File System Layout

The base directory is platform-specific. All Thumper-Run data lives under this root.

~/.local/share/tr-desktop/

~/Library/Application Support/tr-desktop/

%APPDATA%\tr-desktop\

The annotated tree below shows the top-level directories and their purpose.

tr-desktop/
├── apps/                  
    # Installed applications (git repos, venvs)
│   ├── comfyui/
│   ├── sillytavern/
│   └── .../
├── models/                
    # Downloaded model weights (safetensors, GGUF)
│   ├── checkpoints/
│   ├── loras/
│   └── vae/
├── gallery/               
    # Generated images with embedded metadata
├── crdt/                  
    # CRDT state, encryption keys, sync data
├── logs/                  
    # Application and process logs
└── cache/                 
    # Temporary data, thumbnails, download parts

Some third-party tools manage their own directories outside the Thumper-Run base. These are discovered automatically but can be overridden with environment variables.

HuggingFace Cache

HuggingFace Hub stores downloaded repos (model weights, configs) in a shared cache. Thumper-Run discovers models here via repo-level scanning.

~/.cache/huggingface/hub/
# Override: HF_HOME=~/.my-hf-cache

~/.cache/huggingface/hub/
# Same path as Linux; override: HF_HOME

%USERPROFILE%\.cache\huggingface\hub\
# Override: HF_HOME=C:\MyHFCache

Ollama Models

Ollama stores its own model blobs and manifests. Thumper-Run auto-detects the Ollama model directory and can auto-pull required models on app launch.

~/.ollama/models/
# Override: OLLAMA_MODELS=/path/to/models

~/.ollama/models/
# Override: OLLAMA_MODELS=/path/to/models

%USERPROFILE%\.ollama\models\
# Override: OLLAMA_MODELS=D:\OllamaModels

MIOpen Cache (AMD GPUs — Linux only)

MIOpen builds a "find database" of optimized GPU kernels on first run. This cache is specific to your GPU architecture (e.g., gfx1150) and can be large. Only applies to Linux with AMD ROCm GPUs.

~/.config/miopen/
# Typically 200–800 MB per GPU arch

comgr Cache (AMD GPUs — Linux only)

The AMD compiler cache stores compiled GPU code objects. Rebuilds automatically if deleted, at the cost of a longer first run. Only applies to Linux with AMD ROCm GPUs.

~/.cache/comgr/
# Typically ~650 MB

The following environment variables control where Thumper-Run and its dependencies store data.

AI models are large. Here is a rough guide to how much space different components use.