search

lllyasviel / stable-diffusion-webui-forge

GNU Affero General Public License v3.0
8.64k stars 856 forks source link

Latest update is causing GPU degradation #1701

Open GigaByteWide opened 2 months ago

GigaByteWide commented 2 months ago

I never had this issue with the previous updates, I am using both --formers & --Medvram (even tested with --lowvram)

Currently using RTX 2060 Super with 16gb of ram.

Previous updates, I was able to finish generation within 1 minute to 14 minutes.

The latest update is now at 47 minutes to 92 minutes per generation.

The CMD window spams me with GPU degradation warnings no matter which settings I use. I've tried asking people on YouTube what I can do to resolve this issue.

They all just gave me ignorant responses like "Lol git gud, get yourself an rtx 4090 or just give up."

DenOfEquity commented 2 months ago

start here: https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/1181, most likely answer is that you've set GPU Weights too high. Don't use command-line arguments unless you know you need them (needing them with old Forge or A1111 does not mean needing them now); I run on 8GB VRAM with no such settings needed.

skibidiskoobidi commented 2 months ago

Just joined to say. Same issue here since the tiny update this morning (9/5/2024). Was working fabulous last night. I am a n00b and don't change anything from default other than image size. As suggested, I have tried reducing the gpu weight. This seems to be a temporary fix. Meaning that, it will generate some images without error until I restart forge. Upon restarting I get the error again and need to reduce the weight even lower compared to the previous run to clear the warning. This is also has affected hires fix upscaling. It makes it slow to a crawl, 2 to 3 mins per iteration. The system is Windows 10. Intel based 64gb ram. 4070 16gb Ti Super. Samsung NVME 2gb 75% free space.

skibidiskoobidi commented 2 months ago

Replying to myself. The recent push (mid afternoon EST 9/5/2024) seems to have resolved the issue.

Just joined to say. Same issue here since the tiny update this morning (9/5/2024). Was working fabulous last night. I am a n00b and don't change anything from default other than image size. As suggested, I have tried reducing the gpu weight. This seems to be a temporary fix. Meaning that, it will generate some images without error until I restart forge. Upon restarting I get the error again and need to reduce the weight even lower compared to the previous run to clear the warning. This is also has affected hires fix upscaling. It makes it slow to a crawl, 2 to 3 mins per iteration. The system is Windows 10. Intel based 64gb ram. 4070 16gb Ti Super. Samsung NVME 2gb 75% free space.

dramatticdev commented 2 months ago

This has happened to me before. Try clearing your virtual ram, sometimes cached files mess up and conflict with updated ones from my experience. The best way I found to completely clear virtual ram is to remove the paging file in advanced system settings under the virtual ram tab then restarting your pc. The reactivate virtual ram and set the size to what you can. I do not have a good pc at all but I have mine set to 40gb on a ssd, running a 5th gen i5 with 8gb physical Ram and a rtx 3060 and I am getting flux generations in under 4 minutes with 30 steps ,1152x896, gpu weights are set to 10 000MB

LIQUIDMIND111 commented 2 months ago

Replying to myself. The recent push (mid afternoon EST 9/5/2024) seems to have resolved the issue.

Just joined to say. Same issue here since the tiny update this morning (9/5/2024). Was working fabulous last night. I am a n00b and don't change anything from default other than image size. As suggested, I have tried reducing the gpu weight. This seems to be a temporary fix. Meaning that, it will generate some images without error until I restart forge. Upon restarting I get the error again and need to reduce the weight even lower compared to the previous run to clear the warning. This is also has affected hires fix upscaling. It makes it slow to a crawl, 2 to 3 mins per iteration. The system is Windows 10. Intel based 64gb ram. 4070 16gb Ti Super. Samsung NVME 2gb 75% free space.

HELLO, IS IT WORKING NOW? STILL ISSUES?

Snailenjoyer commented 2 months ago

Hey brother, I feel your pain, I believe this issue you are experiencing is caused by the update of your torch and xformers that the UI was requesting so much. Now this is and issue that was hell for me to fix but I will give you a guide here and pray to god that you manage to fix it. Steps are as follows:

  1. Go to your stable diffusion folder and delete the entire venv folder.
  2. Go to edit your webui-user.bat and remove xformers from it so it does not install the latest version that gimps gen speed
  3. Launch your webui-user.bat so that it creates the venv folder again and after the full launch close the UI
  4. Now click on the bar on top with file path and type "cmd"
  5. Once a cmd boots up type the following (line by line hitting enter after each line): cd ./venv/scripts activate pip install torch==2.3.0 torchvision==0.18.0 torchaudio==2.3.0 --index-url https://download.pytorch.org/whl/cu118 ~After the installation finishes dont close out the cmd just yet.
  6. Go to this file path: stable-diffusion-webui-forge>venv>Scripts
  7. Scroll through the files in this folder and you will see 3 files all starting with the name pip. One of them should be called pip(some number.some number eg. 3.10) this is your pip version
  8. Go to the link corresponding with your pip version: PIP3.10: https://download.pytorch.org/whl/cu118/xformers-0.0.26.post1%2Bcu118-cp310-cp310-win_amd64.whl#sha256=a8b401d4e5f3845bd277401506f62250702dedfc5cd7944e87d0bc9c30761281 PIP 3.11: https://download.pytorch.org/whl/cu118/xformers-0.0.26.post1%2Bcu118-cp311-cp311-win_amd64.whl#sha256=2fd882d2dca2e5e814aede20819804041496c938fb395f00c511424ef76dd86e PIP 3.8: https://download.pytorch.org/whl/cu118/xformers-0.0.26.post1%2Bcu118-cp38-cp38-win_amd64.whl#sha256=70d72d7b61dc02463b24140c637ad3dc7281e88d90c319b26d9a2ae2f49fe1ca PIP 3.9: https://download.pytorch.org/whl/cu118/xformers-0.0.26.post1%2Bcu118-cp39-cp39-win_amd64.whl#sha256=2131fef2466cc4b3464113679aa6f701a7bc292bc623b08197fe4001369f8c40
  9. Once the corresponding .whl file downloads place it in your Scripts folder (stable-diffusion-webui-forge>venv>Scripts) and copy its full name
  10. Go back to the cmd prompt and type: pip install full_name_of_the_xformers_file.whl
  11. Now you can close out the cmd prompt and add -xformers back to webui-user.bat which should resolve your slow gen issue completely

I hope I managed to help, I am assuming your issue as it seems similar but in case you didnt update your xformers and pytorch i have no clue.

skibidiskoobidi commented 2 months ago

weird, my venv folder is in \AppData\Local\Programs\Python\Python310\Lib\venv it appears that I have pip3.10 in this folder.

i don't have a venv in my forge or stable diffusion folder at all.

Snailenjoyer commented 2 months ago

That is indeed weird, forge should create a venv folder in the main folder after the first boot. Try launching forge and see if it creates it, if not go to the folder path click it, type cmd and in cmd type "git pull". Maybe that creattes a venv folder for you.