Use dynamic GPU duration based on text length

Reduce GPU duration for generate function from 120 to 10 seconds

Simplify to non-streaming render for ZeroGPU compatibility

Wrap streaming inference in @spaces.GPU, yield chunks after

Remove warmup - no GPU at startup on ZeroGPU

Revert to float32 model loading - codec needs float32, dtype casts in model.py handle ZeroGPU SDPA

Load model in bfloat16 natively

Add streaming audio generation with Web Audio player

trying to cast

trying to cast

tryin relative imports

INIT