Based on hunyuan video wrapper by kijai.
Update 1/1/25
This adds device selection nodes to hunyuan video wrapper.
I have included the workflow.
AND there is a download for the changed files if you want to just manually replace the 2 files in /custom_nodes/Comfyui-HunyuanVideoWrapper/
Or you can download the git from here.
https://github.com/MrReclusive/ComfyUI-HunyuanVWMultiGPU
This has be recoded from the last version to make this all optional, so I could get the pull request approved.
so this IS coming to base hunyuanvideowrapper.
I've modified the nodes to allow selection of device for those with multiple gpu's, this was necessary for me as my system has 2 4090's on a 1x pcie switch, so unloading and loading was slower then generating, with 2 4090's and the 4080 I don't have to unload anything during generation anymore.
Description
This has the 2 files needed just to update hunyuan video wrapper to work with this.
overwrites nodes and nodes_rf_inversion.
Will remove this once pull request is complete.
FAQ
Comments (12)
Hello, i have 2 gpu , 2 RTX 3090 and other
i put the two file and load your workflow without error.
difficult to change cuda device in cuda device selector
and
if i stay with default parameters , i have one error :
HyVideoVAELoader.loadmodel() got an unexpected keyword argument 'cuda_device' is:issue
sorry :
other question , lm studio or other loader can make load sharing on multi gpu with gguf llm , do you know if it can make it in futur on Hunyuan video, or in the roadmap of someone ?
not sure how its difficult to change, its a drop down list, my only guess is what "other" is in your system.. and for the error, its all optional so with nothing connected to the cuda input on the loader it should just be ignored.
I do know the hunyuan guys are working on a multigpu version, but where that is in terms of timing unsure, looks like image to video is next in the roadmap.
Any idea when you will merge with the latest commits? Looking for TeaCache. Thanks!
the version I have on github has teacache and dpm scheduler, I haven't looked today if anything changed, but every time I updated they would update again like a few hours later so I was waiting for that to chill out a bit before I did a full merge again.
I have 4 GPUs but cuda device selector shows only cuda:0 when I click on it. Is there some other configuration that needs to be done?
what gpu's do you have? and are any of them in nvlink?
@mrreclusive3545 4 RTX 6000 Ada. No Nvlink. Transformers and other python programs recognize them.
@softwareweaver glad you solved it, do you by chance have the capability of nvlink? i am curious how comfy and all this sees nvlinked devices, I've been thinking of picking up refurbished 3090's specifically for nvlink and AI video.
@mrreclusive3545 Unfortunately. I don't have access to any devices with NVLink. Best bet would be someone with couple of 3090s.
@mrreclusive3545 if you dont mind use linux, there is a hack driver can use pcie to do a "fake nvlink" p2p mode for multi-gpu, its seem work for llm but im not sure for sd
@omnisaa not sure that would be of any use in my setup, my 2 4090's share a 1x pcie slot(on 4 port pcie riser), which is the reason I had to create this in the first place was because the block swap in the wrapper was so slow because of that, it would actually run faster on the 4080 I have on the board as it has a full x16.
but now with hunyuan being fully integrated in comfy I no longer use this as the other multigpu addons work for it, and with teacache and all that i don't even use the fast models anymore.