LTX_Ripple_FFAF_Edit_v11.json error in the Patch Sage Attention KJ
Hello.
I am using ComfyUI Desktop and tried using LTX_Ripple_FFAF_Edit_v11.json. An error occurred during execution—"ModuleNotFoundError: No module named 'sageattention'"—within the "Patch Sage Attention KJ" node.
I realize that an additional custom node needs to be installed for it to work properly. I tried installing various custom nodes by searching for "sageattention" in the Node Manager, but the error persists.
Could you tell me exactly which custom node I should select in the Node Manager? Or do I need to download some custom node and manually copy it to <drive>:\Comfy-Desktop\ComfyUI-Installs\ComfyUI\ComfyUI\custom_nodes\?
You need sageattention installed. You can bypass those nodes if you don't have it installed.
I have updated the workflow in the workflow_examples folder, where I replaced sageattention with the comfykitchen attention backend.
Thank you very much for the reply.
I managed to install Triton + Sage Attention on ComfyUI Desktop, though it was a bit tricky—the folder containing python.exe in ComfyUI Desktop doesn't include all the necessary "lib" files; most of them are located in the .venv folder.
If anyone is interested, I could create a post with a step-by-step guide for ComfyUI Desktop installation.
Installing Triton + Sage Attention and using the "Patch Sage Attention KJ" node yielded excellent results: on my RTX 3060 Ti 8GB the time per step dropped from 9 seconds to 6.6 seconds.
Thanks for the v13 workflow; I incorporated the "Model Attention Backend" node from it and placing it after the "Patch Sage Attention KJ" node in my workflow, this helped stabilize it.
By the way, for those interested, you can reduce VRAM usage by replacing the diffusion and/or CLIP original models with the modules found at this link -
https://huggingface.co/Winnougan/ltx-2.5-w4a8-convrot-int4-convrot-Winnougan-Blessing