This release updates the validated audio, checkpoint and score-rendering dependencies while keeping the model runtime and UI on tested versions.
- SoundFile 0.14.0 fixes concurrent file-open behavior; safetensors 0.8.0 updates checkpoint handling; abcjs 6.7.1 updates score rendering. Ruff is now 0.17.1 for development checks.
- Gradio is fixed at 6.17.3 to keep the tested UI / Hub combination. Installations using a later Gradio return to 6.17.3. The companion Pinokio installer now respects the app requirement when repairing Gradio assets.
- macOS retains Torch 2.14.1; Linux/CUDA retains 2.10.0. YuE2, yue2.cpp, model weights, NumPy, Transformers, Hub, Accelerate, tiktoken and the separate Cover environment are unchanged.
Upgrade: Pinokio users click Update, then Start. No Reset is required. CLI users should use the platform override file as documented; the intentional upstream metadata mismatches for Torch (macOS), SoundFile and safetensors should not be “fixed” by reinstalling the older pins.
Validation: 309 tests passed in both an upgraded environment and a clean installation, with lint, formatting, type and syntax checks passing. Native audio/checkpoint and UI construction checks passed on macOS, Linux and Windows. A bounded 1,280-token BF16/MPS generation produced 51.2 seconds of audio; tokens, latents and audio were byte-identical to the baseline. Desktop/mobile score screenshots were pixel-identical to the old renderer, and all seven Studio tabs opened without page errors. Full CUDA generation was not rerun; this is not a claim of complete cross-platform model equivalence.
This release updates the validated audio, checkpoint and score-rendering dependencies while keeping the model runtime and UI on tested versions.
Upgrade: Pinokio users click Update, then Start. No Reset is required. CLI users should use the platform override file as documented; the intentional upstream metadata mismatches for Torch (macOS), SoundFile and safetensors should not be “fixed” by reinstalling the older pins.
Validation: 309 tests passed in both an upgraded environment and a clean installation, with lint, formatting, type and syntax checks passing. Native audio/checkpoint and UI construction checks passed on macOS, Linux and Windows. A bounded 1,280-token BF16/MPS generation produced 51.2 seconds of audio; tokens, latents and audio were byte-identical to the baseline. Desktop/mobile score screenshots were pixel-identical to the old renderer, and all seven Studio tabs opened without page errors. Full CUDA generation was not rerun; this is not a claim of complete cross-platform model equivalence.