Nvidia Blackwell 50XX GPU instructions (Windows)

[w-e-w]

As PyTorch have not yet released a compatible builds for Blackwell GPUs
we have been allowed to publish Early Access PyTorch wheels by Nvidia

---

There are several methods to run on Blackwell 50XX GPUs

Method 1: use the a new standalone release

Recommended for new install or for those not familiar with terminals / commands)

we have prepared a standalone release for Windows which can be downloaded here sd.webui-1.10.1-blackwell.7z (1.8GB)

1. Download sd.webui-1.10.1-blackwell.7z
2. Extract using 7z

Windows 11 seems to have it added native 7z support

3. Click run.bat to launch webui

note: do not enable --xformers, currently it is not compatible xformers v0.0.29.post2 is released, see below for instructions

Method 2: use the latest dev branch

Recommended for those who are migrating from existing installation and for those who are familiar with commands and terminals

• As of Blackwell compatibility (Windows only) #16817 the dev branch have been update with auto detection of blackwell GPUs, when detected it will automatically install the appropriate PyTorch which version

1. Follow the guide on how to switch branches see wiki How to switch to different versions of WebUI

instead of the release_candidate in the example substituted with dev to switch to dev branch

2. Add --reinstall-torch to COMMANDLINE_ARGS to tell webui to reinstall PyTorch
3. Launch webui

remember to remove--reinstall-torch from COMMANDLINE_ARGS after it's done reinstalling PyTorch

Method 3: manual upgrade

Meant for developers
The PyTorch wheels are provided at https://huggingface.co/w-e-w/torch-2.6.0-cu128.nv
This should be all the info you need

xformers (may not work)

to use xformers
you need you need to set the environment variable XFORMERS_PACKAGE to xformers==0.0.29.post2
on windows this can be done by adding the following line to to webui-user.bat before call webui.bat

set XFORMERS_PACKAGE=xformers==0.0.29.post2

the webui-user.bat should look like something this (click to expand)

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--xformers
set XFORMERS_PACKAGE=xformers==0.0.29.post2

call webui.bat

note: if you have previously used --xformers then you will need to add --reinstall-xformers to tell webui to reinstall xformers

similer to --reinstall-torch, remeber to remove --reinstall-xformers after it's done

[Andy91xx]

thank you its working i chose the portable installer, when using a detailer i get this in the command prompt

WARNING ⚠️ torchvision==0.20 is incompatible with torch==2.6.██████████████████████████| 40/40 [00:07<00:00, 5.38it/s]
Run 'pip install torchvision==0.21' to fix torchvision or 'pip install -U torch torchvision' to update both.
For a full compatibility table see https://github.com/pytorch/vision#installation

Running that command simply upgraded the torch that is installed in my system not in this portable version, also will this portable version work on linux ?

[w-e-w]

some people seems to have reported similer issues over at adetailer repo

• [Bug]: Ultralytics error when using Torch 2.6 cu128 Bing-su/adetailer#767

[Slider-Override]

The "Method 1" ends in a Instant BlackScreen crash follow by a reset of the PC, looks like there is a issue did someone else try it on Windows ?

[w-e-w]

can you try method 2
clone and then switch to dev branch before running
if mehtod 2 also creash then there might be something wrong with your hardware

what GPU do you have?
have you tried stress testing your GPU
do you have a PSU that has enough wattage
I will try physically reinstalling your GPU, maybe it's not installed properly?

[Kraisith]

I have the same exact issue, tried method 2 as well. Sometimes I can get it to generate 1 or 2 images but then the black screen happens (which then forces me to press the physical reset button on the PC). I have a 5080, FurMark etc. works fine, reseating didn't help.

[persistent-aura]

I'm also having this black screen issue on my new system with an RTX5080 and 1200W PSU. The latest nvidia driver hotfix did not fix the problem.

[Andy91xx]

Yup windows 11 works fine.

