-
-
Notifications
You must be signed in to change notification settings - Fork 134
Install ‐ LINUX ‐ Requirements and DeepSpeed
Depending on your Linux distribution, you'll need to install some packages for AllTalk and its TTS engines to function correctly:
-
Debian-based systems (e.g., Ubuntu):
sudo apt install libaio-dev espeak-ng gcc g++
-
RPM-based systems (e.g., CentOS, Fedora):
sudo yum install libaio-devel espeak-ng gcc g++
For standalone installations, the atsetup.sh
script should automatically install DeepSpeed.
For TGWUI installations, a simplified DeepSpeed installation method is available. Visit DeepSpeed 0.14.2 for Linux for pre-built wheel files and instructions.
-
Compatibility: DeepSpeed must match your:
- Python version (e.g., 3.10, 3.11, 3.12)
- PyTorch version (e.g., 2.0.x, 2.1.x, 2.2.x)
- CUDA version used by PyTorch (e.g., 11.8, 12.1)
-
Filename Explanation: For example,
deepspeed-0.14.2+cu121torch2.3-cp312-cp312-manylinux_2_24_x86_64.whl
breaks down as:- DeepSpeed version:
0.14.2
- CUDA version:
cu121
(meaning CUDA 12.1) - PyTorch version:
torch2.3
- Python version:
cp312-cp312
(meaning Python 3.12) - Linux build:
manylinux_2_24_x86_64
So
deepspeed-
0.14.2
+
cu121
torch
2.3
-
cp312
-cp312-manylinux_2_24_x86_64.whl
- DeepSpeed version:
-
Installation Process:
- Determine your Python, PyTorch, and CUDA versions (use AllTalk's diagnostics tool).
- Download the appropriate wheel file.
- Install using pip:
pip install deepspeed-0.14.2+[version]manylinux_2_24_x86_64.whl
-
CUDA Development Toolkit:
- Required version: 12.1.0 (tested and confirmed working)
- Installation in Conda environments:
conda install nvidia/label/cuda-12.1.0::cuda-toolkit=12.1
-
Important Distinction: The CUDA Development Toolkit is separate from:
- Your graphics card driver version
- The CUDA version used by your graphics driver
- PyTorch or Python CUDA versions
-
Warning Message: You may see a warning message when AllTalk starts with DeepSpeed installed. This is safe to ignore for AllTalk purposes.
- If you encounter symlink issues in conda environments, the script in
/system/config/fixsymlinks.sh
may help resolve them. - Always ensure you're in the correct Python virtual environment before installing or updating packages.
- If you update Python, PyTorch, or the CUDA version, remember to uninstall and reinstall the matching DeepSpeed version.