acheong08 / Diffusion-ColabUI

Choose your diffusion models and spin up a WebUI on Colab in one click
Apache License 2.0
421 stars 64 forks source link

Broken optimizations #9

Closed acheong08 closed 1 year ago

acheong08 commented 1 year ago

xformers and triton are both broken in Automatic1111's webui. I cannot fix this.

Can someone help?

GrennKren commented 1 year ago

Well.. I use xformers in kaggle and colab. Basically, installing in an inside virtual environment with Python 3.10

In Colab, you can choose between Pyenv and Conda for your virtual environment (download manually first).

In Kaggle , Conda was already built in.

Pyenv ``` #Download and installing ! curl https://pyenv.run | bash #Initialization required environtment path import os os.environ['PYENV_ROOT']=f"{os.environ['HOME']}/.pyenv" os.environ['PATH']+=f":{os.environ['PYENV_ROOT']}/bin" #setup virtual environtment and Python 3.10 ! sudo apt-get install -y libffi-dev ! pyenv install 3.10.7 ! pyenv local 3.10.7 ! python --version ! pyenv exec python --version #Download Install xformers from my compiled binaries !curl -L https://github.com/GrennKren/xformers_wheel/releases/download/v0.0.14_py310/xformers-0.0.14.dev0-cp310-cp310-linux_x86_64.whl --output "xformers-0.0.14.dev0-cp310-cp310-linux_x86_64.whl" !pyenv exec pip install xformers-0.0.14.dev0-cp310-cp310-linux_x86_64.whl #Launching webui ! COMMANDLINE_ARGS="--share --xformers" REQS_FILE="requirements.txt" pyenv exec python launch.py ```
Conda ``` #Download and install IF conda wasnt installed ! curl --output conda.sh -L https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh ! bash conda.sh -b # -b for silent install #Setup virtual environtment and python 3.10 ! conda create -y -n YouNameIt python=3.10 ! python --version ! conda run -n YouNameIt python --version #Install Xformers ! conda install -y -n YouNameIt xformers -c xformers/label/dev #Launching Webui ! COMMANDLINE_ARGS="--share --xformers" REQS_FILE="requirements.txt" conda run -n YouNameIt --no-capture-output python launch.py # --no-capture-output is needed because conda run was not produce any output until the python program completed. ```
PreCoi commented 1 year ago

用了ipynb文件后,造成Colab账号封禁。

GrennKren commented 1 year ago

用了ipynb文件后,造成Colab账号封禁。

It may have something to do with that

"model" . My two accounts, kaggle, were also recently locked. Right now I am trying again to make a new notebook without sharing the notebook in public. Edit:: still no sign of being locked when turning off to private
PreCoi commented 1 year ago

能解封吗?不知道能否申诉成功。

acheong08 commented 1 year ago

Hello. I apologize that this has caused the lockdown of your accounts. It may have something to do with the violation of certain Google policies:

This cannot be resolved by me.

acheong08 commented 1 year ago

@PreCoi It might take a few days for your account to be unlocked. In the meantime, I recommend using dummy accounts for Colab. It is also not recommended to keep the colab running for more than two hours at a time.

acheong08 commented 1 year ago

Triton has been fixed. Testing xformers

acheong08 commented 1 year ago

Triton has not been fixed. Error: https://github.com/openai/triton/issues/625

acheong08 commented 1 year ago

Upgrading to nightly triton... Testing...

acheong08 commented 1 year ago

More errors. Seems like default xformers is outdated. Installing newer xformers takes very long

acheong08 commented 1 year ago

Xformers has no improvements considering the time it takes to update the dependencies