xformers,flashattention,sage attention 1 click installer bat file for windows + mod

[FlowDownTheRiver]

After you download the official zip window built
Put this zip file in the main directory next to other bat files. Extract the bat file in the main diretory.
Run, select option 1 . Wait for the installation to finish.(It will inject python header files for python 3.10.6 for triton to work and install necessary packages. You don't need to do anything else.)

Edit : The first script I uploaded was giving you options now the script will show what it will do step by step. Just "press any key" to install... For those who may say that I lied in the title well it is 1 click + 1 button installer :)

Installer :
package installer V3 :

this shall fix all the problems according to my tests vs version 2. Note that it is also installing some extra packages potential for mod updates.

package_installer_V3.zip

Important note :
I have updated the package installer to V3,it shall handle the problems hopefully.
make sure you have a ->>>>>>>>>>>> "Fresh" <<<<<<<<<<<<- zip folder with nothing broken or nothing installed before. Otherwise if some broken library there I can't possibly fix all errors appear for everyone. Because I did not use force upgrade flag for the installer. That is why I recommend you to have a "fresh" zip folder. If you already downloaded the models just copy the "hf_download" folder inside webui folder to your fresh unzipped project folder not to download all the model files again. However if you want to try your luck that it can work for you,it doesn't harm anything. You can try installing without fresh project folder. If it doesn't work that way follow what I have written.

Uninstaller :
Due to high volume of problems occurred from different users I have added a reverse installer. This is for trouble shooting,testing with not enabled stuff etc. It is not a regular uninstaller but it just simply deletes related files.
Tip: If you get an error with the installer script,I suspect you may have a broken installation,this shall clear those libraries.
so use uninstaller first,then try installer to see if it works.

package_uninstaller.zip

The bat file shall look like this :

Your folder structure shall look like this :

Final result :

I decided to add a mod for those who wanna try :

Features :
V2 :

1. Start and End frame Control (from this PR Add Start and End Frame control，works great! #167)
2. Resolution Slider (from this PR Add resolution slider (from 240 -> 720 in 16-step increments) #158) -> Speed Improvement with lower resolution
3. Keep models in the memory after generation for future generations (my implementation) -> Speed Improvement
4. New defaults and theme -> Speed Improvement hopefully with new defaults and ability to test lower values
5. Keep Only Final Video option

@nanafy I applied your fix.Thanks.(V2)

V3 :
Everything V2 has +

1. Random and Last used Seed
2. GPU stats (Nvidia) -> So that you will know how to use your resources and set your parameters.

V4 :
Everything V3 has +

1. Negative prompting and other hidden parameters
2. Fps control with a slider
3. Larger sampling preview

V5 :
Everything V4 has +

1. Latent Window Size parameter added
2. States (Presets Manager)
   a) It can save settings,prompts
   b) Load last state ( Magic :) , on the next launch or page reload it can restore your last used settings,prompts )
   c) Your settings will be inside "states" directory under "webui" folder. You can share your settings and prompts with people.

V6 :
Everything V5 has +

1. Framepack F1 (forward) model support.
   Note 1 : Due to some heath problems and I got my rams giving error all of a sudden,I wasn't able to test to the limits. May include bugs. Hope not :)
   Note 2 : For V5 users, you have your states recorded inside "webui\states" directory. There you will find "ui_settings_manager.ini" file. This file is "not" compatible with V6 because I had to add model choice to the states. So you need to delete it and it will generate a new settings file for you which will be compatible with V6.
   Note 3 : Framepack F1 model will be downloaded on the first generation when you select that model. This will be done just for once. So be patient it may take some time to download depending on your internet speed.

V6_RC_2 :
Everything V5 has +

1. Framepack F1 (forward) model support.
   Note 1 : Due to some heath problems and I got my rams giving error all of a sudden,I wasn't able to test to the limits. May include bugs. Hope not :)
   Note 2 : For V5 users, you have your states recorded inside "webui\states" directory. There you will find "ui_settings_manager.ini" file. This file is "not" compatible with V6 because I had to add model choice to the states. So you need to delete it and it will generate a new settings file for you which will be compatible with V6.
   Note 3 : Framepack F1 model will be downloaded on the first generation when you select that model. This will be done just for once. So be patient it may take some time to download depending on your internet speed.

Notes on this specific version :
This shall act similar to V6. However. No transformers models (framepack original or f1) are loaded at the start of the program.
They are loaded on demand when the UI starts and when you press generate. When you switch models,previous model gets cleared from GPU memory. You can inspect it from gpu stats. So this shall prevent the UI not being able to be launched due to high demand of both models loading at the same time.

On top of that by request I have unlocked maximum resolution to 960. Ohhh Boi! Even on 24 GB vram, low vram is at 10,sampling steps at 15,it took around 5 minutes + just to do the first section. I will probably not use it, but this is for crazy people or for 5090 users? :) I don't feel crazy enough ^^

Please test this release candidate,I would love to get some feedback.

