Skip to content

VibeSeq v0.1.0-alpha.18

Pre-release
Pre-release

Choose a tag to compare

View all tags
@github-actions github-actions released this 18 Jul 15:43
v0.1.0-alpha.18
2d05cee

Desktop runtime validation prerelease

This unsigned prerelease is for clean-machine validation of the VibeSeq desktop runtime, local persistence, arrangement editing, playback, and export.

It embeds the Studio, Stable Audio 3 execution source, a frozen portable CPU runtime, and verified first-run installers, so Node.js, Python, and uv do not need to be preinstalled. Model weights and the optional Windows CUDA runtime are not included in the executable.

On every launch VibeSeq checks the current platform's exact Stable Audio files. If VibeSeq Data is absent, empty, incomplete, or missing a required file, the Inference readiness panel offers a resumable SHA-256-verified download after the model terms are accepted:

Only the Stable Audio model format selected for the running hardware is downloaded. FlashAttention-compatible NVIDIA hardware must use the CUDA route and never silently downgrades to CPU. CPU-only Windows continues to use the public TFLite route. macOS and Linux store models below ~/VibeSeq Data/models; Windows stores models and its optional runtime below VibeSeq Data beside the installed VibeSeq.exe.

MuScriptor remains gated and is not redistributed. Each user accepts its Hugging Face conditions, downloads the pinned config.json and model.safetensors in their browser, and places them directly under SAVE CACHE UNDER. That exact revision folder opens in Finder, File Explorer, or the Linux file manager. VibeSeq does not download, select, or move these files or request or store a Hugging Face token; Verify files in cache only validates the files already in that folder.

Stable Audio and MuScriptor share the base Python/CUDA PyTorch environment and one persistent managed CUDA service on Windows, but install and verify as independent runtime profiles. Missing or broken MuScriptor packages never block Stable Audio readiness or generation. Stable Audio is loaded once and retains VRAM priority across generations; MuScriptor remains GPU-resident only when measured VRAM headroom allows it and otherwise stays on CPU without evicting Stable Audio. CUDA generation uses the official 8-step schedule and exact requested latent duration instead of generating an extra six seconds. FlashAttention 2 remains specific to Stable Audio, and the service never opens a console window.

Windows runtime installation clears and refreshes only the pinned FlashAttention wheel cache before synchronization, then automatically retries with an isolated temporary cache if interrupted installation metadata is incomplete. App shutdown terminates the complete sidecar, CUDA Python, and runtime-installer process trees; the persistent CUDA service also watches its parent process so it cannot remain orphaned and block a later Setup update.

Powered by Stability AI.

Windows is distributed as a current-user assisted NSIS installer so extraction progress is visible and VibeSeq Data remains beside the installed executable. The installed app then shows Local data, Audio engine, Health check, and Studio startup stages while its local service starts. Windows SmartScreen and macOS Gatekeeper may warn because code-signing identities are not configured yet. On Linux, run chmod +x VibeSeq-*.AppImage before launching. Verify the downloaded asset with SHA256SUMS.txt before running it.

macOS installation

This prerelease is not signed or notarized with an Apple Developer ID, so macOS may report that VibeSeq is damaged.

  1. Open the DMG and drag VibeSeq.app into Applications.
  2. Open Terminal and run:
xattr -dr com.apple.quarantine "/Applications/VibeSeq.app"

Double-click VibeSeq in Applications.
This command only removes the browser-added quarantine attribute from VibeSeq. It does not disable Gatekeeper system-wide and does not require sudo.
Only run it after downloading VibeSeq from the official GitHub release and verifying SHA256SUMS.txt.

Assets 7
Loading