…

On Wed, 5 Feb 2025 at 20:05, Slider-Override ***@***.***> wrote: The "Method 1" ends in a Instant BlackScreen crash follow by a reset of the PC, looks like there is a issue did someone else try it on Windows ? — Reply to this email directly, view it on GitHub <#16818 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BFGWKZ6CZITMNLU26H6WB4D2OJVIRAVCNFSM6AAAAABWFKHJG2VHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTEMBXGQYDIOI> . You are receiving this because you commented.Message ID: <AUTOMATIC1111/stable-diffusion-webui/repo-discussions/16818/comments/12074049 @github.com>

[Slider-Override]

which card do you have ?
on my side it runs direkt into a Blackscreen and the PC reboots

[w-e-w]

I thinks the has a 5090

• UserWarning: NVIDIA GeForce RTX 5090 with CUDA capability sm_120 is not compatible with the current PyTorch installation #16822

[Slider-Override]

Yeah Its a 5090 MSI Trio, its possible thats a issue with the Driver, because see as well Blackscreen reports from other users.

[Andy91xx]

I have a gaming trio and works fine for me. I’m a fresh install of windows though due to having issues after installing the card.

…

On Thu, 6 Feb 2025 at 07:30, Slider-Override ***@***.***> wrote: Yeah Its a 5090 MSI Trio, its possible thats a issue with the Driver, because see as well Blackscreen reports from other users. — Reply to this email directly, view it on GitHub <#16818 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BFGWKZZA3C55PKIZGRZKQJL2OMFRJAVCNFSM6AAAAABWFKHJG2VHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTEMBXHA2DCNQ> . You are receiving this because you commented.Message ID: <AUTOMATIC1111/stable-diffusion-webui/repo-discussions/16818/comments/12078416 @github.com>

[PicceloStudios]

I'm on a 5090 FE from a 3090, on an old Windows 11 installation, all seem to work fine after I did a clean installation from the Method 1. share here.

Although I was having issues at the beginning trying to use a "Latent tile batch size" of 8, It was taking the memory to my RAM and taking ages.
Which was working perfectly fine previously in my 3090 with 24gb of VRAM Then I remember that I had xformers installed on my previous version which is still not supported for the 50 series, its a shame! I hope it gets xformers support soon.

Thanks for sharing this information was life saving!

[PicceloStudios]

forgot to say, give me feedback on if it works or not

Hello and sorry for my late reply.

It didnt work not sure if I did something wrong I did follow all your steps, upgraded CUDA from 12.1 to 12.6 and also included the xformers into the webui-user.bat, it did install them into the SD folder and when executing SD it doesnt show any errors.

Environment:

webui-user.bat

Maybe I have missed something? I suppose the pytorch 2.6 is already implicit with the Method 1 standalone Folder?

[w-e-w]

Maybe I have missed something?

from what I can see from screenshot looks fine

It didnt work not sure if I did something wrong

I need you to explain how is not working, how do you know it's not working
do dose the UI runs what dose the footer shows?

for 50xx it should show something like

if possible upload sysinfo.json goto settings > sysinfo > download system info
and past in any consol logs

---

I did follow all your steps, upgraded CUDA from 12.1 to 12.6

well, I don't recall including upgrading CUDA as a part of the instructions, pytorch includs CUDA runtime so shouldn't as long as your driver is reasonably up to date

I suppose the pytorch 2.6 is already implicit with the Method 1 standalone Folder?

the pytorch in the standalone package pytorch 2.6 with CUDA 12.8 not 12.6
and it doesn't matter you if the CUDA runtime of your system is 12.8 or 12.6, it uses CUDA runtime bundled with pytorch whitch is 12.8

[PicceloStudios]

Hello and thanks for your prompt response,

I'm sorry, I think the 12.6 cuda I read it from here:

here are my findings and why I think is not working for me:

Here is an example with the xformers activated:


1min.57.7sec.

and here is the same parameters without xformers:


1min.54.7sec.

I mention is not working because I'm not seen any actual performance uplift from having xformers ON or OFF in terms of speed or being able to use the x8 LTBS (Latent tile batch size) since xformers allow for less usage of Vram, which I was previously able to use with my 3090 with less Vram.

Not sure if this is something to do with this being a preview version?

