After upgrading to 5060ti I found out that prebuilt xformers for torch 2.7.0 doesn't support cuda, so I publish this .whl package for interested people.
For first version during building I included ONLY 12.0 compute capacity support, so first version supports BLACKWELL arch only.
Xformers is not actual after torch 2.7.0, since built-in optimizations offer same speed.
If you use higher python versions and cuda-13 consider using Sage-attention instead, it offers even bigger speed compared to xformers.
Xformers has small effect on computations' precision, this may slightly alter resulting image, keep this in mind.
Xformers at my setup offers around 50% speed boost compared to doggetx.
Installation:
open cmd in stable diffustion venv/scripts folder
type activate.bat
type pip uninstall xformers
put unpacked .whl file from zip archive to venv/scripts folder
type pip install xformers- and hit tab to automaticall fill filename (If you want to use this version with higher python version you can try adding --ignore-requires-python flag after filename).
launch webui-user.bat with --xformers argument
xFormers 0.0.30+4cf69f09.d20260302
memory_efficient_attention.ckF: unavailable
memory_efficient_attention.ckB: unavailable
memory_efficient_attention.ck_decoderF: unavailable
memory_efficient_attention.ck_splitKF: unavailable
memory_efficient_attention.cutlassF-pt: available
memory_efficient_attention.cutlassB-pt: available
[email protected]: available
[email protected]: available
[email protected]: unavailable
[email protected]: unavailable
memory_efficient_attention.triton_splitKF: available
indexing.scaled_index_addF: unavailable
indexing.scaled_index_addB: unavailable
indexing.index_select: unavailable
sp24.sparse24_sparsify_both_ways: available
sp24.sparse24_apply: available
sp24.sparse24_apply_dense_output: available
sp24._sparse24_gemm: available
[email protected]: available
[email protected]: available
swiglu.dual_gemm_silu: available
swiglu.gemm_fused_operand_sum: available
swiglu.fused.p.cpp: available
is_triton_available: False
pytorch.version: 2.7.0+cu128
pytorch.cuda: available
gpu.compute_capability: 12.0
gpu.name: NVIDIA GeForce RTX 5060 Ti
dcgm_profiler: unavailable
build.info: available
build.cuda_version: 1208
build.hip_version: None
build.python_version: 3.10.9
build.torch_version: 2.7.0+cu128
build.env.TORCH_CUDA_ARCH_LIST: 12.0
build.env.PYTORCH_ROCM_ARCH: None
build.env.XFORMERS_BUILD_TYPE: None
build.env.XFORMERS_ENABLE_DEBUG_ASSERTIONS: None
build.env.NVCC_FLAGS: None
build.env.XFORMERS_PACKAGE_FROM: None
build.nvcc_version: 12.8.61
source.privacy: open sourceDescription
Comments (17)
Any possibility to offer python 3.12 or 3.13 versions for rtx 30+ series?
what workaround do you use with python 3.12? i have python 3.10.9 installed and latest versions of kohya-ss and automatic1111 work fine. torch 2.7.0, which is compatible with latest rtx 50 series, works too
Does it work for RTX 50? I switched to a 5090 and Xformer no longer works; ControlNet Deep was giving an error.
yes, it does. i specifically compiled it for my rtx 5060 ti
if i know right 5070 is Blackwell Arch, so it's compute capacity is same as 5060ti, so yeah. i use latest automatic1111 build on torch 2.7.0, my best xformers is 0.0.30. i got 50% speed boost compared to doggettx and no opt. you can see showcase screenshot for confirmation.
@sevanidze Hi, do I have to put the xformers and xformers-0.0.30+4cf69f09.d20260302.dist-info folders into the venv script? Or do I have to put the contents of these two folders?
@OtakuStorm_Ai i described installation instructions in model description. put contents of archive in venv/scripts folder, open cmd in same folder, type activate.bat, type pip uninstall xformers, after that pip install xformers-0.0.30+4cf69f09.d20260302 . now launch your workflow and configure it to use xformers instead of other optimization packages
I wanted to try it, but I'm currently using Python 3.13 (and having compatibility issues).
i will try to build needed version on portable 3.13 python after today's night shift. please, specify your gpu here
@sevanidze rtx 5090
@WrackerRioter cuda and torch version?
xformers devs switch to torch internal opt method which are as good as xformers. they also disabled possibility to include prebuilt flash attention methond strating from 0.0.35 version, the only version compatible with newer python. nothing i can do here, sorry
I have Python 3.13 and RTX5090 and i'm using this:
xformers-0.0.33%2Bcu130torch2.9-cp39-abi3-win_amd64.whl
Cuda13.0 and Torch2.9.0
https://huggingface.co/Wildminder/AI-windows-whl/blob/main/xformers-0.0.33%2Bcu130torch2.9-cp39-abi3-win_amd64.whl
@orzechowy3334318 Does ControlNet Depth work?
@WrackerRioter With a little patch: https://github.com/Fannovel16/comfyui_controlnet_aux/issues/598#issuecomment-3608184221
@WrackerRioter Now i'm using DepthAnythingV3.