VSA cant work, to_gate_compress layers missing.
hello, this is from comfyui terminal:
[WARNING] VSA: the model has no to_gate_compress layers; running the fine stage without the coarse branch
it is --fallbacking-- to dense if the user using comfyui native "model sparse attention node" in vsa mode (use verbose mode if cant reproduce).
it is --failing-- if the user using "h3 vsa attention (fast h3)" custom node from Zironic H3-Optimizations.
The fallback is fine and it still works and speeds up with not much output difference. I'm adding the 'unstable' version that has them. The issue is on release they embedded them in the fl2va only, while it seems they actually would work fine or the distillation at least works fine to sample reference. Adding them also relies on the global timestep which on reference just kills reference and also hybridized stuff. So the UNSTABLE version is pretty much some kind of weird fl2va thing that looks and sounds good but reference can be inconsistent on larger sizes so you really need to run it very low res like 576, but it looks good for the small size.