Here is my sysinfo:
sysinfo-2025-02-07-09-45.json

And thank you so much for taking the time to help me out on this, I really appreciate the effort!

[w-e-w]

unfortunately I don't really have a good answer
all I can say is that is seems to be working fin on my 3090
I guess I have to update the post to
xformer unverified

on my system running with sd1.5 a 512x768 + 2x upscale + hires fix with another sd1.5 model

with xformers

without

with xformers processing time and memory usage is significantly lower

[PicceloStudios]

unfortunately I don't really have a good answer all I can say is that is seems to be working fin on my 3090 I guess I have to update the post to xformer unverified

on my system running with sd1.5 a 512x768 + 2x upscale + hires fix with another sd1.5 model

with xformers

without

with xformers processing time and memory usage is significantly lower

Thank you so much for your time and effort, I hoped you had a magic wand, in any case it is really appreciated.

Dont get me wrong, it is super fast compared to the 3090 thats for sure is just that of course I would like to squeze the performance properly, I guess I will have to wait until there is a stable version for me to test that actually enable the xformers on my 5090.

Here is an example with similar parameters to yours

and here is the same example with added parameters to yours (sd1.5 at 512x768 + 2x upscale + Tiled diffusion + Tiled VAE + Control net)

My issue is that I use SD to upscale large imagery pretty large up to 8K, so it takes time to upscale them and in this profesion (Archviz) every minute counts! I was hoping with xformers that I didnt have to split the images in smaller chunks and upscale them separately but rather do it in one go, it seems for now I will have to stick to this workflow.

Again I'm deeply grateful for your help! thank you so much!

[tripleS-Dev]

Im using 5090,

When it starts

gradio 3.41.2 requires pillow<11.0,>=8.0, but you have pillow 11.0.0 which is incompatible.
xformers 0.0.29.post2 requires torch==2.6.0, but you have torch 2.7.0.dev20250127+cu118 which is incompatible.

