I don't think anyone has bypassing cloned subgraphs through.

I'm not lodging this as a bug, because I'm not sure exactly how the desired functionality could be implemented.

Pictures might be quicker here:

So image a horizontal stack of LoRA subgroups, all cloned. Each of them ingeniously adds the `low_noise` LoRA to the `high_noise` LoRA, and other things (removed from this example, because... complicated)

When they are bypassed using a handy Fast Group Bypass node so that only one is in operation.... things tend to go badly. Initially it bypasses all the nodes inside all the subgroups (which are just 1 cloned subgroup), and I would have to open one of them, and unbypass the nodes inside.

Then things mostly work (Kijai's LoRA processing system seems to honor the bypass state of the subgroup that holds the LoRA selection node -- I'm sure that's entirely accidentaly) but the other nodes run. That's why you can see a Regex and a PyList node outside the subgraph.

The root problem here is, of course, "how can ComfyUI cleanly bypass a subgroup?" Bypassing all the nodes inside is a very bad solution, even if they're not all clones.

The best I can come up with, is to write a node to get the bypass state of the subgraph from within the subgraph and use an "if then/else" node, to create manual bypasses. But that will only work so long as Comfy doesn't mess with the bypass state of the nodes inside too much (currently it seems to only try that once). How Comfy decides to handle this may change going forward, which is why I raise this as an issue here.

The report names Fast Group Bypass, cloned LoRA subgroups, and subgraph bypass state, but no source file or test. Start by reproducing the workflow shown in the attached diagrams and trace how bypass state propagates into cloned subgraphs. Done means subgroup bypassing behaves predictably without incorrectly changing unrelated nodes, with a regression test covering the workflow.

Written by the indexing model from the issue text.