About This Project

The open-source ACE-Step v1.5 music model was forked and extended with modal_app.py — a production-grade serverless deployment script. It packages the model into a Docker image, caches Hugging Face model weights during image build to avoid re-downloading on every cold start, and uses Modal's CRIU-based GPU memory snapshotting so warm starts are near-instant. GPU tier (L4, A10G, or A100) is automatically selected based on the configured LM model size (0.6B, 1.7B, or 4B). The app scales to zero when idle.

A stateless HTTP proxy that sits between the browser and the Modal inference layer. All requests are validated with Pydantic, rate-limited per session with slowapi (5 generations/min), and routed to the upstream ACE-Step API over an HTTP/2 persistent connection managed by httpx.AsyncClient. Audio is streamed back to the client via FastAPI StreamingResponse, keeping the internal Modal URL and API key away from the browser at all times.

A responsive web app built around a single prompt box — everything else is optional. Async job status is tracked via long-polling with exponential backoff. The waveform player is built on wavesurfer.js. All API shapes are validated client-side with Zod schemas matching the backend's Pydantic contracts, catching interface drift at the boundary. The frontend never contacts the Modal API directly — all traffic flows through the FastAPI backend.

A multi-layered pipeline from commit to production. Pre-commit hooks (detect-secrets, Bandit, Semgrep, Ruff) catch issues locally before they reach CI. Every PR triggers SAST scanning, dependency auditing, container scanning (Trivy), and the full test suite with 100% coverage enforcement. Dependabot keeps all dependencies current. Release Please automates semantic versioning and changelog generation across the monorepo on every merge to main.