Mount gpu in container
Hello, I'd like to use runpod to perform blender render tasks. I wish to use this docker image:
https://hub.docker.com/r/linuxserver/blender
The documentation say this about hardware accelaration: this container is capable of supporting accelerated rendering with /dev/dri mounted in
For Nvidia, it requires the installation of Nvidia container tools
linuxserver/blender - Docker Image | Docker Hub
This repository doesn't have a description
23 Replies
It won’t work. We can’t do full pass of device as it requires elevated permission and that brings security issues.
https://discord.com/channels/912829806415085598/948767517332107274/1183808597260521512
I remember that madiator had posted something about blender before and i think this was discussed by another community member
maybe an options to look into
https://discord.com/channels/912829806415085598/1015038737425633330/1183226103897264158
Oh i found it! someone had made a yt video and madiator made the repo / docker container to support the workflow ~
I mean it's gradio app but still has bugs but it works
Thank you for your help, I found the repo but there is no documentation with it ?
However, the yt video is very explicit
what repo?
Just to clarify when I shared the links it was just to share something i read as a lead to look into but not as a solution since it was just something i remember happening as a member of the server:
1) The yt video is a community member who was just showing something off
2) The github repo is just a fun project that was created to implement the project that was shown in the YT video separately from the original person
3) If u launch the runpod template it looks like it will start up a web interface.
You can play from there. This is not an official supported thing just a community idea from a while back that someone did for blender and I saw that was what u were interested in if u wanted a lead. But again, this is purely community interest and not official in any way.
I recommend just read through some the discord talk in the thread but this is just rlly ppl hacking things tgt - not something rlly flushed out like ur original linked project
Do you think I can create a custom dockerfile to perform the rendering serverless based on the linuxserver/blender dokcer image of my original post ?
( to avoid downloading blender each time)
My understanding is that setup_gpu.py detect the gpu itself and I don’t if it will works in serverless ?
If it sounds right to you and works as expected , I can open source the resulting dockerfile
Oh and I need to use the eevee engine, not cycles
Honestly, I am not too sure about how anything blender works sorry. Definitely not my speciality
No problem, I think Madiator is more capable about Blender 🙂
mayby I can try though it would be locked to single version
I won’t take your time, your theoric approve is enough for me, I’m starting it.
If it works, I can create a simple script to generate multiple image tags with different blender versions from the hardcoded array
I mean you can use my github as start point https://github.com/kodxana/gradio-Blender
GitHub
GitHub - kodxana/gradio-Blender: gradio Blender Render for RunPod
gradio Blender Render for RunPod. Contribute to kodxana/gradio-Blender development by creating an account on GitHub.
Yes I'm already on it of course 🙂
in the conversation of the creator of the original yt video, you mention GPU usage percentage. How can I find it to make sure it works at least a little bit ?
https://developer.nvidia.com/nvidia-system-management-interface
Nvidia SMI might be a good way to see it through CLI - other than the UI built into runpod website when you run a pod, it will tell you on the runpod website UI
@Madiator2011 I think you made a mistake at this line:
https://github.com/kodxana/gradio-Blender/blob/main/blender_gradio_app.py#L193
The variable after
-o should be output_file_path and not output_file_base_pathGitHub
gradio-Blender/blender_gradio_app.py at main · kodxana/gradio-Blender
gradio Blender Render for RunPod. Contribute to kodxana/gradio-Blender development by creating an account on GitHub.
fell free to sumbit PR
Oh yeah better way
nvm it looks fine
it saves frame to frame directory
Okay, then the variable is unused https://github.com/kodxana/gradio-Blender/pull/1
I managed to get something using cycles rendering
Keep going to eevee rendering and uploading
oh nice