How to decode CSM tokens into audio tensors for streaming

I built a streaming pipeline for CSM-1B that handles the token-to-audio decode. The key issue is that HF’s StaticCache uses index_copy_ which breaks CUDA graphs. Replacing it with slice assignment + a persistent backbone cache gets you reduce-overhead compilation. Full code with patches and a demo server: https://github.com/D3velop-llc/csm-rtx5090