When TeaCache4HunyuanVideo?

[shivshankar11]

https://github.com/LiewFeng/TeaCache/tree/main/TeaCache4HunyuanVideo

[jepjoo]

This works for me although not with Enhance-a-vid:
#226

[Fredd4e]

Works for me too, however it does not seem to work with the smaller fp8 model, and also as you say not with enhance a video.

[ctf05]

I found the enhance a video bug today. Im going to look into it sometime in the next two days. Other than that, I think it works fine. Tell me if there are any other bugs. It works on fp8 for me.

[Fredd4e]

I did some more testing, as i should have done beforehand.
Turns out its the auto cpu offload that does not work well for me with fp8 model. Apologies :)

[kroketagamer]

Hello, How do you do it? workflow?

[ctf05]

Hello, How do you do it? workflow?

Im not at my pc to give a workflow right now but its pretty easy. Pull the PR branch talking about teacache to your custom nodes. Teacache is not yet in the main repo branch. It adds a new node named TeaCache. Connect this to the video sampler. Values between .1- .2 seem best for a speed to quality ratio.

[ctf05]

I did some more testing, as i should have done beforehand.
Turns out its the auto cpu offload that does not work well for me with fp8 model. Apologies :)

All good. Glad it works. :)

[kroketagamer]

Hello, How do you do it? workflow?

Im not at my pc to give a workflow right now but its pretty easy. Pull the PR branch talking about teacache to your custom nodes. Teacache is not yet in the main repo branch. It adds a new node named TeaCache. Connect this to the video sampler. Values between .1- .2 seem best for a speed to quality ratio.

My Video Samplet node doesn't have the input for Teacache, I have the node but no connector on the video sampler

[Fredd4e]

I do have a question, I just dont seem to be able to get any decent results, I only have an GTX 1080ti and have been using SDPA since my old gpu dont support the faster thing. Does this work with SDPA or could that be why my results look like its been through no more than 2 iterations?

Any ideas?

[0000111100001111]

This works for me although not with Enhance-a-vid: #226

That PR is just giving me tensor location issues. Is it maybe not gonna be compatible with block swap?

Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument mat1 in method wrapper_CUDA_addmm)