Dependency Aware Node Caching for low RAM/VRAM machines #7509
Conversation
…of its decendents have executed. This allows users with lower RAM to run workflows they would otherwise not be able to run. The downside is that every workflow will fully run each time even if no nodes have changed.
|
Tested locally and appears to work as expected. Re-running the same workflow seems to cause a temporary rise in RAM before getting cleaned up, but that might be a quirk of python garbage collection. A 'proper' solution in the future would be to have a way to mark which nodes should not cache their results/display used RAM/VRAM of cache on each node, but the |
|
Can you fix the ruff check? |
There was a problem hiding this comment.
I figured out how to commit to the PR branch - ruff should now be passing @comfyanonymous
|
I did the commit wrong the first time, but here was the passing ruff run for it: https://github.com/comfyanonymous/ComfyUI/actions/runs/14392860510 Should be good to merge now |
Adds a new node caching option that will offload a cached node once all of it's children have executed in order to reduce the amount of RAM/VRAM a workflow needs. The downside is that a workflow will fully execute each time, so for an end user, it will seem like there is no cache at all.
This caching option is enabled using a new --cache-none flag (the name was chosen as for users, it seems like no caching occurs at all).
I created this pull request due to the resources that the new video models use. see: #7465