NOTE: This README is a work in progress.
Rumple Dreambox runs on MacOS, Windows, and Linux.
NOTE: On Linux, run the custom torch install below before running the scripts/init-venv.sh
script.
brew install pyenv
pyenv global 3.12.0
brew install python3.12 # 10 and 11 are probably ok
brew install portaudio
brew install cmake
python3 -m pip install jax-metal
On H200 GPU's, you need to run nvidia-smi
and note the CUDA Version:
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 550.127.08 Driver Version: 550.127.08 CUDA Version: 12.4 |
|-----------------------------------------+------------------------+----------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 NVIDIA GH200 480GB On | 00000000:DD:00.0 Off | 0 |
| N/A 39C P0 253W / 700W | 9446MiB / 97871MiB | 14% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
And run the appropriate torch install, or modify scripts/install-torch-cuda
. The cu124 below matches nvidia-smi CUDA version 12.4:
win-scripts
has setup scripts.
Download model to client/audio/models:
You need to run the client and the server in two separate terminals. Before running each, you need to install the dependencies:
scripts/init-venv.sh
NOTE: To run the GUI, additionally you need to pip install gui-requirements.txt
Then activate the `venv`::
```bash
source scripts/activate-venv.sh
Then you can run both of these, except in separate terminals:
scripts/client.sh
scripts/server.sh [model-name] # optpional model name
Typical config looks like this:
"xl": {
"model_name": "stabilityai/stable-diffusion-xl-refiner-1.0",
"pipeline_class": "AutoPipelineForImage2Image",
"parameters": {
"strength": 0.33, # 0 incoming image is strong, 1 weak
"num_inference_steps": 20, # Lower for speed, higher for quality
"guidance_scale": 12.0, # How strongly to follow prompt
"width": 640,
"height": 512,
"negative_prompt": "painting, anime, illustration"
},
"scheduler": {
"type": "EulerAncestralDiscreteScheduler"
}
}
On Windows/Mac/Linux, you need to set:
export HF_API_KEY=<your key>
export SAFETENSOR_HOME=<safetensor directory>
- Configs for server-side image generation models are in
server/config.json
- Default prompt is in
client/gui/main_window.py
, lame but will fix later
keys in bashrc comment out pyqt and