And When I want to train Lora

loading vae: <All keys matched successfully>
FETCH ComfyRegistry Data: 15/32
loading text encoder: _IncompatibleKeys(missing_keys=[], unexpected_keys=['text_model.embeddings.position_ids'])
Enable xformers for U-Net
Traceback (most recent call last):
  File "C:\Users\hj6ch\Downloads\ComfyUI_cu128_50XX\ComfyUI\custom_nodes\Lora-Training-in-Comfy\sd-scripts\train_network.py", line 1012, in <module>
    trainer.train(args)
  File "C:\Users\hj6ch\Downloads\ComfyUI_cu128_50XX\ComfyUI\custom_nodes\Lora-Training-in-Comfy\sd-scripts\train_network.py", line 236, in train
    vae.set_use_memory_efficient_attention_xformers(args.xformers)
  File "C:\Users\hj6ch\AppData\Local\Programs\Python\Python310\lib\site-packages\diffusers\models\modeling_utils.py", line 273, in set_use_memory_efficient_attention_xformers
    fn_recursive_set_mem_eff(module)
  File "C:\Users\hj6ch\AppData\Local\Programs\Python\Python310\lib\site-packages\diffusers\models\modeling_utils.py", line 269, in fn_recursive_set_mem_eff
    fn_recursive_set_mem_eff(child)
  File "C:\Users\hj6ch\AppData\Local\Programs\Python\Python310\lib\site-packages\diffusers\models\modeling_utils.py", line 269, in fn_recursive_set_mem_eff
    fn_recursive_set_mem_eff(child)
  File "C:\Users\hj6ch\AppData\Local\Programs\Python\Python310\lib\site-packages\diffusers\models\modeling_utils.py", line 269, in fn_recursive_set_mem_eff
    fn_recursive_set_mem_eff(child)
  File "C:\Users\hj6ch\AppData\Local\Programs\Python\Python310\lib\site-packages\diffusers\models\modeling_utils.py", line 266, in fn_recursive_set_mem_eff
    module.set_use_memory_efficient_attention_xformers(valid, attention_op)
  File "C:\Users\hj6ch\AppData\Local\Programs\Python\Python310\lib\site-packages\diffusers\models\attention_processor.py", line 395, in set_use_memory_efficient_attention_xformers
    raise ValueError(
ValueError: torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is only available for GPU
Traceback (most recent call last):
  File "C:\Users\hj6ch\AppData\Local\Programs\Python\Python310\lib\runpy.py", line 196, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "C:\Users\hj6ch\AppData\Local\Programs\Python\Python310\lib\runpy.py", line 86, in _run_code
    exec(code, run_globals)
  File "C:\Users\hj6ch\AppData\Local\Programs\Python\Python310\lib\site-packages\accelerate\commands\launch.py", line 996, in <module>
    main()
  File "C:\Users\hj6ch\AppData\Local\Programs\Python\Python310\lib\site-packages\accelerate\commands\launch.py", line 992, in main
    launch_command(args)
  File "C:\Users\hj6ch\AppData\Local\Programs\Python\Python310\lib\site-packages\accelerate\commands\launch.py", line 986, in launch_command
    simple_launcher(args)
  File "C:\Users\hj6ch\AppData\Local\Programs\Python\Python310\lib\site-packages\accelerate\commands\launch.py", line 628, in simple_launcher
    raise subprocess.CalledProcessError(returncode=process.returncode, cmd=cmd)
subprocess.CalledProcessError: Command '['C:\\Users\\hj6ch\\AppData\\Local\\Programs\\Python\\Python310\\python.exe', 'C:/Users/hj6ch/Downloads/ComfyUI_cu128_50XX/ComfyUI/custom_nodes/Lora-Training-in-Comfy/sd-scripts/train_network.py', '--enable_bucket', '--pretrained_model_name_or_path=C:\\Users\\hj6ch\\Downloads\\ComfyUI_cu128_50XX\\ComfyUI\\models\\checkpoints\\henmixReal_v6b.safetensors', '--train_data_dir=A:/temp/memory', '--output_dir=models/loras', '--logging_dir=./logs', '--log_prefix=seoah_v1', '--resolution=512,512', '--network_module=networks.lora', '--max_train_epochs=100', '--learning_rate=1e-4', '--unet_lr=1e-4', '--text_encoder_lr=1e-5', '--lr_scheduler=cosine_with_restarts', '--lr_warmup_steps=0', '--lr_scheduler_num_cycles=1', '--network_dim=32', '--network_alpha=32', '--output_name=seoah_v1', '--train_batch_size=5', '--save_every_n_epochs=50', '--mixed_precision=fp16', '--save_precision=fp16', '--seed=22', '--cache_latents', '--prior_loss_weight=1', '--max_token_length=225', '--caption_extension=.txt', '--save_model_as=safetensors', '--min_bucket_reso=256', '--max_bucket_reso=1584', '--keep_tokens=0', '--xformers', '--shuffle_caption', '--clip_skip=2', '--optimizer_type=AdamW8bit', '--persistent_data_loader_workers', '--log_with=tensorboard']' returned non-zero exit status 1.
Train finished
Prompt executed in 13.01 seconds

What is problem and How can I fix it?

[w-e-w]

I believe you're asking in the wrong repo...

[persistent-aura]

On a 5080. My screen turns black the moment I try to generate an image. On low res like 512x512 the screen will return after a few seconds but on higher resolutions it stays black and I need to restart the PC to fix the issue. The image still generates. Today's nvidia driver hotfix didn't fix the problem.

[PicceloStudios]

I know is a silly question but have you double checked all your connections? that are properly plug, the Displayport/Hdmi to the screen. it seems to be a weird artifact that of yours, only happen with SD? not while playing video games or any other GPU Intensive tasks?

I have seen collegues with 5090's FE that have issues with Photoshop, appears to blink the screen in the background while moving the cursor around.

Edit: My 5090 is working fine so far with the 572.42, I dont want to update the GPU Driver just in case something goes terribly wrong! I prefer to stay safe for the moment!

I had actually 1 issue while my computer was in sleep mode and after waking up turns out on a Blue screen and restarted by it self.

