Mike0021/zonos2

Move Gradio CSS to launch

Use eager DAC snake activation on ZeroGPU

Lazy-load ZONOS2 inside ZeroGPU request

Remove explicit CUDA library preloads

Lazy-load tokenizer worker

Add startup checkpoints

Use torch top-k routing fallback

Use eager MoE fallback on ZeroGPU

Add CUDA 12 cuBLAS dependency

Preload CUDA 12 cuBLAS for sgl_kernel

Add CUDA 12 NVRTC compatibility for sgl_kernel

Use tmp cache paths for ZeroGPU

Initial ZONOS2 ZeroGPU Space

initial commit