Run Roop unleashed on Runpod

[SakoJutin]

Hello dears
I want to run Roop unleashed on Runpod
Can you explain the way please

[Swigler]

Hi,

You need to run the ubuntu:latest template, you can also try with the Ubuntu server 20.04 template

apt update && apt upgrade && apt install nano && apt install git

apt install build-essential && apt install g++-12

apt install nvidia-cuda-toolkit && apt install ffmpeg && export PATH=$PATH:$HOME/bin

wget https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh && bash Miniconda3-latest-Linux-x86_64.sh

Restart the pod.

conda create -n roop python=3.10 cudatoolkit=11.8 && conda activate roop

git clone https://github.com/C0untFloyd/roop-unleashed.git && cd roop-unleashed

-edit the main.py file in the ui folder, at line 78 replace share=roop.globals.CFG.server_share with share=True then at line 82 set the parameter to True, this is done to set gradio to generate the share link

nano ui/main.py

-install the app

pip install -r requirements.txt

Run the app

python run.py

-if you get any errors when starting run.py simpy restart the app, if the issues persist nano ui/main.py and make sure the code did not mush into a single line for what ever reason.

[Nevaehni]

To add to what Swigler mentioned:

Using ubuntu:20.04 won't work, as it fails at step apt install g++-12. To resolve this, you'll need to use ubuntu:22.04. Be aware that installing ubuntu:latest version will include a newer version of Python, which also leads to installation errors.

[sirdavidjcooper]

just to mention, I setup a tensorDock on Windows in just a few minutes :) then you can just RDP to it.

[Swigler]

Smooth :))

[MattStarfield]

I'm a bit of a n00b here. How do you guys deal with all of these dependencies not being retained across restarts in RunPod?

When I'm doing the steps:

wget https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh && bash Miniconda3-latest-Linux-x86_64.sh

Restart the pod.

conda create -n roop python=3.10 cudatoolkit=11.8 && conda activate roop

I immediately get the error that conda is not found after restart.

[lyranmahir4]

You should not restart the runpod pod, instead before you are about to run python run.py after going through all the steps, put conda deactivate conda activate roop and then run python run.py

[lyranmahir4]

Also on Step 3: Install Miniconda you can do this:

curl -LO https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh -b
eval "$('/root/miniconda3/bin/conda' shell.bash hook)"
conda activate
/root/miniconda3/bin/conda init bash
conda --version

[MattStarfield]

Thank you for this insight. I appreciate that Swigler provided those instructions (above) because I certainly wouldn't have known to take all of those steps (being a n00b and all), but it's frustrating that they explicitly say to 'Restart the Pod' when that will torpedo the entire process.

Also, given that Swigler was able to provide those detailed instructions in the first place, I would logically assume that they would be aware that restarting the pod would wipe out all the work that was performed in the preceding steps. I'm still not sure how that could've been overlooked. I mean, I've spent a grand total of about 2 hours working with RunPod and I learned this lesson immediately (seemingly confirmed by lyranmahir4).

In the end, I just got things up and running on TensorDock, like sirdavidjcooper siggested, instead of RunPod. TensorDock retains the state of the VMs between restarts, which seem infinitely easier than dealing with Pods getting wiped at every shutdown. Honestly, I don't know why/how people tolerate working with this constraint in RunPod when they can just use TensorDock and not have to worry about it. Seriously, why does RunPod seem to be everybody's go-to service for Cloud GPUs when TensorDock seems like a no-brainer in comparison?

I think the "proper" way to deal with this in RunPod would be to create my own Docker Image that has the container already setup with all of the changes listed in Swigler's instructions so those steps don't need to be performed every start up. Like I said, I'm very new to this and would need to invest some time into understanding if that is even the right move, and if so, how to do it. Using TensorDock avoids this whole rabbithole, so unless there's a good reason to get off of TensorDock, I'll probably just keep running with them.

[lyranmahir4]

I know, runpod had more docker that were made and curated by other users, but they slowly stopped making them, because that site is such a pain in the ass, not to mention they keep removing your pods, I actually learned the workarounds while installing facefusion, which is another faceswapper for video, it's good as roop, but installing that brings even more problems, and after one month of consistent trying i got it, only to find out another site called RunDiffusion has that built in.

