Triton for xformers on Windows?

[Silvicultor]

Hello everyone!

I would like to ask if it is possible now to use xformers+triton on Windows with forge? (since Triton is now available for Windows)

Now I know that xformers alone doesn’t really improve gen speed with forge. But I suppose with Triton it could?

So the question would be is there a Triton version for Windows that could be used with forge and especially the 2.3.1 version of pytorch?
And if yes, could I just install it into my venv and would xformers then recognize and use Triton, or would further steps be necessary?

[SvenErik1968]

You can install the latest version of Triton for Windows in the venv, and Triton will see that it is installed. Not sure if it helps with anything though... 😉

venv\Scripts\activate.bat
pip install triton-windows
Using cached triton_windows-3.2.0.post17-cp310-cp310-win_amd64.whl.metadata (1.5 kB)
Using cached triton_windows-3.2.0.post17-cp310-cp310-win_amd64.whl (40.0 MB)
Installing collected packages: triton-windows
Successfully installed triton-windows-3.2.0.post17

Then you can run the following command while in the venv to check:
python -m xformers.info
and you should get something like this:

xFormers 0.0.27
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:               available
memory_efficient_attention.cutlassB:               available
memory_efficient_attention.decoderF:               available
[email protected]:        available
[email protected]:        available
memory_efficient_attention.smallkF:                available
memory_efficient_attention.smallkB:                available
memory_efficient_attention.triton_splitKF:         available
indexing.scaled_index_addF:                        available
indexing.scaled_index_addB:                        available
indexing.index_select:                             available
sequence_parallel_fused.write_values:              available
sequence_parallel_fused.wait_values:               available
sequence_parallel_fused.cuda_memset_32b_async:     available
sp24.sparse24_sparsify_both_ways:                  available
sp24.sparse24_apply:                               available
sp24.sparse24_apply_dense_output:                  available
sp24._sparse24_gemm:                               available
[email protected]:                        available
swiglu.dual_gemm_silu:                             available
swiglu.gemm_fused_operand_sum:                     available
swiglu.fused.p.cpp:                                available
is_triton_available:                               True
pytorch.version:                                   2.3.1+cu121
pytorch.cuda:                                      available
gpu.compute_capability:                            8.9
gpu.name:                                          NVIDIA GeForce RTX 4070 Laptop GPU
dcgm_profiler:                                     unavailable
build.info:                                        available
build.cuda_version:                                1201
build.hip_version:                                 None
build.python_version:                              3.10.11
build.torch_version:                               2.3.1+cu121
build.env.TORCH_CUDA_ARCH_LIST:                    6.0+PTX 7.0 7.5 8.0+PTX
build.env.PYTORCH_ROCM_ARCH:                       None
build.env.XFORMERS_BUILD_TYPE:                     Release
build.env.XFORMERS_ENABLE_DEBUG_ASSERTIONS:        None
build.env.NVCC_FLAGS:                              -allow-unsupported-compiler
build.env.XFORMERS_PACKAGE_FROM:                   wheel-v0.0.27
build.nvcc_version:                                12.1.66
source.privacy:                                    open source

See the is_triton_available: True status where xformers confirms that Triton is available.

[Silvicultor]

Thanks for the quick response! I was able to install Triton into the venv.
But xformers does not recognize it.
On startup it still says:
ModuleNotFoundError: No module named 'triton'

But Triton was definitely installed into the venv. If I try to re-install it says “requirement already satisfied”.

Maybe the problem is related to this warning massage that xformers.info gives me?