[persistent-aura]

Everything is plugged in well. Furmark benchmark tests come back fine. Games play without issues. I can run large LLMs locally like deepseek-r1 32B.
I've only had this issue while generating with stable diffusion so far. It also happens in comfyui. Doesn't seem to make a difference if I use lighter models like sd1.5 or heavier ones like SDXL.

[PicceloStudios]

Everything is plugged in well. Furmark benchmark tests come back fine. Games play without issues. I can run large LLMs locally like deepseek-r1 32B. I've only had this issue while generating with stable diffusion so far. It also happens in comfyui. Doesn't seem to make a difference if I use lighter models like sd1.5 or heavier ones like SDXL.

It seems this GPUs are still not optimized for SD at the moment... Mine runs fine, but Xformers doesnt work for me, I suppose we will have to wait until they drop a stable version thatworks properly with the 50 series.

[Zedrath]

Using Reforge it downloads the official release of a nightly for Cuda 12.8 for windows that was released a few days ago. But here, same issue. It seems semi stable with certain checkpoints, while others crashes the GPU into a black screen it cant recover from on first generation every time. I've been able to run some realistic models for maybe 20 generations, but then randomly crash.

I've got a 5080 as well. It seems support and optimization is horrible and we might have to wait a month or so more before they figure out how to run it stable.

[Panchovix]

It seems that nightly torch 2.7 from 21 February works fine. I can modify it on reforge, but torchvision has to be built (since doesn't seem windows have builds out of the box)

If linux, it should work fine with nightlies.

[Panchovix]

Added some builts torchvision for nightly 2.7 + cu128 for windows, it shouldn't give black images Panchovix/stable-diffusion-webui-reForge@ebe6e5e

[Fulll3]

Hi,

Does anyone know if Blackwell version is available also for Linux?

[ijash]

Hi,

Does anyone know if Blackwell version is available also for Linux?

Idk, but i have ran a1111 on blackwell.

first, install nvidia driver from nvidia site.

then, i follow the regular steps. run bash webui.sh and it will error cannot find kernel module for nvidia or somehing... .

abort the terminal, then activate the venv
and run pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128
then run bash webui.sh again.

tested on linux mint latest using rtx 5080.

[zlo899nv]

@Fulll3 Yes there is Linux support... Linux x86_64, Linux SBSA, and Windows through the command @ijash posted. Thanks!

[Fulll3]

Thanks guys. i got it running but without xformers it seems to perform same as 4090

[PicceloStudios]

Thanks guys. i got it running but without xformers it seems to perform same as 4090

Its the same for me, I got it running although it has Xformers running it doesnt seem to be working since the time generation is the same with them or without.

[Andy91xx]

Alternatively if using an Arch based distribution simply install the nvidia open dkms drivers from pacman or yay then follow these instructions although I installed a nightly build to get mine working

…

On Wed, 19 Mar 2025 at 17:00, Jastria Rahmat ***@***.***> wrote: Hi, Does anyone know if Blackwell version is available also for Linux? I have. first, install nvidia driver from nvidia site. theni follow the regular steps. then activate the venv and run pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128 then run bash webui.sh. tested on linux mint latest using rtx 5080. — Reply to this email directly, view it on GitHub <#16818 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BFGWKZZR7TSXF75NHSY7TVT2VGPC5AVCNFSM6AAAAABWFKHJG2VHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTENJVGQ4DKNA> . You are receiving this because you commented.Message ID: <AUTOMATIC1111/stable-diffusion-webui/repo-discussions/16818/comments/12554854 @github.com>

[Andy91xx]

