
Krea 2 / Krea 2 Turbo SageAttention guard patch for ComfyUI-KJNodes
I put together a small patch package for ComfyUI-KJNodes that makes the `Patch Sage Attention KJ` node safer to use with local Krea 2 / Krea 2 Turbo workflows.
This is not a standalone custom node. It patches the existing KJNodes SageAttention node.
What it does:
- detects Krea 2 model patchers
- only applies SageAttention to allowlisted diffusion attention paths
- skips/falls back for unsupported shapes, masks, dtypes, head counts, or SageAttention failures
- avoids touching likely text-fusion/Qwen attention paths
- adds logging and optional dry-run validation
- includes sample Turbo and RAW workflows
Repo:
https://github.com/SurrealByDesign/comfyui-krea2-sageattention-guard
I also opened an upstream PR to ComfyUI-KJNodes so hopefully this can be incorporated directly.
Tested on my setup with:
- Krea 2 Turbo FP8
- Krea 2 RAW
- qwen3vl_4b_fp8_scaled text encoder
- CLIPLoader type `krea2`
- sageattention==1.0.6
Please back up your KJNodes file before patching. Feedback and testing reports are welcome, especially on different KJNodes/SageAttention versions.