WARNING[XFORMERS]: xFormers can't load C++/CUDA extensions. xFormers was built for:
    PyTorch 2.4.0+cu121 with CUDA 1201 (you have 2.4.1+cpu)
    Python  3.10.11 (you have 3.10.6)
  Please reinstall xformers (see https://github.com/facebookresearch/xformers#installing-xformers)

Already tried to re-install xformers as the warning message suggests, but it gives me a dependency conflict:

ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
torchvision 0.18.1+cu121 requires torch==2.3.1+cu121, but you have torch 2.4.0+cu121 which is incompatible.
Successfully installed torch-2.4.0+cu121 xformers-0.0.27.post2

Don’t understand why this is happening. I have installed the recommended version of forge with pytorch=2.3.1 but here it says I have 2.4.0? Any ideas how to resolve this?

[SvenErik1968]

That first error message is similar to what I get if I try to run python -m xformers.info outside of the venv.

C:\Server\Docker\stable-diffusion-webui-forge>python -m xformers.info
WARNING[XFORMERS]: xFormers can't load C++/CUDA extensions. xFormers was built for:
    PyTorch 2.6.0+cu124 with CUDA 1204 (you have 2.6.0+cpu)
    Python  3.10.11 (you have 3.10.6)
  Please reinstall xformers (see https://github.com/facebookresearch/xformers#installing-xformers)
  Memory-efficient attention, SwiGLU, sparse and more won't be available.
  Set XFORMERS_MORE_DETAILS=1 for more details
xFormers 0.0.29.post3
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]:             unavailable
[email protected]:             unavailable
[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:                  unavailable
sp24.sparse24_apply:                               unavailable
sp24.sparse24_apply_dense_output:                  unavailable
sp24._sparse24_gemm:                               unavailable
[email protected]:                 available
[email protected]:                        available
swiglu.dual_gemm_silu:                             unavailable
swiglu.gemm_fused_operand_sum:                     unavailable
swiglu.fused.p.cpp:                                not built
is_triton_available:                               False
pytorch.version:                                   2.6.0+cpu
pytorch.cuda:                                      not available
dcgm_profiler:                                     unavailable
build.info:                                        available
build.cuda_version:                                1204
build.hip_version:                                 None
build.python_version:                              3.10.11
build.torch_version:                               2.6.0+cu124
build.env.TORCH_CUDA_ARCH_LIST:                    6.0+PTX 7.0 7.5 8.0+PTX 9.0a
build.env.PYTORCH_ROCM_ARCH:                       None
build.env.XFORMERS_BUILD_TYPE:                     Release
build.env.XFORMERS_ENABLE_DEBUG_ASSERTIONS:        None
build.env.NVCC_FLAGS:                              -allow-unsupported-compiler
build.env.XFORMERS_PACKAGE_FROM:                   wheel-v0.0.29.post3
source.privacy:                                    open source

If I am inside the venv, I get this:

(venv) C:\Server\Docker\stable-diffusion-webui-forge>python -m xformers.info
xFormers 0.0.27
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:               available
memory_efficient_attention.cutlassB:               available
memory_efficient_attention.decoderF:               available
[email protected]:        available
[email protected]:        available
memory_efficient_attention.smallkF:                available
memory_efficient_attention.smallkB:                available
memory_efficient_attention.triton_splitKF:         available
indexing.scaled_index_addF:                        available
indexing.scaled_index_addB:                        available
indexing.index_select:                             available
sequence_parallel_fused.write_values:              available
sequence_parallel_fused.wait_values:               available
sequence_parallel_fused.cuda_memset_32b_async:     available
sp24.sparse24_sparsify_both_ways:                  available
sp24.sparse24_apply:                               available
sp24.sparse24_apply_dense_output:                  available
sp24._sparse24_gemm:                               available
[email protected]:                        available
swiglu.dual_gemm_silu:                             available
swiglu.gemm_fused_operand_sum:                     available
swiglu.fused.p.cpp:                                available
is_triton_available:                               True
pytorch.version:                                   2.3.1+cu121
pytorch.cuda:                                      available
gpu.compute_capability:                            8.9
gpu.name:                                          NVIDIA GeForce RTX 4070 Laptop GPU
dcgm_profiler:                                     unavailable
build.info:                                        available
build.cuda_version:                                1201
build.hip_version:                                 None
build.python_version:                              3.10.11
build.torch_version:                               2.3.1+cu121
build.env.TORCH_CUDA_ARCH_LIST:                    6.0+PTX 7.0 7.5 8.0+PTX
build.env.PYTORCH_ROCM_ARCH:                       None
build.env.XFORMERS_BUILD_TYPE:                     Release
build.env.XFORMERS_ENABLE_DEBUG_ASSERTIONS:        None
build.env.NVCC_FLAGS:                              -allow-unsupported-compiler
build.env.XFORMERS_PACKAGE_FROM:                   wheel-v0.0.27
build.nvcc_version:                                12.1.66
source.privacy:                                    open source

