[Suggestion] Blackwell with cu128 xformer version

[RokuDoan]

After digging around a bit with my 5090 cu128-slim, it seem the following versions of torchs and xformers are working for me:

pip install torch==2.7.0 torchvision==0.22.0 torchaudio==2.7.0 xformers==0.0.30 --index-url https://download.pytorch.org/whl/cu128

Also with following flash-attn for most samplers to work:

pip install https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.4.11/flash_attn-2.8.3+cu128torch2.7-cp312-cp312-linux_x86_64.whl

All of my samplers seem to be working fine as well as PulID. Maybe more peeps can test these out but I think these should be the usable versions for all blackwell cards.

[YanWenKun]

Thank you very much for your advise!

I've added flash-attn binary from mjun0812 to the megapak images.

Although I can't just rollback PyTorch version in the images to 2.7, the newly released PyTorch 2.9.1 may have fixed the bugs for Blackwell GPUs.