Installing in windows, currently impossible ?

[rough abyss]

Hi.
I tried many hours to install under windows 11, with CUDA 12.1
but seems not possible due to issues around xformers :

conda install -c xformers xformers returns** following packages are not available from current channels**
*so moving to *
pip install xformers==0.0.28.post2 returns Filename longer than 260 characters in ninja which seems known..https://github.com/facebookresearch/xformers/issues/1132
so moving to
using prebuilt https://github.com/facebookresearch/xformers/archive/refs/tags/v0.0.28.post2.zip ends in
File "C:\Users\37019Administrator\anaconda3\envs\unsloth_env\Lib\shutil.py", line 571, in copytree with os.scandir(src) as itr: ^^^^^^^^^^^^^^^ FileNotFoundError: [WinError 3] The system cannot find the path specified: 'C:\\Users\\Administrator\\Downloads\\xformers-0.0.28.post2\\third_party\\flash-attention\\flash_attn' [end of output]

Triton is succesfully installed in my environment.
I tried with torch 2.4 but doesn't seem the issue..
Any success ? What combination works in 12.1 under Windows ?

[velvet magnet]

I'd suggest you to use WSL instead of windows

[rough abyss]

That was my attempt yesterday but I stop when I realize that it was kind of hard to move back to Visual Studio and make the virtual environment work back under Windows.

I just have found a repo that moves to Docker. Worth considering it
https://github.com/Jiar/jupyter4unsloth

[velvet magnet]

I don't use WSL tbh but most people who do use windows here do with WSL

They said VS Code has support for WSL

[rough abyss]

Thanks..

[velvet magnet]

https://code.visualstudio.com/docs/remote/wsl

[bright copper]

you don't need to compile xformers, there are wheels for it. get torch 2.5.1, cuda 12.4 and run this:

pip3 install -U xformers --index-url https://download.pytorch.org/whl/cu124

[rough abyss]

Thanks, but I have multiple stuff already binded to 12.1
I guess there is no way to find an alternative for that right ?

[bright copper]

what do you have binded besides pytorch? xformers only supports 12.4 on windows

[rough abyss]

made it following your advice.
pip install -U xformers --index-url https://download.pytorch.org/whl/cu124

it installed xformers 0.0.28.post3
Then as soon as I use the notebook it turns..

`ImportError: Unsloth: If you are in Colab, we updated the top cell install instructions - please change it to below then press Disconnect Runtime and then Restart it.

%%capture

Installs Unsloth, Xformers (Flash Attention) and all other packages!

!pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
!pip install --no-deps "xformers<0.0.27" "trl<0.9.0" peft accelerate bitsandbytes

Otherwise in local machines, your xformers version of 0.0.28.post3 is too new.
Please downgrade xformers via pip install --force-reinstall \"xformers<0.0.27\""

[bright copper]

I use 0.0.28.post2

[rough abyss]

For sure, that's what it suppose to work (Linux).
But I spent the full day yesterday and there is no combinatiuon working so far. And the suggestion does not work

[bright copper]

I use that version on windows and it works fine with unsloth

[rough abyss]

There is no way to have it installed

(xformers==0.0.28.post2)

[bright copper]

that command will use the wrong index url

[bright copper]

what is your python version?

if it's 12.1 then use this wheel https://download.pytorch.org/whl/cu124/xformers-0.0.28.post2-cp312-cp312-win_amd64.whl#sha256=a8f52e4519640f58c92a2b1e1ad0e1c479b8406eb50f56920d97c3143039f0d2

[rough abyss]

NO.
I have 3.11

I found this wheel (xformers==0.0.28.post2) :

pip install https://download.pytorch.org/whl/cu124/xformers-0.0.28.post2-cp311-cp311-win_amd64.whl#sha256=852f1fa29d68a8eca8d7a100f95f5ca663c5ceda3f2060b7f00d390145f379c9

installed, but noticed that it automatically performs:
Uninstalling torch-2.5.1

and leaves a 2.5.0 even without cuda..
It gets worst.

[bright copper]

always use these pip arguments when doing stuff like this:
--no-dependencies --disable-pip-version-check --upgrade

[rough abyss]

let me retry. I have a fresh colab ready, but my perseverance will make it over this $hit, and my 3090 will be party tonight or hopefully tomorrow.

[half tinsel]

I didnt read this entire thread, but just install WSL2 if you didnt already.

You cant run these things on windows

WSL2 gives you linux.

And you can connect to vscode with built in support

https://github.com/ashishpatel26/Cuda-installation-on-WSL2-Ubuntu-20.04-and-Windows11

https://learn.microsoft.com/en-us/windows/wsl/install

[rough abyss]

Thanks for this: highly appreciated.
As I already have wsl 2 in place for docker in my environment, also already working CUDA for some containers, I guess it's "almost there" and just need the VSCode part.
Dumb question : all the necessary input (dataset) and output (weights etc) is still used from the system drive or else (Ubuntu, ..) ?

[bright copper]

to get files in ubuntu you can use "sudo systemctl start ssh" then connect via filezilla

[rough abyss]

I think I can get an idea about how to take it from here, but it is probably way outside my comfort zone, where even the most stupid thinkg like switching back and forth to Ubuntu , can easily take hours. It's just because of my limits, nothing bad related to unsloth here. Maybe I am wrong but a tutorial could help simplifying this