Clean up debugging output

Add debugging for voice file discovery

Add voice and text prompt support for audible output

Pad audio to frame_size multiple to avoid CUDA graph shape mismatch

Fix streaming: use context managers and single mimi instance

Simplify: load models to GPU inside @spaces.GPU each call

Fix: cache mimi_decoder on CPU, cast to float for std()

Fix: device is positional arg in LMGen, not keyword

Optimize: monkeypatch CUDA to enable CPU model caching

Add device parameter to LMGen constructor

Enter streaming mode before reset_streaming calls

Move all model loading inside @spaces.GPU function

Fix moshi import path: loaders is in moshi.models

Implement PersonaPlex ZeroGPU demo

initial commit