You could try to look through the open and closed issues at their site, and if you don't find anything, open an issue there.

[Silvicultor]

OK, activating the venv before executing python -m xformers.info helps. Now it's displaying:
is_triton_available: True
But, still xformers isn't actually being able to use Triton.
Still saying on forge startup:
ModuleNotFoundError: No module named 'triton'
... despite Triton correctly being installed in the venv. Maybe it's an issue with xformers?

[SvenErik1968]

I don't have that error during startup, so I suggest you ask about this on the triton-windows site.

Here is what I see (xformers mentioned several times, but triton is not mentioned):

venv "C:\Server\Docker\stable-diffusion-webui-forge\venv\Scripts\Python.exe"
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug  1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: f2.0.1v1.10.1-previous-659-gc055f2d4
Commit hash: c055f2d43b07cbfd87ac3da4899a6d7ee52ebab9
Installing requirements
CUDA 12.1
Launching Web UI with arguments: --xformers --force-enable-xformers --allow-code --skip-google-blockly
Total VRAM 8188 MB, total RAM 65256 MB
pytorch version: 2.3.1+cu121
xformers version: 0.0.27
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce RTX 4070 Laptop GPU : native
Hint: your device supports --cuda-malloc for potential speed improvements.
VAE dtype preferences: [torch.bfloat16, torch.float32] -> torch.bfloat16
CUDA Using Stream: False
Using xformers cross attention
Using xformers attention for VAE
ControlNet preprocessor location: C:\Server\Docker\stable-diffusion-webui-forge\models\ControlNetPreprocessor
[-] ADetailer initialized. version: 25.3.0, num models: 59
C:\Server\Docker\stable-diffusion-webui-forge\venv\lib\site-packages\timm\models\layers\__init__.py:48: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
CivitAI Browser+: Aria2 RPC started
[ForgeCouple] - INFO - Loaded Adv. Presets...
14:36:45 - ReActor - STATUS - Running v0.7.1-b3 on Device: CUDA
Extension: Extra Schedulers: adding new schedulers
2025-03-24 14:36:48,018 - ControlNet - INFO - ControlNet UI callback registered.
[ERROR]: Config states C:\Server\Docker\stable-diffusion-webui-forge\config_states\civitai_subfolders.json, "created_at" does not exist
Model selected: {'checkpoint_info': {'filename': 'C:\\Server\\Docker\\stable-diffusion-webui-forge\\models\\Stable-diffusion\\SDXL 1.0\\Realistic\\naturalustV3_v30.safetensors', 'hash': '3775ba6d'}, 'additional_modules': ['C:\\Server\\Docker\\stable-diffusion-webui-forge\\models\\VAE\\sharpspectrumvaexl_v1.safetensors'], 'unet_storage_dtype': None}
Using online LoRAs in FP16: False
Running on local URL:  http://127.0.0.1:8000

To create a public link, set `share=True` in `launch()`.

[E2GO]

May be this reddit post will be helpful somehow.
Automatic installation of Pytorch 2.8 (Nightly), Triton & SageAttention 2 into Comfy Desktop & get increased speed: v1.1: https://www.reddit.com/r/StableDiffusion/comments/1jivngj/automatic_installation_of_pytorch_28_nightly/