AMD GPU + Windows + ComfyUI! How to get comfyUI running on windows with an AMD GPU! 

You are very welcome! I’m glad it worked and fixed your problems! Thank you for watching!

😂😂 yessss. I love it. I am hoping to get some stuff in my website before too long for people to either generate or post SD images. And I would absolutely love to see frogs in suits that you generated. Thank you so much for watching!

I second this.. many suited frogs are being made. Been trying to get this to work for ages and this by far seems to be the easiest set up. Thanks!@@FE-Engineer

You are very welcome! I’m really glad you got it working and are having fun with it! Still working on the website and another site. But will be getting the ability for folks to send up photos or images from AI sometime in the not too distant future!

Exactly what i was going to post!

Merry Christmas and happy holidays to you and your family as well. Thank you so much I am glad it helped! :)

You are welcome. Thank you so much for watching! I am waiting for ROCm on windows and then everyone can basically do everything with all of the different tools out there without really any compromises. Good speed and good support. One of these days….

You are very welcome thank you for watching!

You are very welcome! Sorry it took a while I’ve been debating about what and how much to include and do for comfyui. And I wanted to make sure it could work on windows without ROCm as a requirement.

I’m glad it helped! Thank you for watching and the kind words! Happy new years to you and your family as well!

You are very welcome! Thank you for watching!

It worked for me and was mostly straight forward :)

Thank you so much. Merry Christmas and happy holidays to you and your family as well! Spend time with your loved ones!

You are welcome! Thank you for watching and for the kind words. :)

also can run from bat file if conda is on path, i followed this tutorial: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-zFKD2Q9m_nQ.html

Yea. For some reason my conda was added to path but was being finicky. I ended up more recently largely ditching conda because for videos I end up installing stuff a lot and conda overall was becoming a bit more hurtful than helpful overall in my specific scenario. :-/

Thank you so much! :) I’m glad this helped you!

Glad it helped! Thank you for watching!

And DirectML is saying I've only got 1GB VRAM, but I'm also using a 7900XTX

@@ajphilippineexpat Same error

@@ajphilippineexpat same for a friend of mine as well. Anyone knows how to fix this?

Thank you so much! I am glad this helped. Comfyui is awesome but definitely has a learning curve. So I tried to keep it really straight forward as there is a ton of creators going in depth for comfyui.

Glad it worked! Heh recently someone had a comment saying this tutorial does not work anymore? So interesting that it did work without issue for you. Glad you got it up and running! :) thank you for watching!

same error

solved it with '' pip install -r requirements.txt' command

Hooray! Thanks for watching! :)

You are very welcome! I’m glad you liked it and found this helpful! :)

Yes. Sdxl does not work or at least I have never gotten it to work as of yet. The only place I have gotten sdxl to work is in Linux with ROCm

Rocm is basically AMD equivalent of cuda. It takes special cuda code and lets it work for an amd card.

You are very welcome! Thank you so much for watching and your support!

I have never used comfyui manager. What is it exactly?

You are very welcome! Merry Christmas and happy holidays to you and your family!

No it does not.

Auto1111 with zluda Or Sd.next with zluda will be the fastest most likely.

@@FE-Engineerthe auto1111 works for me, but sd.next gives me some issues

Might be a problem with torch and directml would be my guess.

wow, how? mine on comfy is much worse, only ~2it/sec while in A1111 i got up to ~14it/s on my 7800xt

You are welcome! That’s crazy, for me it is considerably slower. But I’m glad it is working well for you!

Code changed. I’m looking into it. If you roll back the git code it will work.

@@FE-Engineer ah, that explains my problems.

I have never seen that. What were you doing when you got that error?

Same issue, when it hits ksampler

Edit line 31: C:\Users\[USER]\miniconda3\envs\comfyui\Lib\site-packages\torchsde\_brownian\brownian_interval.py From: generator = torch.Generator(device).manual_seed(int(seed)) To: generator = torch.Generator().manual_seed(int(seed)) Restart comfyui from scratch

@@double.parker same here Error occurred when executing KSampler:

You are very welcome! I’m glad it helped!

In general I would assume so. But I have not tinkered with this enough to say for sure.

Lower your image resolution. Whatever you are doing is trying to use more VRAM than what your video card actually has.

