Peak-normalize MelGAN audio so clips are audible

Patch T5Attention cache_position for SpectrogramDiffusionPipeline

Include traceback in MIDI inference errors

Surface payload details when MIDI inference fails

Export ModelMixin so encoder from_pretrained can resolve

Shim spectrogram_diffusion encoder classes for Diffusers

Put snapshot on sys.path so spectrogram_diffusion shim imports

Drop protobuf pin so note-seq can install

Install note-seq before loading SpectrogramDiffusionPipeline

Add custom handler dependencies (diffusers, note-seq, onnxruntime)

Add custom Inference Endpoint handler (SpectrogramDiffusionPipeline)

Duplicate from google/music-spectrogram-diffusion