SageAttention 3 on Z-IMAGE-TURBO and got a broken output image

[Demuzx]

Hello, team! I tested SageAttention 3 on Z-IMAGE-TURBO and got a broken output image. Also found out it doesn’t speed up inference at all.

[START] Security scan
[ComfyUI-Manager] Using uv as Python module for pip operations.
[DONE] Security scan

ComfyUI-Manager: installing dependencies done.

** ComfyUI startup time: 2025-12-11 21:14:37.911
** Platform: Linux
** Python version: 3.12.11 | packaged by conda-forge | (main, Jun 4 2025, 14:45:31) [GCC 13.3.0]
** Python executable: /venv/main/bin/python
** ComfyUI Path: /workspace/ComfyUI
** ComfyUI Base Folder Path: /workspace/ComfyUI
** User directory: /workspace/ComfyUI/user
** ComfyUI-Manager config path: /workspace/ComfyUI/user/default/ComfyUI-Manager/config.ini
** Log path: /workspace/ComfyUI/user/comfyui.log

Prestartup times for custom nodes:
0.0 seconds: /workspace/ComfyUI/custom_nodes/rgthree-comfy
0.8 seconds: /workspace/ComfyUI/custom_nodes/ComfyUI-Manager

Checkpoint files will always be loaded safely.
Total VRAM 15840 MB, total RAM 96522 MB
pytorch version: 2.8.0+cu129
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce RTX 5070 Ti : cudaMallocAsync
Using async weight offloading with 2 streams
Enabled pinned memory 91696.0
Using sage attention
Python version: 3.12.11 | packaged by conda-forge | (main, Jun 4 2025, 14:45:31) [GCC 13.3.0]
ComfyUI version: 0.4.0
ComfyUI frontend version: 1.33.13

But SageAttention 2.2 works fine. An 8-second inference drops to 6 seconds on a 5070 Ti when SageAttention 2.2 is enabled, and the image doesn’t get mangled into a mosaic.