Blackwell Support
Bug
Command
Disclaimer: I'm not expert on this area, it should be fixes with proper ways
I wanted to play with webui dev branch (probably change to forge instead, gonna try future) I can't build from pip and since its the efficient modules that can save memory while generating images, so I try to fix it to myself by building from sources
I created a environment with venv and torch installed, and here are the scripts
py -V:3.10 -m venv venv
call .\venv\Scripts\activate
# update pip to latest version
python -m pip install --upgrade pip
# build tools (setuptools < 82 due to others)
pip install setuptools<82 ninja
# tested torch-2.13.0.dev20260429
# --force-reinstall --no-deps
pip install --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/cu132
pip install -U "triton-windows<3.7"
# init for ninja + nvcc
SET MAX_JOBS=1
SET NVCC_THREADS=5
SET FLASH_ATTN_CUDA_ARCHS=75;120
SET CUDA_ARCH_LIST=7.5+PTX
SET TORCH_CUDA_ARCH_LIST=%CUDA_ARCH_LIST%
# flash-attention init
SET DISTUTILS_USE_SDK=1
SET FLASH_ATTENTION_FORCE_BUILD=TRUE
git clone --single-branch --depth 1 --filter=blob:none https://github.com/Dao-AILab/flash-attention.git src\flash-attention
cd src\flash-attention
git submodule update --init --recursive --depth 1 --filter=blob:none
# flash-attention 2
python setup.py install
# flash-attention 3 (not working, due to 128 alignment issue)
cd hopper
python setup.py install
# back to workdir
cd ../../../
git clone --single-branch -b v0.0.35 --depth 1 --filter=blob:none https://github.com/facebookresearch/xformers.git src\\\xformers
cd src\\\xformers
git submodule update --init --recursive --depth 1 --filter=blob:none
pip install --no-build-isolation -e .after successfully installed xformers package, run with python -m xformers.info shows that fa2/fa3/blackwell was unavailable
I do trying to install flash-attn by installing from source (hopper/fa3 I can't build with CUDA v13.1 for now, there is issues while building, finding a way to fix it) and now shows with
(venv) D:\AI\src\\\xformers>python -m xformers.info
xFormers 0.0.35+03b91d7.d20260430
memory_efficient_attention.ckF: unavailable
memory_efficient_attention.ckB: unavailable
memory_efficient_attention.ck_splitKF: unavailable
memory_efficient_attention.cutlassF-pt: available
memory_efficient_attention.cutlassB-pt: available
memory_efficient_attention.cutlassF-blackwell: unavailable
memory_efficient_attention.cutlassB-blackwell: unavailable
[email protected]: available
[email protected]: available
[email protected]: unavailable
[email protected]: unavailable
[email protected]: unavailable
memory_efficient_attention.triton_splitKF: availablethen I get back to webui, generating image shows that my GPU are too new
so I found out that at too new appears related to CUDA_MAXIMUM_COMPUTE_CAPABILITY
and it shows that at xformers/ops/fmha/cutlass.py and xformers/ops/fmha/flash3.py written CUDA_MAXIMUM_COMPUTE_CAPABILITY = (9, 0)
then I patched both files with the line to CUDA_MAXIMUM_COMPUTE_CAPABILITY = (12, 0), reinstall with pip install --no-build-isolation -e .
now the image can be generated normally
hope that there is patches coming up
referenced: https://github.com/facebookresearch/xformers/issues/1356#issuecomment-3602435421
Environment
PyTorch version: 2.13.0.dev20260429+cu132
Is debug build: False
CUDA used to build PyTorch: 13.2
ROCM used to build PyTorch: N/A
OS: Microsoft Windows 11 Pro (10.0.26200 64-bit)
Python version: 3.10.11 (tags/v3.10.11:7d4cc5a, Apr 5 2023, 00:38:17) [MSC v.1929 64 bit (AMD64)] (64-bit runtime)
Python platform: Windows-10-10.0.26200-SP0
Is CUDA available: True
CUDA runtime version: 13.2.78
CUDA_MODULE_LOADING set to:
GPU models and configuration: GPU 0: NVIDIA GeForce RTX 5070
Nvidia driver version: 596.21
cuDNN version: Could not collectSource: facebookresearch/xformers