force-reinstall torch==2.7.0.dev20250307+cu128 torchvision==0.22.0.dev20250308+cu128 torchaudio==2.6.0.dev20250308+cu128 --index-url https://download.pytorch.org/whl/nightly/cu128 --extra-index-url https://download.pytorch.org/whl/nightly/cu128
pause

[jackeduardo]

this command is perfect, thanks bro

[reddwen]

nice! it's really helpful

[dimonnwc3]

any idea how to run flux in fp4 using onnx from here https://huggingface.co/black-forest-labs/FLUX.1-dev-onnx
or where to get safetensors fp4 checkpoint?

[edisonchan]

any idea how to run flux in fp4 using onnx from here https://huggingface.co/black-forest-labs/FLUX.1-dev-onnx or where to get safetensors fp4 checkpoint?

https://github.com/mit-han-lab/nunchaku

[shidarin]

Using @HDANILO's dockerfile I've been able to get most of comfyui working with my 5080, though sage attention is still broken, even after building from scratch. @kijai's WAN workflows also have a tendency of crashing with this container, even with sageattention turned off.

Here's my dockerfile so far, with comments:

# This image is based on the latest official PyTorch image, because it already contains CUDA, CuDNN, and PyTorch
FROM nvcr.io/nvidia/pytorch:25.02-py3
# The CUDA and PyTorch in this are both outdated.  The PyTorch version number Nvidia says is in this container (2.7.0a)
# cannot be trusted due to PyTorch specifying that version since early Jan, it'll need to be updated below.

# Installs Git, because ComfyUI and the ComfyUI Manager are installed by cloning their respective Git repositories
RUN apt update --assume-yes && \
    apt install --assume-yes \
        git \
        sudo

# Updates CUDA 12.8 to 12.8.1
# CUDA in the base container is outdated
RUN wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-ubuntu2204.pin
RUN mv cuda-ubuntu2204.pin /etc/apt/preferences.d/cuda-repository-pin-600
RUN wget https://developer.download.nvidia.com/compute/cuda/12.8.1/local_installers/cuda-repo-ubuntu2204-12-8-local_12.8.1-570.124.06-1_amd64.deb
RUN dpkg -i cuda-repo-ubuntu2204-12-8-local_12.8.1-570.124.06-1_amd64.deb
RUN cp /var/cuda-repo-ubuntu2204-12-8-local/cuda-*-keyring.gpg /usr/share/keyrings/
RUN apt-get update
RUN apt-get -y install cuda-toolkit-12-8

# Clones the ComfyUI repository and checks out the latest release
RUN git clone https://github.com/comfyanonymous/ComfyUI.git /opt/comfyui && \
    cd /opt/comfyui && \
    git checkout tags/v0.3.18

# Clones the ComfyUI Manager repository and checks out the latest release; ComfyUI Manager is an extension for ComfyUI that enables users to install
# custom nodes and download models directly from the ComfyUI interface; instead of installing it to "/opt/comfyui/custom_nodes/ComfyUI-Manager", which
# is the directory it is meant to be installed in, it is installed to its own directory; the entrypoint will symlink the directory to the correct
# location upon startup; the reason for this is that the ComfyUI Manager must be installed in the same directory that it installs custom nodes to, but
# this directory is mounted as a volume, so that the custom nodes are not installed inside of the container and are not lost when the container is
# removed; this way, the custom nodes are installed on the host machine
RUN git clone https://github.com/ltdrdata/ComfyUI-Manager.git /opt/comfyui-manager && \
    cd /opt/comfyui-manager && \
    git checkout tags/3.27.3

# Installs the required Python packages for both ComfyUI and the ComfyUI Manager
RUN pip install \
    --requirement /opt/comfyui/requirements.txt \
    --requirement /opt/comfyui-manager/requirements.txt

# Blackwell torch torchvision torchaudio
RUN pip install \
    --pre torch==2.7.0.dev20250308+cu128 torchvision torchaudio \
    --index-url https://download.pytorch.org/whl/nightly/cu128 \
    --force-reinstall
    # Use force-reinstall instead of --ignore-installed so that it fully uninstalls earlier versions of pytorch etc.