Yikes. I apologize. I have not seen that error come up. Were you maxing out your vram? Try running it with something up to monitor your vram. My guess is you might have tried to generate an image that was too big for it to handle. This is a big guess though. Monitor system resources and redo and see how it looks. Might give you some clues. Also. Do something like barebones stock. Stock prompt. Stock model. Stock pipeline setup. Stock image dimensions. Make sure you can run it with everything being as controlled as possible to see if you are changing something somewhere that is causing this weird errors. That’s my suggestion.

Sounds like a python version issue, or potentially a permissions issue. I would make sure you have things setup appropriately and using the correct python version.

Sounds like you aren’t using directml torch to me?

Hey try this: python main.py --directml Otherwise I had the same error! Comfy tried to call torch.cuda.current_device(), and it could not of course: [...] File "G:\AI\ComfyUI\comfy\model_management.py", line 83, in get_torch_device return torch.device(torch.cuda.current_device()) File "G:\AI\VirtualEnvs\win_comfy\lib\site-packages\torch\cuda\__init__.py", line 674, in current_device _lazy_init() File "G:\AI\VirtualEnvs\win_comfy\lib\site-packages\torch\cuda\__init__.py", line 239, in _lazy_init raise AssertionError("Torch not compiled with CUDA enabled") AssertionError: Torch not compiled with CUDA enabled

Automatic1111 directML with ONNX might be faster. It usually is faster. But it comes with a lot of drawbacks. You have to manually optimize models. No SDXL. And most extensions like control net do not work. :-/

Hi, i do not know how you can Install ComfyUI in a other Path but you can ReDirect the model folders used by ComfyUI with editing the "extra_model_paths.yaml.example" in the main Directory.

According to the instructions, Comfyui has been installed on drive C. You can check drive C and it will be there. The checkpoint must be saved on drive C.

I apologize. I have no idea. This sounds complicated and prone to produce errors as one repo could update and require an updated package where another one might not or might break. Interesting idea though. I might play around with it and see if I can get it up and running…but again. Sounds prone to break and cause issues long term.

An other solution is to use symlink to link the venv folder to the one from A1111. Personally I use the symlink for everything that is shareable environment model ect.. So these folder are just link that point out to the real one somewhere else. I also use that to split the AI models in two category the one I use often are on the SSD while the other are on a HDD. Then they are linked back to A1111 and Comfy model folder.

Hi, this is the "Enviroment" Path you have to go into ComfyUI (C:User/Name/ComfyUI)

Torch settings: Using directml with device: Total VRAM 1024 MB, total RAM 32695 MB pytorch version: 2.3.1+cpu Set vram state to: NORMAL_VRAM

I would be skeptical about getting this to run. (for reference, that MSI looks like it is a monitor, not a graphics card). It might work, but even if it does, I imagine it would be incredibly slow. If you wanted to go AMD, right now in general the 7800xt looks like the absolute best bang for the buck using newer cards, especially if you find one potentially on sale a little. On the flipside, if you really wanted to do a lot of stable diffusion and AI, a used nvidia 3080 would be a really nice choice if you can get one for a good price. 3080 12G model would be even better. Please keep this in mind though, these are extremely vague recommendations, you should absolutely spend time thinking about how much you want to realistically spend, and what you want the graphics card to do and be used for, as those things may significantly impact what the best choice is for you particularly. I would mostly avoid RDNA2 cards as they are not nearly as well equipped as RDNA3 (rx7000 cards) for AI workloads. I also personally struggle to really recommend new nvidia cards as they are very expensive. But even with that, new Nvidia Super cards are coming out literally next month, and the prices look to be better than before in general, so some price shifting may be occurring with Nvidia 4000 series cards here VERY soon!

@@FE-Engineer Thanks a lot for your detailed reply.

No. I don’t use comfyui too much honestly. I find the spaghetti…messy.

Delete your existing directory and everything in it. You already have one cloned.

Try using the medvram flag. See if that helps. Although if you are running dual video cards you might run into weird issues. I have never tried with dual video cards and unfortunately in that type of instance I don’t know how much help I can honestly provide :-/

@@FE-Engineer Hi, I'm totally new to this thing and I ran into this low vram thing too. Everything installs fine, the UI is on but it says I have only 1GB and I can't render even a tiny (128x128) image. I have an RX570 with 4GB of Vram and 32GB of RAM on MB, that should be enough. Where and how should I use that "medvram" flag? Thanks.

Is it when you try to load a model?

So notes: SDXL will not work to my knowledge. Shark from nod . Ai is easy but is miserable to use. SD . Next is pretty straight forward. But is somewhat limited overall. I think this might be the best overall right now. Automatic 1111 directML is good but converting to ONNX is irritating. You give up samplers in ONNX but get a big speed boost. Comfyui is comfy ui. Complicated to use. DirectML will net you about ~6 it/s and it’s a little finicky. I think that’s the most common current windows run down.

