gary4local v0.1.17

gary4local v0.1.17 pairs with Gary4JUCE v4.0.3.

• Adds seed support for Carey lego, complete, and cover requests.
• Adds optional ScragVAE download and toggle for ACE-Step generation — huge thanks to scragnog for the quality boost.
• Moves LoRA training timestep μ into advanced settings and keeps the ACE-Step/Side-Step default at -0.4, with 0.0 available for vocal datasets.

installer SHA-256: fef042b9d13c5e09666f24e8b31af9c00b4c5f90e8eb44b8569e19a9c9fa4ecb

gary4local v0.1.16

this is a small hotfix for carey environment rebuilds on clean machines.

• pins the last known-good trove-classifiers build dependency after a malformed upstream wheel began breaking Carey's isolated Hatchling setup
• treats omitted and explicit null sampling controls the same way in gary
• fixes Stable Audio loop requests with an explicit bar count
• clarifies carey's completion-duration API description

compatible with gary4juce v4.0.2.

installer SHA-256: 6334a9d62d4a106c64b4ec3eedc83c1a8ccc15f060fed0320136f0cdaa535fb0

gary4local v0.1.15

we've got ACE-Step LoRA training now.

• caption, prepare, and edit ACE-Step training datasets inside gary4local
• train LoRA or DoRA adapters against regular base or XL-base
• use Min-SNR loss weighting, best-checkpoint tracking, VRAM preflight, and the experimental balanced attention + MLP profile
• keep standard and XL LoRAs isolated automatically
• swap carey models more freely with improved model offloading

fair warning... this trainer has only been tested on a 5070 laptop GPU with training runs using ace-step-v15-base. plz let me know if you have any issues with xl-base.

compatible with gary4juce v4.0.2.

installer SHA-256: b71b567120f7f82d31d6c81ce5a23da10572757f5037a9224bca0959dd493b51

gary4local v0.1.14

What's changed

• Makes under-permissioned Hugging Face token errors explicit for gated Stable Audio 3 downloads.
• Adds the gated-token permission guide directly to the SA3 model screen.
• Repairs older SA3 environments by installing missing LoRA training dependencies before preprocessing.
• Cancels SA3 LoRA training without flashing PowerShell or taskkill windows.

Compatible with Gary4JUCE v4.0.2.

SHA-256: 62198c911d6888d7a4a9dcfcf6edb72197858fefeaf92eeef32650444b58d734

gary4local v0.1.13

gary4local v0.1.13 makes Hugging Face onboarding and gated model downloads more reliable on Windows.

Highlights

• Adds a compact visual guide showing the required public gated-repository permission for fine-grained Hugging Face tokens.
• Surfaces actionable model-download errors directly in the model list, including gated-token permission failures.
• Uses Hugging Face's official snapshot downloader for resumable downloads, Xet/LFS handling, and correct cache layout.
• Adds disk-space preflight checks and prevents incomplete SA3 snapshots from appearing ready.
• Loads complete SA3 Medium and bundled T5Gemma files directly from the local cache, avoiding unnecessary Hub checks during inference.
• Keeps the token-help popup visible in both configured and unconfigured token layouts.

Compatible with gary4juce v4.0.1.

SHA-256: 8266B104A416B7E9D53A2180446EEC52B0B28A83D4F17417EB05CDB47B750424

gary4local v0.1.12

Stable Audio 3 LoRA training has arrived on Windows.

Highlights

• Train SA3 LoRAs directly from the gary4local control center using a focused integration of dada-bots' underfit.
• Edit optional text-sidecar prompts, monitor selectable auto-scrolling logs, preserve job status across app restarts, and cancel preprocessing or training cleanly.
• Automatically register completed .safetensors adapters for immediate use in Gary4JUCE.
• Optionally enable experimental per-track latent-RMS loudness normalization during pre-encoding, with a recommended base-model target of 0.90.
• Harden SA3 decoding across CFG and preview paths by aligning latent device and dtype with the decoder.
• Reduce transient GPU memory pressure by converting half-precision models before transfer to CUDA.
• Improve service-row and log-viewer readability throughout the control center.

Compatible with gary4juce v4.0.0.

SHA-256: 82AEEDA164E51CF42B004EE21F3D21282E5DA8219F17705B514CD83CE65F1B49

gary4local v0.1.11

gary4local v0.1.11 adds the local Stable Audio 3 backend for Gary4JUCE v4.0.0.

• Adds SA3 model download, token reuse, and gated-access handling.
• Adds SA3 generation, loop, transform, and continuation endpoints.
• Adds experimental latent-prefix continuation mode.
• Adds SA3 LoRA registration and prompt-pool building from training sidecars.
• Adds advanced SA3 output shaping controls for hot LoRA/base outputs.

v0.1.10

• Fixes startup crashes caused by runtime icon.png copy/sync failures on some Windows machines.
• Carey cover requests now use cover-nofsq for clearer source detail.
• Lego tasks now use acestep-v15-base instead of the SFT model for more reliable results.

RETIRED - v0.1.7 DO NOT INSTALL

⚠️ RETIRED RELEASE - DO NOT INSTALL

This v0.1.7 installer was withdrawn after reports of startup crashes/white-screen failures on some Windows machines. The stable updater feed has been rolled back to v0.1.6 while v0.1.10 is validated and released.

If you installed v0.1.7 and gary4local will not open, install v0.1.6 from the previous release or wait for v0.1.10.

Original intent of this release:

• Carey covers route through cover-nofsq.
• Lego uses the regular base model instead of SFT.

v0.1.6

• Adds the new Carey local LoRA manager for gary4local, including dynamic local LoRA registry generation and caption-pool building for Gary4JUCE.
• Supports separate checkpoint folders and captions/source folders so Side-Step style training workflows import cleanly.
• Adds standard-versus-XL LoRA family filtering so Gary4JUCE only shows adapters that match the current Carey XL mode.
• Requires gary4juce v3.0.4 or newer for the new local LoRA functionality.