WanVideoWrapper with SageAttention2.2.0 results in Noise Output

[wong00]

System: Linux
GPU: H800
CUDA: 12.4
Torch: 2.6
Triton: 3.2.0
SageAttention: 2.2.0

My ComfyUI workflow is like:

The Diffusion Model is: Wan2_2-S2V-14B_fp8_e4m3fn_scaled_KJ.safetensors

when I use SageAttention==1.0.6, the result is normal.

But when I use SageAttention==2.2.0, the result is noise:

I want know how to deal with this problem?

[wong00]

I run the test_sageattn.py and the result is below:

q (4, 32, 64, 128) cuda:0 torch.float16
sage vs math: mean_rtol=1 max_rtol=2 mean_atol=5.02 max_atol=5.02
The above (except max_rtol) should be < 0.05 (on RTX 20xx/30xx) or < 0.1 (on RTX 40xx/50xx)

I want know what's the problem with Pytorch or GPU

[jt-zhang]

Issue resolved! Please update to the latest code. Thank you!

[jokerz0624]

Issue resolved! Please update to the latest code. Thank you!

Hello, I have solved this bug caused by the mistake code of a commit which supports torch.compile.
However, when I go to write a PR, the code motioned below is reverted.
Is there any plan to revert the support of torch.compile again? and I'm grad to fix it