Magcache Support #1877
Magcache Support #1877
Conversation
|
Looks interesting. I have two requests:
|
ai_diffusion/workflow.py
Outdated
| # Apply MagCache as a model patch if enabled | ||
| if checkpoint.magcache_enabled: | ||
| if arch in [Arch.flux, Arch.flux_k]: | ||
| print(f"Applying MagCache patch for {arch}") |
|
|
||
| model_type = "flux_kontext" if arch is Arch.flux_k else "flux" | ||
|
|
||
| try: |
There was a problem hiding this comment.
no need to wrap in try/except, this should not fail (or do so loudly if the code is broken)
ai_diffusion/ui/settings.py
Outdated
| else "" | ||
| ) | ||
|
|
||
| has_wave_speed = hasattr(client.features, 'wave_speed') and client.features.wave_speed |
There was a problem hiding this comment.
the hasattr is not needed
(see also below)
ai_diffusion/api.py
Outdated
| @@ -64,7 +63,12 @@ class CheckpointInput: | |||
| self_attention_guidance: bool = False | |||
| dynamic_caching: bool = False | |||
| tiled_vae: bool = False | |||
|
|
|||
| magcache_enabled: bool = False | |||
| magcache_thresh: float = 0.24 | |||
There was a problem hiding this comment.
Please remove the parameters also from here and settings.py.
They can be dug up from git if ever needed, but I'd rather not have them in the code without a good reason.
And if more configurability is desired at some point, a single "strength" value or some preset would be a more user friendly way to control it anyway.
…lear printing messages and other fix
|
Thanks for Acly's detailed review. I'll try to address these issues. I think we might keep a threshold setting to accommodate models that will use MagCache in the future, but clean up unnecessary code |
Added support for MagCache. It has been tested locally link (custom server) with ComfyUI. MagCache is commonly used in video models such as FramePack and Wan, and it also supports models like Flux and Flux Kontext. This support includes updates to nodes and adjustments to detailed parameters.
https://github.com/Zehong-Ma/ComfyUI-MagCache