I use Garuda with a 5090

…

On Wed, 19 Mar 2025 at 17:08, Andy L ***@***.***> wrote: Alternatively if using an Arch based distribution simply install the nvidia open dkms drivers from pacman or yay then follow these instructions although I installed a nightly build to get mine working On Wed, 19 Mar 2025 at 17:00, Jastria Rahmat ***@***.***> wrote: > Hi, > > Does anyone know if Blackwell version is available also for Linux? > > I have. > > first, install nvidia driver from nvidia site. > > theni follow the regular steps. > > then activate the venv > and run pip install --pre torch torchvision torchaudio --index-url > https://download.pytorch.org/whl/nightly/cu128 > then run bash webui.sh. > > tested on linux mint latest using rtx 5080. > > — > Reply to this email directly, view it on GitHub > <#16818 (reply in thread)>, > or unsubscribe > <https://github.com/notifications/unsubscribe-auth/BFGWKZZR7TSXF75NHSY7TVT2VGPC5AVCNFSM6AAAAABWFKHJG2VHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTENJVGQ4DKNA> > . > You are receiving this because you commented.Message ID: > <AUTOMATIC1111/stable-diffusion-webui/repo-discussions/16818/comments/12554854 > @github.com> >

[9wow]

Is it normal for xformers to not show up in the Cross attention optimization list (Settings\Optimizations) and for A1111 to say its using Doggettx in the terminal? xformers==0.0.29.post2 is installed and i've ran '--reinstall-xformers'

[Fateroachinii]

If the console displays the message "Applying attention optimization: Doggettx... done," it means that xformers did not activate. I have the same issue and don't know how to fix it—I followed the instructions exactly while installing the portable version. My GPU is a 5070 ti.

[TITORDFOX]

If the console displays the message "Applying attention optimization: Doggettx... done," it means that xformers did not activate. I have the same issue and don't know how to fix it—I followed the instructions exactly while installing the portable version. My GPU is a 5070 ti.

I have the same problem as you
Doggettx is displayed even after activating --xformers with the variable
set XFORMERS_PACKAGE=xformers==0.0.29.post2
and I tried with the new version it too 0.0.29.post3
but I have the same result

the only difference is that with --xformers activated
the generation of the 3K image takes 1 minute longer
and uses 10GB more Vram

my GPU is a 5090

[PicceloStudios]

If the console displays the message "Applying attention optimization: Doggettx... done," it means that xformers did not activate. I have the same issue and don't know how to fix it—I followed the instructions exactly while installing the portable version. My GPU is a 5070 ti.

I have the same problem as you Doggettx is displayed even after activating --xformers with the variable set XFORMERS_PACKAGE=xformers==0.0.29.post2 and I tried with the new version it too 0.0.29.post3 but I have the same result

the only difference is that with --xformers activated the generation of the 3K image takes 1 minute longer and uses 10GB more Vram

my GPU is a 5090

Unfortunately (Don't understand why) Xformers are not working within the Blackwell 5090 GPU's Architecture (Not sure about the other GPU's of the 50 series) Xformers are meant to reduce the amount of VRAM used for SD Image generation rather than increasing it, as well as speeding up the initial steps processes, so If you are having an extra minute or so by enabling them there most be something wrong with your installation, I repeat Xformers are not working so far, will need to wait until a fix comes around and activate those for our Blackwell GPUs, I also have a 5090, and use to do SD with a 3090 with Xformers and was able to do 12K without a problem, ofcourse with an increase on time generation.

In my personal case, it takes the same time to generate an image with or without Xformers activated, no matter the size of the image, even though it shows in the console that is working with Xformers activated, its a pain obviusly for someone that was accustomed to quick image generation starting processes.

I hope this helps.

[MadMan2k]

If the console displays the message "Applying attention optimization: Doggettx... done," it means that xformers did not activate. I have the same issue and don't know how to fix it—I followed the instructions exactly while installing the portable version. My GPU is a 5070 ti.

I have the same problem as you Doggettx is displayed even after activating --xformers with the variable set XFORMERS_PACKAGE=xformers==0.0.29.post2 and I tried with the new version it too 0.0.29.post3 but I have the same result
the only difference is that with --xformers activated the generation of the 3K image takes 1 minute longer and uses 10GB more Vram
my GPU is a 5090

Unfortunately (Don't understand why) Xformers are not working within the Blackwell 5090 GPU's Architecture (Not sure about the other GPU's of the 50 series) Xformers are meant to reduce the amount of VRAM used for SD Image generation rather than increasing it, as well as speeding up the initial steps processes, so If you are having an extra minute or so by enabling them there most be something wrong with your installation, I repeat Xformers are not working so far, will need to wait until a fix comes around and activate those for our Blackwell GPUs, I also have a 5090, and use to do SD with a 3090 with Xformers and was able to do 12K without a problem, ofcourse with an increase on time generation.

