Version 5.0.0 beta on Ubuntu: temporary install steps

[iwr-redmond]

These instructions are intended to be used if the new network installer isn't working correctly for you. They will be unnecessary once the bugs in the network installer are ironed out.

Prepare the environment

1. You can set up Python 3.13 on your version of Ubuntu using the Deadsnakes repository:

sudo add-apt-repository -y ppa:deadsnakes/ppa
sudo apt-get update
sudo apt-get -y install python3.13-full

2. You will also need some additional dependencies installed and configured:

sudo apt-get -y install git pipx
pipx ensurepath

Set installation options

3. Set the correct Torch index according to your GPU:

• NVIDIA CUDA:

pytorch_url="https://download.pytorch.org/whl/cu126"

• AMD ROCm:

pytorch_url="https://download.pytorch.org/whl/rocm6.4"

• Intel XPU:

pytorch_url="https://download.pytorch.org/whl/xpu"

4. Set the desired extras using a comma-separated list:

• NVIDIA CUDA:

dgen_options="ncnn,bitsandbytes,gpt4all_cuda,xformers"

• AMD ROCm:

dgen_options="ncnn,gpt4all,xformers"

• Intel XPU:

dgen_options="ncnn,gpt4all"

The available options can be reviewed here.

Install dgenerate

5. You can now install dgenerate with pipx:

pipx install dgenerate[$dgen_options]@git+https://github.com/teriks/dgenerate@version_5.0.0 --pip-args "--extra-index-url $pytorch_url" --python python3.13

[Teriks]

This looks good, thank you.

There are definitely probably some things lurking to fix.

Just fixed extra index URL detection for when nvidia-smi reports cuda 13+

Attempting to allow the use of the installer environment for training scripts: https://dgenerate.readthedocs.io/en/version_5.0.0/manual.html#utilizing-the-python-environment-for-training

[iwr-redmond]

Hopefully they won't be needed very much, as the new installer is shaping up nicely. On the training support,xformers is not available for XPUs, and may need to be an optional extra.

[Teriks]

I’ve transitioned it into extra, no mps support or rocm as well. Though it is not a requirement for training at least, just helps with memory overhead

[iwr-redmond]

Thanks @Teriks. Looking at the index, PyTorch for ROCm 6.4 does include xformers.

[Teriks]

I think this support is very new, possibly just for torch 2.8.0 currently?

[iwr-redmond]

It's present in the ROCm 6.2.4 and ROCm 6.1 indexes as well.