is support stable diffusion?

[libai-lab]

No description provided.

[blepping]

@libai-lab

if you use or can use ComfyUI, see #11 - i made an extension to use SageAttention there (and can be used with Stable Diffusion). doesn't currently work with SD15 models, only some SDXL attentions work (but there is a performance improvement). i didn't test with other models like Flux.

of course, this could also probably be ported to other frontends like A1111 but someone else would have to do it.

note: i'm just a random person, no affiliation with the SageAttention project.

[KartavyaBagga]

Hi @jt-zhang,
Will there be a support for SD15, head_dim 40 for hidden_states ?

[jason-huang03]

supporting head dim 40 is hard because int8 tensor core requires mma inner dimension of 32. In the case of hidden_states 40 we recommend use the fp16 baseline attention kernel.