Closed wtomin closed 5 months ago
Good. The generated images by sd1.5+fa on 910b are also ok?
Good. The generated images by sd1.5+fa on 910b are also ok?
sdv1.5 does not use flashattention by default. Therefore my exps with sdv1.5 were just using vanilla attentions.
Because I found a better solution to this problem, this PR is closed to avoid confusion.
After the modifications, I have run
text_to_image.py
on both 910A(ms2.1) and 910B(ms2.2) with SDv1.5 and SDv2.0. All the experiments worked without errors, and the generated images are of good quality.However, the experiments only related to
CrossAttention
. I haven't tested withCrossFrameAttention
, because I'm not sure whenCrossFrameAttention
is applied.PS: I just found that SDv1.5 does not use flash-attention by default. Therefore the experiments with SDv1.5 were not meaningful.