linux or docker instructions about everydream2trainer

[chris-aeviator]

I‘m happy to share steps later for xf from a different docker file

…

Am 12.01.2023 um 22:01 schrieb Antonios Karamitros ***@***.***>:  could anyone attempt to make a docker image and if it works, share the Dockerfile here? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.

[chris-aeviator]

Here's what I use for xformers in docker, lmk if it needs something else, removed many lines specific to my setup. I'm unsure if ED2 needs bitsandbytes (8bit adam). You might also not need huggingface transformers if the repo doesnt rely on the diffusers lib.

FROM nvidia/cuda:11.7.1-cudnn8-devel-ubuntu20.04
LABEL maintainer="Tofu AI"

ENV DEBIAN_FRONTEND=noninteractive

RUN apt update && \
    apt install -y bash \
                   build-essential \
                   git \
                   git-lfs \
                   curl \
                   ca-certificates \
                   python3.8 \
                   python3-pip \
                   python3.8-venv && \
    rm -rf /var/lib/apt/lists

# make sure to use venv (this might not be needed)
RUN python3 -m venv /opt/venv
ENV PATH="/opt/venv/bin:$PATH"

RUN python3 -m pip install --no-cache-dir --upgrade pip && \
    python3 -m pip install --no-cache-dir \
        torch \
        torchvision \
        torchaudio \
        --extra-index-url https://download.pytorch.org/whl/cu117 && \
    python3 -m pip install --no-cache-dir \
        accelerate \
        datasets \
        hf-doc-builder \
        huggingface-hub \
        modelcards \
        numpy \
        scipy \
        tensorboard \
        git+https://github.com/huggingface/transformers.git@main


RUN pip install ninja bitsandbytes
RUN pip install -v -U git+https://github.com/facebookresearch/xformers.git@main#egg=xformers

## install ED2 from here
``

## EDIT:

this did not take into consideration that ED2trainer needs Python 3.10 - I'm currently setting up an ECS instance (linux) with ED2 - will be able to post a install log here 

[chris-aeviator]

things would be even easier with a requirements.txt file

[victorchall]

Yeah the dependencies are a bit tricky, particularly xformers, I need to get runpod going then I can revisit linux/docker.

[chris-aeviator]

xformers, bitsandbytes an accellerate should all install fine via any official pytorch image. xformers requires you to build (via a pip command) if you don't use the conda repos, but it usally takes only 2 minutes.

EDIT:
trick to avoid xformers build failing due to a misisng nvcc is to use the developers version of pytorch docker images

[AntouanK]

could anyone attempt to make a docker image and if it works, share the Dockerfile here?

[victorchall]

I use bits and bytes to support 8bit adam (significant performance improvement and ~1.5GB VRAM less) but that should be simple on linux, just install bitsandbytes 0.35.0, or newer versions may work as well on linux.

If bitsandbytes isn't installed, it will fall back to normal AdamW optimizer from torch, so its not strictly necessary. 8bit adamw is significantly faster though, and seems to be worth it.

[UglyStupidHonest]

I have it running fine with just installing the dependencies from I read in the bat file except the Windows stuff. The only tricky bit was that I had to update triton to get the xformers to work . (Ubuntu 20.4 lts)

[victorchall]

The windows_setup.cmd could be pretty easily stripped to requirements.txt, the patch_bnb is not required, and a different xformers install would be needed. That's most if it. Getting xformers to work in the biggest pain point.

[chris-aeviator]

what's the issue with xformers you face? It installed just fine with the official instructions every time I installed it, hayppy to assist

[victorchall]

Qslug has setup a container image here:

https://ghcr.io/qslug/ed2runpod

(should actually be agnostic to provider though, including a local docker instance)

Probably a few more hints here based on runpod/vast setup that could be useful for you:

https://github.com/victorchall/EveryDream2trainer/blob/main/doc/CLOUD_SETUP.md