Mod Installation :
1 )
Inside webui folder backup these :
demo_gradio.py
diffusers_helper (folder)

2 )
Use the mod file :
(If you did it in the past skip this step) Put pynvml_installer_for_gpu_stats.bat next to other bat files (framepack_cu126_torch26 main directory) and run it.
put demo_gradio.py and diffusers_helper(folder)in webui folder(replace files)

FlowDownTheRiver_Mod_v5.zip

(V5 will be removed from the post after V6 confirmed to work properly)

FlowDownTheRiver_Mod_v6.zip

FlowDownTheRiver_Mod_v6_RC_2.zip

Your UI will look like this :

All Credits where it belongs to and great minds behind this project...

Have Fun!

[ljubomirb]

I kind of feel that fixing these type of problems would be easier using a docker container.
Developers, please, please, make one compose magic!
(hopefully with some persistent folder location for models)

[FlowDownTheRiver]

@ljubomirb I have added a new version of the bat file which addresses triton python.h file missing. Now everything works. Make sure to use this updated file.

[LIQUIDMIND111]

@ljubomirb I have added a new version of the bat file which addresses triton python.h file missing. Now everything works. Make sure to use this updated file.

ok sorry, i have ADD and read well but daydream a lot, lol, so this bat DOES NOT ADDS sage attention and triton, it just make them usable? or it DOES install ALL OF THEM accelerators?

EDIT, ok i just read OPTION 4..... so it does ALL for you.......

[FlowDownTheRiver]

@ljubomirb I have added a new version of the bat file which addresses triton python.h file missing. Now everything works. Make sure to use this updated file.

ok sorry, i have ADD and read well but daydream a lot, lol, so this bat DOES NOT ADDS sage attention and triton, it just make them usable? or it DOES install ALL OF THEM accelerators?

EDIT, ok i just read OPTION 4..... so it does ALL for you.......

Yes it does what it is showing automatically. Practically installing what are necessary and adding missing files...

[dallas4u]

Worked perfect. Thank you!

[LIQUIDMIND111]

@ljubomirb I have added a new version of the bat file which addresses triton python.h file missing. Now everything works. Make sure to use this updated file.

ok sorry, i have ADD and read well but daydream a lot, lol, so this bat DOES NOT ADDS sage attention and triton, it just make them usable? or it DOES install ALL OF THEM accelerators?
EDIT, ok i just read OPTION 4..... so it does ALL for you.......

Yes it does what it is showing automatically. Practically installing what are necessary and adding missing files...

ye sorry, i had to read it a few times... i appreciate your patience.

[LIQUIDMIND111]

@ljubomirb I have added a new version of the bat file which addresses triton python.h file missing. Now everything works. Make sure to use this updated file.

ok sorry, i have ADD and read well but daydream a lot, lol, so this bat DOES NOT ADDS sage attention and triton, it just make them usable? or it DOES install ALL OF THEM accelerators?
EDIT, ok i just read OPTION 4..... so it does ALL for you.......

Yes it does what it is showing automatically. Practically installing what are necessary and adding missing files...

hello, i got this at the end:

Successfully installed flash-attn-2.7.4
Successfully completed command.

NOTE: Skipping cleanup of downloaded file as it resides in the Hugging Face cache.

======================================================================
== Python Headers/Libs Copied
'All' is not recognized as an internal or external command,
operable program or batch file.

Press any key to continue . . .

IS THIS FINE?

[LIQUIDMIND111]

DONE!!

Currently enabled native sdp backends: ['flash', 'math', 'mem_efficient', 'cudnn']
Xformers is installed!
Flash Attn is installed!
Sage Attn is installed!

[LIQUIDMIND111]

can you also use this to install sage on portable COMFY UI?>

[FlowDownTheRiver]

can you also use this to install sage on portable COMFY UI?>

It needs modification as far as I can remember the comfyui portable environment structure. I don't have access at the moment to the files. I may take a look at it later on. However comfyui scripts mostly comes with their own installers or manager handles the installing. With a few path adjustments this script can also work on comfyui I assume.

Edit : There is also this problem. ComfyUI environment may have different version of python 3.11 instead of 3.10 so the precompiled packages needed to be changed depending on what pytorch version and cuda version is used. So it is a bit complicated than just editing the paths... 😊

[Keidan1993]

This worked for xformers and flash attention but is still not detecting sage attention as installed for some reason.

[FlowDownTheRiver]

This worked for xformers and flash attention but is still not detecting sage attention as installed for some reason.

I will take a look again later on by removing and reinstalling these libraries to replicate your problem or not. Tomorrow possibly.

Edit : Unfortunately I wasn't able to reproduce the error you get.

2025-04-20.07-18-56.mp4

[LIQUIDMIND111]

This worked for xformers and flash attention but is still not detecting sage attention as installed for some reason.

works well for me.....

[BBird6]

You are my hero for today!

I was struggling with SageAttention, now everything works fine.
Went from 12s/it down to 4s/it. (4060S/16GB)
Thank you Sir!