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.

[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...