Nvidia CUDA on Windows - Guide

[shadeau9]

Not sure what the best way to add this in is so I thought I'd make a discussion. Every release I follow these steps to get GPU acceleration to work with my Nvidia GPU:

1. Ensure that you've already installed Nvidia's Cuda Toolkit (you only need to do this the first time or when you're updating the Cuda Toolkit)
2. Download source code zip
3. Extract to folder of my choosing
4. Run ebook2audiobook.cmd in extracted folder (I did not run as admin)
5. Wait for setup to complete
6. Close command prompt where setup was running
7. Open anaconda prompt
8. Activate python_env in extracted folder: conda activate "Path\To\Python_Env"
9. Run pip uninstall torch torchvision torchaudio -y to remove the automatically installed version of torch, torchvision, and torchaudio
10. Run pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128 (I'm using nightly and cu128 because of my GPU requirements, but you should install the version that makes sense for you based on your GPU and the Pytorch install instructions) to install the Cuda compatible version of torch, torchvision, and torchaudio
11. Run ebook2audiobook.cmd in extracted folder (I did not run as admin)

So far this has consistently worked for me after updating to new versions of ebook2audiobook. On an RTX 5090 a 7hr 13min audiobook can be created in 1hr 54min.

[ROBERT-MCDOWELL]

you don't need to 6. Close command prompt where setup was running
stay inside the ebook2audiobook folder.

the rest is what it needs to do when the current torch installed by coqui-tts is not compatible with your GPU.

[LexiconCode]

I had to do this as well with a 2070 rtx which is something I think would still be supported yet.

[nl255]

Just a note on something that tripped me up. You must use conda activate .\python_env for this to work, don't just use conda activate (I guess I am too used to software that uses venv).

[ROBERT-MCDOWELL]

no because you can have many other python env by name. specify ./ means you are pointing to a folder, not a name

[nl255]

You can, but unless someone is not using the standard ebook2audibook setup (such as by editing the cmd/ps1 launch file) the python env it uses is in [ebook2audiobook directory]\python_env. By default, the one used by ebook2audiobook doesn't show up on 'conda list' at all (at least it doesn't for me). Thus the following, once ebook2audiobook has gone through the setup process is what I did to get it working on Windows. (In my case I just stick with cu118, adding --pre to pip install may be necessary depending on which one you need)

conda activate .\python_env
pip uninstall torch torchvision torchaudio -y
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/[whatever cuda version you need]
conda decativate

[ROBERT-MCDOWELL]

conda info --envs should show you .\python_env

[nl255]

conda info --envs should show you .\python_env

It didn't for me, as of v25.3.10 which is when I wrote that, but it seems to do so (though there is only a path given, no actual name) with a fresh install of v25.3.22.

[ROBERT-MCDOWELL]

ok good to know

[DrewThomasson]

so.... should we make a wiki of this or.. idk

[DrewThomasson]

u know this thinggy

https://github.com/DrewThomasson/ebook2audiobook/blob/v25/Dockerfile

[DrewThomasson]

Cause that does include all the working torch variants I collected ☝️

at least for Debian linux docker...

[ROBERT-MCDOWELL]

yes we can add a major.minor cuda version as an --option.
major, minor are important as some GPU are supported on 11.4 and dropped on 11.8 for ex.

[DrewThomasson]

oh interesting I didn't know it could get that fine-grained

[ROBERT-MCDOWELL]

usually it;'s named cu118 etc...

[jestarray]

@shadeau9 https://developer.nvidia.com/cuda-downloads Is this the nvidia toolkit? I did your steps and I still have a problem with nvidia gpu not being detected

[ROBERT-MCDOWELL]

eb2ab is running under its own python env which is python 3.12. running ebook2audiobook.sh will install everything for you.
then you must enter under its python env:
conda activate ./python_env (linux/mac) .\python_env (windows)
then uninstall torch/torchaudio/torchvision and install your torch/torchaudio/torchvision

[shadeau9]

@jestarray Do the same check for the eb2ab python environment and see what it says. As stated, when you set up eb2ab it creates its own python environment and you'll need to make sure that torch version is compatible with your cuda version.

[jestarray]

Yes, I ran the said commands under conda activate .\python_env on windows and the output is still the same. It is still not detecting the gpu though

[shadeau9]

If it gave you the same output then it isn't activating the environment correctly since the environment is python 3.12 instead of 3.13. If you're running just conda activate .\python_env then I'd recommend using the full path instead. There should be a "python_env" folder inside your eb2ab installation folder and that's what you need to activate. If you're having issues with Powershell then try an Anaconda prompt instead.

[shadeau9]

If it gave you the same output then it isn't activating the environment correctly since the environment is python 3.12 instead of 3.13. If you're running just conda activate .\python_env then I'd recommend using the full path instead. There should be a "python_env" folder inside your eb2ab installation folder and that's what you need to activate. If you're having issues with Powershell then try an Anaconda prompt instead.

[DrewThomasson]

Made a Wiki of this btw

https://github.com/DrewThomasson/ebook2audiobook/wiki/GPU-ISSUES

[shadeau9]

Thank you for putting that together!

[SyamsQ]

Failed

[ROBERT-MCDOWELL]

pip cache purge. your issue has nothing to do with eb2ab or torch but your OS using another process. or reboot.