🤖 AI Generated Codebase

This project was entirely designed and implemented by Claude 4.5 Opus, Claude 4.5 Sonnet, and Gemini 3.0 Pro Preview, acting as autonomous software engineering agents under human supervision.

ML model services for the halfremembered agentic music production system.

Each service = one process, one model, one bespoke API.

Note: This repo is hand-crafted for a specific AMD ROCm setup on Arch Linux. If you're running different hardware, you'll need to adapt the PyTorch/ROCm configuration.

┌─────────────────────────────────────────────────────────┐
│  Your Application / MCP Server / CLI                    │
│  - Orchestrates model calls                             │
│  - Combines outputs (MIDI → audio, etc.)                │
└──────────────────────┬──────────────────────────────────┘
                       │ HTTP (localhost:200x)
                       ▼
┌─────────────────────────────────────────────────────────┐
│  Model Services (Python/FastAPI)       ← this repo      │
│  - One process per model                                │
│  - Independent venvs (uv)                               │
│  - Bespoke APIs per model                               │
│  - Ports 2000-2099                                      │
│  - Managed via systemd user units                       │
└─────────────────────────────────────────────────────────┘

Each service is standalone - start what you need, ignore the rest.

# Install just (task runner)
# Arch: pacman -S just
# Mac: brew install just

# Set up a service
just sync clap

# Run a service
just run clap

# Check all services
just status-all

# Run tests
just test clap

This repo is developed on:

• CPU: AMD Ryzen AI MAX+ 395 (32 CUs)
• GPU: AMD Radeon 8060S (gfx1151, RDNA 3.5, 40 CUs)
• VRAM: 96GB unified memory (shared CPU/GPU)
• Memory bandwidth: ~240 GB/s (the bottleneck for LLM inference)
• OS: Arch Linux (rolling release)
• Python: 3.13 (via python313 package)
• ROCm: 7.10.0 via official AMD wheels
• PyTorch: 2.9.1+rocm7.10.0 from repo.amd.com/rocm/whl/gfx1151/

Key constraint: Memory bandwidth limits 7B model inference to ~13-17 tok/s regardless of GPU utilization.

The AMD gfx1151 wheel index distributes the rocm metapackage as a source tarball. uv's resolver struggles with its transitive dependencies. We work around this with override-dependencies in each service's pyproject.toml:

[tool.uv]
override-dependencies = [
    "rocm==7.10.0",
    "rocm-sdk-core==7.10.0",
    "rocm-sdk-libraries-gfx1151==7.10.0",
]

See CLAUDE.md for full pyproject.toml configuration.

halfremembered-models/
├── hrserve/                 # Shared serving library
│   ├── pyproject.toml
│   ├── hrserve/
│   │   ├── model_base.py    # ModelAPI base class
│   │   ├── audio_utils.py   # Audio encoding
│   │   ├── midi_utils.py    # MIDI encoding
│   │   └── ...
│   └── tests/
│
├── services/
│   ├── clap/                # Each service is self-contained
│   │   ├── pyproject.toml   # Own deps, hrserve as editable
│   │   ├── api.py           # LitAPI implementation
│   │   ├── server.py        # Bootstrap
│   │   └── tests/
│   ├── orpheus-base/
│   ├── musicgen/
│   └── ...
│
├── systemd/                 # Service units
├── justfile                 # Task runner
└── CLAUDE.md               # Agent instructions

Each service exposes:

• POST /predict - Model inference (JSON in, JSON out)
• GET /health - Returns {"status": "ok"} when ready

Services are independent - call them directly via HTTP.

The source code in this repository is released under the MIT License. See LICENSE for details.

This repository contains code to run various ML models. The model weights themselves are subject to their own licenses (e.g., CC-BY-NC, Apache 2.0, Meta Research License).

• Orpheus Models: Custom trained, CC-BY-NC 4.0.
• MusicGen: Meta/Facebook Research (CC-BY-NC 4.0 / MIT).
• CLAP: LAION (Apache 2.0 / MIT).
• Demucs: Meta/Facebook Research (MIT).
• YuE: Open source (Apache 2.0).
• Qwen2.5-VL: Alibaba Cloud (Apache 2.0).

Please consult the individual service directories or original model repositories for specific weight licensing.

This codebase was generated by Claude 4.5 Opus, Claude 4.5 Sonnet, and Gemini 3.0 Pro Preview, acting as autonomous software engineering agents under human supervision.