OBSIDIAN Neural Provider

Related Repositories

Repository	Description
obsidian-neural-central	Central inference server
obsidian-neural-provider ← you are here	Provider kit — run a GPU node on the network
obsidian-neural-frontend	Storefront & dashboard
obsidian-neural-controller	Mobile MIDI controller app
ai-dj	VST3/AU plugin (client)

---

Overview

OBSIDIAN Neural is an open source VST3/AU workstation for AI music generation designed for live performance. This repository contains the provider kit: a containerized FastAPI inference server that now supports 8 specialized AI models simultaneously. By running this kit, you contribute your GPU to the distributed network and earn a share of the platform's revenue.

🧠 The Multi-Model Engine

Each provider node is now a versatile workstation capable of switching between 8 specialized "brains" on the fly:

1. Stable Audio Open 1.0 — General purpose foundation for full-mix textures.
2. Foundation-1 — Surgical tag-based control for melodic and harmonic phrasing.
3. Audialab EDM Elements — High-energy EDM leads, supersaws, and plucks.
4. RC Infinite Pianos — High-fidelity grand and electric piano performances.
5. RC Vocal Textures — Choral and operatic vocal chord progressions.
6. SAO Instrumental — Modern indie, rock, and lofi stems.
7. StableBeaT — Advanced trap drum machine and 808 grooves.
8. Gluten-V1 — Specialized loop engine for trap and wavy melodic motifs.

Auto-Config Technology: The provider script automatically extracts optimal parameters (Steps, CFG Scale, Conditioning Duration) directly from each model's internal model_config.json to guarantee the best possible audio quality.

---

How it works

Musician in their DAW
↓ types a prompt or draws on the canvas
OBSIDIAN Neural central server
↓ finds an available Multi-Model GPU Provider
Your machine (provider)
├── LLM inference (Gemma 4 via Ollama)
│       ↓ optimizes the prompt / analyzes the drawing
│       ↓ returns structured JSON response
└── Audio generation (Dynamic Model Stack)
        ↓ Loads weights (.safetensors) for the requested model
        ↓ Generates audio using lab-tested settings (from model config)
        ↓ returns validated WAV
Musician receives the sound in real time

Subscription revenue is redistributed equally among all eligible providers each month via Stripe Connect, after deduction of a 15% platform fee covering infrastructure costs (fal.ai fallback, hosting, maintenance).

---

Requirements

Component	Specification
NVIDIA GPU	RTX 3070+ (8 GB VRAM min, 12 GB+ recommended)
RAM	16 GB
Storage	~40 GB (required for the full 8-model suite)
OS	Windows / Linux
CUDA	11.8+
Docker	20.10+ with NVIDIA Container Toolkit

---

What your provider runs

Each provider runs two inference stacks:

Stack	Model / Capabilities
Audio	8 Specialized Models (On-demand loading)
LLM	gemma4:e2b via Ollama for prompt optimization and vision

Jobs are mutually exclusive — your provider processes one request at a time (LLM or Audio) to ensure maximum VRAM availability and stability.

---

Quick start

1 — Benchmark your GPU

Verify your GPU can handle the high-quality multi-model generation:

docker run --rm --gpus all \
  -e HF_HOME=/root/.cache/huggingface \
  innermost47/obsidian-neural-provider:latest \
  python benchmark.py

2 — Join the network

1. Send an email to contact@obsidian-neural.com with your GPU model and public URL.
2. Once approved, the admin will send you your activation token (OBSIDIAN_TOKEN).

---

3 — Network setup

Your machine must be reachable over HTTPS with WebSocket support.

3a — Get a free domain (DDNS)

If you don't have a static IP, use DuckDNS:

Linux Setup:

mkdir -p ~/duckdns
cat > ~/duckdns/duck.sh << 'EOF'
echo url="https://www.duckdns.org/update?domains=YOUR_DOMAIN&token=YOUR_TOKEN&ip=" | curl -k -o ~/duckdns/duck.log -K -
EOF
chmod +x ~/duckdns/duck.sh
(crontab -l 2>/dev/null; echo "*/5 * * * * ~/duckdns/duck.sh >/dev/null 2>&1") | crontab -

3b — Port forwarding

On your router admin:

• Forward port 443 → your machine's local IP:443
• Forward port 80 → your machine's local IP:80 (for SSL)

3c — Install nginx + SSL

Linux Example:

sudo apt install nginx certbot python3-certbot-nginx -y
sudo certbot --nginx -d myprovider.duckdns.org

Edit your nginx config to enable WebSocket proxying to port 8000:

location / {
    proxy_pass http://localhost:8000;
    proxy_http_version 1.1;
    proxy_set_header Upgrade $http_upgrade;
    proxy_set_header Connection "upgrade";
    proxy_read_timeout 300s;
}

---

4 — Run the provider

docker run -d \
  --name obsidian-provider \
  -e OBSIDIAN_TOKEN=your_activation_token \
  -e CENTRAL_SERVER_URL=https://central.server.url.com \
  --gpus all \
  -p 8000:8000 \
  -v obsidian_data:/data \
  --restart unless-stopped \
  innermost47/obsidian-neural-provider:latest

The container will:

• Activate itself and download weights for the 8 specialized models.
• Start Ollama for prompt analysis.
• Connect to the central registry via WebSocket.

---

Security & verification

The central server ensures network integrity via:

Audio proof-of-work — periodic spectrogram comparisons against reference banks for all 8 models. LLM conversation echo — verifies LLM responses haven't been tampered with. Canary tests — random invalid requests to verify provider validation logic.

---

Monthly redistribution

Redistribution Formula: Monthly revenue - 15% platform fee = Distributable amount ÷ nb eligible providers

Eligibility:

1. Presence — Online ≥ 8h on at least 80% of active days.
2. Activity — Processed at least 1 real job (non-fallback) during the month.

---

Building from source

git clone https://github.com/innermost47/obsidian-neural-provider.git
cd obsidian-neural-provider
docker build --build-arg HF_TOKEN=your_hf_token -t obsidian-neural-provider .

---

Public Data Dashboard

Track active subscribers and monthly redistribution history at: obsidian-neural.com/public.html

---

License

GNU Affero General Public License v3.0

---

Made with 🎵 for the future of sampling — obsidian-neural.com