Question about VIBEVOICE_USE_MEAN

[zzrdwj]

why VLLM add this to support deterministic mode but HF code doesn't? Will this cause performance drop?

[pengzhiliang]

VIBEVOICE_USE_MEAN was added for debugging purposes in the vLLM plugin. The actual training uses .sample(), which is why the HF code doesn't include this option. Using .sample() is the intended behavior and should not cause performance degradation.