[lyranmahir4]

A better guide:
Please note, in Runpod, restarting the pod results in losing all progress, so avoid restarting at any point during the setup.

Step 1: Run the Template

To begin, use one of the following templates:

• ubuntu:latest template
• Ubuntu server 20.04 template

Step 2: Update System and Install Essential Packages

First, update the system and install the necessary packages by running the following commands:

apt update && apt upgrade -y
apt install -y nano git build-essential g++-12 nvidia-cuda-toolkit ffmpeg
export PATH=$PATH:$HOME/bin

Step 3: Install Miniconda

Next, install Miniconda by executing these commands:

curl -LO https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh -b
eval "$('/root/miniconda3/bin/conda' shell.bash hook)"
conda activate
/root/miniconda3/bin/conda init bash
conda --version

Step 4: Create and Activate a Conda Environment

Create a new Conda environment named roop with Python 3.10 and CUDA toolkit 11.8 by running:

conda create -n roop python=3.10 cudatoolkit=11.8
conda activate roop

Step 5: Clone the Repository

Clone the repository for the program you need to run and navigate into its directory:

git clone https://github.com/C0untFloyd/roop-unleashed.git
cd roop-unleashed

Step 6: Edit main.py File

Navigate to the ui/main.py file and make the following modifications:

• At line 79, replace share=roop.globals.CFG.server_share with share=True.
• At line 82, set the parameter to True to enable Gradio's share link generation.

Step 7: Install Dependencies

Install the required dependencies by running:

pip install -r requirements.txt

Step 8: Run the App

To run the application, use the following commands:

conda deactivate
conda activate roop
python run.py

Step 9: Handling Errors

If you encounter any errors when starting run.py, try restarting the app using:

conda deactivate
conda activate roop
python run.py

If the problem persists, check the ui/main.py file to ensure the code hasn't merged into a single line unexpectedly.

[MattStarfield]

You are a saint. This thread desperately needs a more complete, cleaned up version of these instructions. Thank you!

What about Nevaehni's comment:

Using ubuntu:20.04 won't work, as it fails at step apt install g++-12. To resolve this, you'll need to use ubuntu:22.04. Be aware that installing ubuntu:latest version will include a newer version of Python, which also leads to installation errors.

Do your updated instructions address this? Or is this not really a problem?

[lyranmahir4]

Using ubuntu:20.04 won't work, as it fails at step apt install g++-12. To resolve this, you'll need to use ubuntu:22.04. Be aware that installing ubuntu:latest version will include a newer version of Python, which also leads to installation errors.

Yeah I am using that, didn't use other runpod instances with other ubuntu versions, The one I am using is runpod/pytorch:2.1.0-py3.10-cuda11.8.0-devel-ubuntu22.04 but this could be problematic as well, especially I can run the program fine, processed two videos, but suddenly I choose to use Codeformer and then it goes into out of memory errors, and I am using an RTX 4090 instance
RuntimeError: Error in execution: Non-zero status code returned while running Conv node. Name:'/blocks.21/conv_out/Conv' Status Message: /onnxruntime_src/onnxruntime/core/providers/cuda/cuda_call.cc:121 std::conditional_t<THRW, void, onnxruntime::common::Status> onnxruntime::CudaCall(ERRTYPE, const char*, const char*, ERRTYPE, const char*, const char*, int) [with ERRTYPE = cudaError; bool THRW = true; std::conditional_t<THRW, void, onnxruntime::common::Status> = void] /onnxruntime_src/onnxruntime/core/providers/cuda/cuda_call.cc:114 std::conditional_t<THRW, void, onnxruntime::common::Status> onnxruntime::CudaCall(ERRTYPE, const char*, const char*, ERRTYPE, const char*, const char*, int) [with ERRTYPE = cudaError; bool THRW = true; std::conditional_t<THRW, void, onnxruntime::common::Status> = void] CUDA failure 2: out of memory ; GPU=0 ; hostname=6fd294868569 ; file=/onnxruntime_src/onnxruntime/core/providers/cuda/cuda_allocator.cc ; line=47 ; expr=cudaMalloc((void**)&p, size);

[lyranmahir4]

We gotta try out other pytorch instances and check how well do they handle these problems

[lyranmahir4]

@MattStarfield