ComfyUI with Cuda on Freebsd

NapoleonWils0n · Friday at 5:53 PM

Installing ComfyUI and ComfyUI-Manager with Cuda on Freebsd using Podman with the freebsd-cuda project.

ComfyUI

GitHub - Comfy-Org/ComfyUI: The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface.

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. - Comfy-Org/ComfyUI

github.com

ComfyUI-Manager

GitHub - Comfy-Org/ComfyUI-Manager: ComfyUI-Manager is an extension designed to enhance the usability of ComfyUI. It offers management functions to install, remove, disable, and enable various custom nodes of ComfyUI. Furthermore, this extension prov

ComfyUI-Manager is an extension designed to enhance the usability of ComfyUI. It offers management functions to install, remove, disable, and enable various custom nodes of ComfyUI. Furthermore, th...

github.com

Freebsd Cuda project

GitHub - NapoleonWils0n/freebsd-cuda: Run Linux command line and gui applications on FreeBSD with full Nvidia Cuda hardware acceleration.

Run Linux command line and gui applications on FreeBSD with full Nvidia Cuda hardware acceleration. - NapoleonWils0n/freebsd-cuda

github.com

Freebsd Cuda - ComfyUI

GitHub - NapoleonWils0n/freebsd-cuda: Run Linux command line and gui applications on FreeBSD with full Nvidia Cuda hardware acceleration.

Run Linux command line and gui applications on FreeBSD with full Nvidia Cuda hardware acceleration. - NapoleonWils0n/freebsd-cuda

github.com

ffmpeg-rust-scripts

GitHub - NapoleonWils0n/ffmpeg-rust-scripts: Statically compiled FFmpeg Rust scripts for automated video and audio processing

Statically compiled FFmpeg Rust scripts for automated video and audio processing - NapoleonWils0n/ffmpeg-rust-scripts

github.com

realesrgan-ncnn-vulkan

cerberus/real-esrgan-ncnn-vulkan/realesrgan-ncnn-vulkan-video-upscale.org at master · NapoleonWils0n/cerberus

cerberus code library. Contribute to NapoleonWils0n/cerberus development by creating an account on GitHub.

github.com

freebsd root config

GitHub - NapoleonWils0n/freebsd-root-xps: freebsd root dotfiles for dell xps 15 2019

freebsd root dotfiles for dell xps 15 2019. Contribute to NapoleonWils0n/freebsd-root-xps development by creating an account on GitHub.

github.com

freebsd-dotfiles

GitHub - NapoleonWils0n/freebsd-dotfiles-xps: freebsd dotfiles dell xps 15 2019

freebsd dotfiles dell xps 15 2019. Contribute to NapoleonWils0n/freebsd-dotfiles-xps development by creating an account on GitHub.

github.com

freebsd bin

GitHub - NapoleonWils0n/freebsd-bin-xps: freebsd-bin-xps

freebsd-bin-xps. Contribute to NapoleonWils0n/freebsd-bin-xps development by creating an account on GitHub.

github.com

pixaroma

pixaroma

Designing at the Speed of Innovation. I teach ComfyUI and test the latest AI models like Flux, Qwen, Wan, etc. This channel is about building better workflows for image and video generation. I share everything I learn so you can use these tools for your own creative work. My 5-hour ComfyUI...

www.youtube.com

cracauer@ · 2026-06-11T12:06:02+0100

Any idea why CUDA works in this context (podman) but not in a plain Linuxulator?

NapoleonWils0n · 2026-06-11T13:59:01+0100

When you try and pull down python libraries that depend on Cuda using the Linuxulator for example

Python can see your on Freebsd and so fails to pull down the wheels,
which are only built for Linux

Podman has options to specify the platform and os as linux
and because it a container this tricks libraries into thinking they are running on linux

And by mapping the Linuxulator libraries and using devfs rules to unhide the gpu from Freebsd into the Podman container
you dont need to install the nvidia driver in the container

So podman uses the freebsd kernel and nvidia driver with native performance
without the overhead you would get using virtualization

The other advantage is unlike bhyve you only need one gpu
you dont need to be rich and fork out for a second gpu to pass though to a bhyve vm

Also because you are using the wayland or x11 sockets you have hardware accelerated graphics
which isnt possible with a vnc connection

you would have buy another monitor to use with the byhve vm and pass that through to get hardware accelerated graphics

And the audio uses a pulseaudio socket so you get audio with no latency
and dont have to buy another audio device and pass it through to the vm

You also have dockerfiles and compose.yaml to automate building the podman images
which you can distribute, unlike a bhyve vm which has to be manually built

So in short

The container uses the nvidia driver and libraries from the freebsd host
so when you update the driver on the freebsd side it will continue working with podman container

You have native performance with no overhead and hardware accelerated graphics using wayland or x11 sockets
and audio with no latency using a pulseaudio socket

You can distribute podman containers and build them locally
which you cant do with a bhyve vm

The podman container uses your user id and guid from the freebsd host in the container
so all the files are owned by your user on freebsd

You can mount directories from freebsd into the container for file access
like mounting the downloads directory from freebsd with a firefox podman container

So when you download files using firefox running in a podman container
the downloaded files are in your downloads directory on freebsd and owned by your user on freebsd

Its free and requires no additional hardware

unlike bhyve where you would have to but another gpu, monitor and audio device
which is going to cost at least 500 pounds or so

There are some things that wont work with using the Linuxulator and Podman
because the kernel doesnt support the syscalls like python jax for example

But i have mamaged to get the following working

1) Firefox with hardware accelerated video and Widevine DRM support.

2) ffmpeg with hardware-accelerated nvenc encoding.

3) WhisperX audio transcription with Cuda acceleration.

GitHub - m-bain/whisperX: WhisperX: Automatic Speech Recognition with Word-level Timestamps (& Diarization)

WhisperX: Automatic Speech Recognition with Word-level Timestamps (& Diarization) - m-bain/whisperX

github.com

4) Ollama using Cuda.

GitHub - ollama/ollama: Get up and running with Kimi-K2.6, GLM-5.1, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models.

Get up and running with Kimi-K2.6, GLM-5.1, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models. - ollama/ollama

github.com

5) kokoro-onnx using Cuda.

GitHub - thewh1teagle/kokoro-onnx: TTS with kokoro and onnx runtime

TTS with kokoro and onnx runtime. Contribute to thewh1teagle/kokoro-onnx development by creating an account on GitHub.

github.com

6) ComfYU

GitHub - Comfy-Org/ComfyUI: The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface.

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. - Comfy-Org/ComfyUI

github.com

I did get a comment on one of the videos from Dr Owain Kenway

The Head of Research & Development (Platform Technologies) within ARC at UCL.

Saying

It makes FreeBSD a viable platform for ML stuff with PyTorch/JAX etc. If I can find time (and a free GPU which is probably harder) I'm tempted to have a go on our private cloud platform since I have FreeBSD working there.

However i did try python jax and couldnt get it work which maybe due to the fact i have an old turing nvidia card
or it could be an issue with the linuxulator syscalls