How keplar-404/ace-step1.5-for-cloud-deployment works
ACE-Step 1.5 is an AI music-generation application with CLI and persistent API entry points. The central workflow accepts generation parameters, optionally uses the 5Hz language model for lyrics or metadata, runs DiT audio inference, and returns or saves generated audio. The README also documents covers, segment editing, multi-track generation, source separation, and track completion; these are represented as handler capabilities. Some internal wiring is inferred only where excerpts or documentation support it; unsampled subsystem links are omitted.
Main entry points: handler.py Generation handler · inference.py Generation API · llm_inference.py Lyrics and metadata · model_downloader.py Model downloader
Components 18
User Interfaces
- Interactive CLI
cli.py - Inference API
server.py
Generation Workflow
- Generation API
acestep/inference.py - Generation handler
acestep/handler.py - Lyrics and metadata
acestep/llm_inference.py - Audio output
acestep/inference.py - Audio covers
acestep/handler.py - Segment editing
acestep/handler.py - Multi-track generation
acestep/handler.py - Source separation
acestep/handler.py - Track completion
acestep/handler.py
Model Runtime
- DiT models
acestep/models - 5Hz LM
acestep/llm_inference.py - nano-vLLM backend
acestep/third_parts/nano-vllm/nanovllm/llm.py - Model downloader
acestep/model_downloader.py
Customization
- LoRA management
acestep/core/generation/handler/lora_manager.py - LoRA registry and scaling
acestep/core/lora/service.py
Other components
- Music creator
Connections 20
- Music creator → Interactive CLI: requests music
- Music creator → Inference API: submits request
- Interactive CLI → Generation API: starts generation
- Inference API → Generation API: calls generation
- Generation API → Lyrics and metadata: generates lyrics and metadata
- Generation API → Generation handler: requests audio inference
- Generation handler → DiT models: runs diffusion
- Generation API → Audio output: saves generated audio
- Generation API → Lyrics and metadata: uses optional LM
- Lyrics and metadata → 5Hz LM: runs language model
- 5Hz LM → nano-vLLM backend: uses backend
- Generation handler → Audio covers: supports cover tasks
- Generation handler → Segment editing: supports repaint tasks
- Generation handler → Multi-track generation: supports instrument tasks
- Generation handler → Source separation: supports separation
- Generation handler → Track completion: supports track extension
- Model downloader → DiT models: downloads model files
- Model downloader → 5Hz LM: downloads model files
- Generation handler → LoRA management: manages adapters
- LoRA management → LoRA registry and scaling: uses registry and scaling