Ah thanks on the advice. Been trying shark to little success. Been having 1111 DirectML generating a few images but I am 90% sure it is actually using the CPU to do the generations instead of the GPU... practically zero GPU activity and my CPU stats and fan speed is high. So try and get SD Next working instead sounds best? What about under Linux? 1111 with the ROCM it can just work out of the gate, right? @@FE-Engineer

@@Knightedskull On Linux usually everything work. SDXL control net lora, dreambooth training, LLM, Wishper, Coqui TTS . Also the performance is usually pretty good. It get about 18It/s on 512px image, up to 23Its if I turn off medvram. With LLM It does over 60T/s with 7Bq4 model, on a 7900xt. It can also run everything at once while playing a game and recording at same time. I made a video playing Minecraft with shader while trying to follow the instruction from an AI that have a voice and 3D Vtuber avatar. I managed to get everything to run at once on a 7900xt. Though seriously a XTX would be better because It was tight on the Vram side.

@@Knightedskull u have used the --use-directml on the commandline?

yeah i believe so but i've switched back to linux @@ubuu90

Activate Conda environment. Run the start script. You can run conda list to see all conda environments that you have. Maybe that will help?

For me. On windows. Using directml and NOT using ROCm (like I use in Linux). It is about 40% of my normal it/s speed under similar circumstances. So it is a big decrease in performance. But. It works with everything I think. It is not limited to ONNX, and it should support all the fun stuff like inpainting controlnet etc as best I can tell. I have not played with those things on it in windows though so I will have to actually test to be sure. If someone absolutely wants all functionality and refuses to go to Linux. This is probably the best bet right now.

@@FE-Engineer i dont care speed in windows rather i need stability. i always run into vram issues on windows when using fork of a1111. only thing i want is seamless experience hence i have dual boot ubuntu

Might be the version of python you are using. But code has been updated since that video. Check the video description.

De-Install Torch and Re-Install it again that will fix it... I had same Problem ✌(Un-Install = pip uninstall torch) Then use the Line in the Video to Re-Install

:):) happy to help! Thanks for watching!

Yea. Looks like I need to make a new video it seems

@@FE-Engineer can you show which version you use? so it's possible to reproduce the exact things. i only have 7900 xt

I think that even now with all the progress that is being made constantly. Not having ROCm on windows. And using directml. I think it really isn’t compatible with everything that is available when using cuda yet. With that being said. I think PyTorch with rocm is good. And Rocm on windows I think is very close. But until we get there I think you will just find a lot of little things like this that are virtually unsolvable. :-/. Just library incompatibilities and things missing that do odd things.

You are very welcome! Thank you so much for watching!

might want to try the low or med vram options?

@@FE-Engineer im pretty sure it says somewhere but how do i do it?

Strange Path... DeInstall Torch and proper Install it into Pyton

So I have updated my GPU and installed a new model. If I load comfyui up fresh from the mini conda terminal I can generate an image and it works. However, if I try to swap to a different model in comfyui I will get a blank image again.

It might be the model. Some get a bit weird. Try the standard 1.5 model and make sure that works reliably. Some of it is a bit of trial and error. Some need a vae. Some need clip skip. Unfortunately it’s hard to say why this is happening exactly.

@@FE-Engineer thanks, i tried a different model and it worked :)

I think you can actually reconfigure the models directory to be a different place. And that would be a better overall way to do it versus trying to setup something like sym links that could end up with weird side effects.

Yes I use symlink for everything that is shared and also put the model I use less often on a HDD but still have them to show up in the model folders. .

@@leucome nice, good to know

have you found a solution for this problem ? i'm actually in the same situation

@@mizuhahato6215 I dont use Comfy anymore, went back to A1111 directml fork, with Lobe theme, and dont have as many issues as before, ofcourse sometimes, depending on prompt length I get net enough VRAM error, but it works.

@@ZeroNyte ok ok, i just found that if you launch comfy with --lowvram, it works fine

It’s ok. It is a misprint. Mine says that as well.

Probably a python version problem. That’s usually the cause of that error.

Hi @@FE-Engineer - appreciate the video and all the help! I am a complete noob at this so i hope that i can ask how to then change my python version?

For windows or Linux?

@@FE-Engineer windows

Works like a Charm... I'm just done it myself and im rendering Pictures while i write this.