# Builds Triton from source
RUN git clone https://github.com/triton-lang/triton.git /opt/triton && \
    cd /opt/triton && \
    pip install ninja cmake wheel pybind11 # build-time dependencies && \
    pip install -e python

# Build SageAttention
RUN git clone https://github.com/thu-ml/SageAttention.git /opt/sageattention
    #cd /opt/sageattention
    #python setup.py install
# SageAttention won't build in the docker build environment- claims it can't figure out what GPU we have. It builds just fine
# when the container is finished, which is how I got the wheel below

# Copy prebuilt SageAttention2 wheel
COPY ./sageattention-2.1.1-cp312-cp312-linux_x86_64.whl /opt/sageattention/sageattention-2.1.1-cp312-cp312-linux_x86_64.whl
RUN pip install /opt/sageattention/sageattention-2.1.1-cp312-cp312-linux_x86_64.whl

# Build FlashAttention3 because why not
#RUN git clone https://github.com/Dao-AILab/flash-attention.git --recursive /opt/flashattention && \
#    cd /opt/flashattention && \
#    git checkout v2.7.4.post1 && \
#    cd /opt/flashattention/hopper && \
#    python setup.py install
# Broken for now

# Sets the working directory to the ComfyUI directory
WORKDIR /opt/comfyui

# Exposes the default port of ComfyUI (this is not actually exposing the port to the host machine, but it is good practice to include it as metadata,
# so that the user knows which port to publish)
EXPOSE 8188

# Adds the startup script to the container; the startup script will create all necessary directories in the models and custom nodes volumes that were
# mounted to the container and symlink the ComfyUI Manager to the correct directory; it will also create a user with the same UID and GID as the user
# that started the container, so that the files created by the container are owned by the user that started the container and not the root user
ADD entrypoint.sh /entrypoint.sh
ENTRYPOINT ["/bin/bash", "/entrypoint.sh"]

# On startup, ComfyUI is started at its default port; the IP address is changed from localhost to 0.0.0.0, because Docker is only forwarding traffic
# to the IP address it assigns to the container, which is unknown at build time; listening to 0.0.0.0 means that ComfyUI listens to all incoming
# traffic; the auto-launch feature is disabled, because we do not want (nor is it possible) to open a browser window in a Docker container
CMD ["python", "main.py", "--listen", "0.0.0.0", "--port", "8188", "--use-sage-attention", "--disable-auto-launch"]

This still results in the 'sm_120' is not a recognized processor for this target error, although we've now updated all components and built Triton and SageAttention from scratch. :/

[shidarin]

YES! I was looking for those wheels- someone mentioned using "the kijai built wheels" but didn't provide a link. Thanks

[HDANILO]

If you get the updated dockerfile working @shidarin, would you mind submitting a PR? this seems to be a good improvement on the overall experience of the docker

[edisonchan]

I did not set those env variables since I'm not building pytorch- I'm using the nightly wheels from https://download.pytorch.org/whl/nightly/cu128

Oh yeah actually I do have it working without building torch using the nightly... that error definitely only got fixed by building Triton from source, but now I'm unsure if the env vars have anything to do with it. This was few weeks ago and I did a rebuild to update few days ago. I've shared the first build wheels here, couple of people have used them successfully:

https://huggingface.co/Kijai/PrecompiledWheels/tree/main

This triton wheel would cause comfyui-easy-use and svdquant(https://github.com/mit-han-lab/nunchaku) can not be imported.

[adamreading]

also, about kijai, I'm running kijai myself with success, you do have to increase your ram limits on docker though, especially if there's a lot of offloading happening, containers crashes are 99% of the time related to hitting ram limits...

Currently I'm at 24gb ram for docker, and sometimes I can hit the limit with big input images on wan21

I spend most of my time in video gen with Wan/Hunyaun/ltx and speak with KJ most days. I had to up my wsl ram allowance to 70gb to stop random crashes, and employ every purge node and after run reset available for the high res workflows. I’m writing story boarding gem workflows with multiple iterations within a single flow. Often mixing models as well. ram use in comfy is difficult to control. The VRAM purges nicely but the Ram doesn’t. I end up sitting at 99% for ages whilst it swap files if I’m not careful. I do find the Native workflows have better Ram control.

[HDANILO]

I'm also on the storyboarding bandwagon! But yeah, I've decided to split things up in different workflows for this exact purposes, RAM doesn't purge well because of the "smart" cache from comfyui, but if you click on the top right vacuum clear icon, it showed purge the cache too