In my personal case, it takes the same time to generate an image with or without Xformers activated, no matter the size of the image, even though it shows in the console that is working with Xformers activated, its a pain obviusly for someone that was accustomed to quick image generation starting processes.

I hope this helps.

Same thing here with the 5070 Ti. No luck with xformers 0.0.29.post2. The system says it is activated, but there is no noticeable difference in generation speed

[LoksliSpb]

Hello everyone, I have installed a working build for the NVIDIA 5080 series graphics card using method 1. Thank you very much to those who made it. But only the SD work. FLUX and XL models not, for some reason they are not available in this build. Or is there something I don't understand?

[ijash]

Can you post the log?

[MadMan2k]

Could someone make ReActor work with this version of automatic1111?

I have this output

RuntimeError: D:\a\_work\1\s\onnxruntime\python\onnxruntime_pybind_state.cc:857 onnxruntime::python::CreateExecutionProviderInstance CUDA_PATH is set but CUDA wasnt able to be loaded. Please install the correct version of CUDA andcuDNN as mentioned in the GPU requirements page (https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html#requirements), make sure they're in the PATH, and that your GPU is supported.

[MadMan2k]

And yes, if it’s relevant, I start my SD using the run.bat file

[ijash]

Just maybe..

1. Disable xformers
2. Update to the latest nvidia driver.

That's it for now.
Again, report back with your commands, and logs. Hopefully me or others can help until it solved.

[MadMan2k]

1. I've tried even clean install, so xformers was not even enabled
2. I have the latest one 572.83

[MadMan2k]

Anyway, thank you for trying to help me

[MadMan2k]

I tried to run ReActor on ComfyUI, like you did, and it works after manually updating torch and torchvision. As far as I can tell (I'm not familiar with ComfyUI), it seems to work - at least I got a relevant output and encountered no errors. However, I still had no luck with Automatic1111

[daniel-krueger12]

SD Blackwell works for my 5090 but did you find a way to install xformers. or are there other ways for faster generation.
In another git I read that someone compared xformers with sdpa and that there were hardly any differences.
the problem is i can hardly find somethink about sdpa do you may know what that is and how to set it up with rtx 50 card.
Do you think we will have an easy way to install stable deffusion and xformers by Easter? i would love to spent my free time with sd.

[MadMan2k]

It says here that you can use xformers v0.0.29.post2, but many users (including me) have reported that even though it says xformers is being used, there is practically no change. So, for now, we probably just need to wait

[tiekom]

Hey, I got mine working. It went from 37.6 seconds on my RTX 4090 to 25 on my new 5090.

1. Install the dev branch on your PC, and then you need to make a few tweaks.

2. Install a beta version of xformers here. Just add these options to your webui-user.bat if you already had xformers delete it from venv\Lib\site-packages directly.

   set COMMANDLINE_ARGS=--force-enable-xformers --xformers
   set XFORMERS_PACKAGE=xformers==0.0.30.dev1005
   

3. Then you must ensure your torch version is: torch: 2.6.0+cu128.nv

4. If you do everything, you should see the option in the Optimization Tab in the settings.

****

I hope it fixes your issue. I think that the xformers next 0.0.30 version will fix the problems with the new cards.

[MODFireball]

@tiekom's solution worked for me, though I did it a bit lazier - just used the standalone and added "--reinstall-xformers" first to the commandline_args, similarly to first xformers instructions at the top. From quickly testing, takes 30% less time on my 5080.

[daniel-krueger12]

which dev branch blackwell SD or xformers

[LoksliSpb]

Hello everyone, I'm still using the stable Forge build from method 1, but I still haven't been able to add support for the FLUX and XL models. Is there any way to do this?