What problem does it solve? Setting up and running the HeartMuLa open-source music generation stack involves fragile dependency pins, required source-code patches, GPU memory constraints, and strict input formatting rules. This Skill packages the complete installation, patching, and generation workflow so you can produce full songs from lyrics and tags without rediscovering every pitfall. ## Core Features & Use Cases - Guided Installation & Patching: Clone heartlib, create a Python 3.10 venv, upgrade datasets/transformers, and apply the required RoPE cache and HeartCodec loading patches. - VRAM-Aware Generation: Run the 3B model on GPUs with as little as 8GB VRAM using lazy loading, or split models across multiple GPUs. - Input Formatting & Quality Rules: Format lyrics with bracketed structural tags ([Verse], [Chorus]) and comma-separated style tags, while enforcing float32 for HeartCodec to avoid audio quality degradation. - Use Case: A user asks for an open-source Suno alternative to generate a 4-minute romantic piano song from custom lyrics on a single consumer GPU, producing a 48kHz stereo MP3 in roughly real time. ## Quick Start Generate a song from my lyrics file and the tags "piano,happy,wedding" using the HeartMuLa 3B model with lazy loading enabled.
