实时生成速度有点慢

[iamyishan]

大佬好，后续有安排MOSS‑TTS‑Realtime支持vllm推理的计划吗？目前实时推理有点慢
[WS] sent END, total chars=83

[WS] first audio chunk, TTFB=3386ms
[WS] audio chunk #1, samples=23040, duration=0.96s
[WS] audio chunk #2, samples=23040, duration=0.96s
[WS] audio chunk #3, samples=23040, duration=0.96s
[WS] audio chunk #4, samples=23040, duration=0.96s
[WS] audio chunk #5, samples=23040, duration=0.96s
[WS] audio chunk #6, samples=23040, duration=0.96s
[WS] audio chunk #7, samples=23040, duration=0.96s
[WS] audio chunk #8, samples=23040, duration=0.96s
[WS] audio chunk #9, samples=23040, duration=0.96s
[WS] audio chunk #10, samples=23040, duration=0.96s
[WS] audio chunk #11, samples=23040, duration=0.96s
[WS] audio chunk #12, samples=23040, duration=0.96s
[WS] audio chunk #13, samples=23040, duration=0.96s
[WS] audio chunk #14, samples=23040, duration=0.96s
[WS] audio chunk #15, samples=23040, duration=0.96s
[WS] audio chunk #16, samples=23040, duration=0.96s
[WS] audio chunk #17, samples=23040, duration=0.96s
[WS] audio chunk #18, samples=23040, duration=0.96s
[WS] audio chunk #19, samples=23040, duration=0.96s
[WS] audio chunk #20, samples=23040, duration=0.96s
[WS] audio chunk #21, samples=23040, duration=0.96s
[WS] audio chunk #22, samples=23040, duration=0.96s
[WS] received END, total audio chunks=22

──────────────────────────────────────────────────
首包延迟=3386ms 总耗时=34.16s 音频时长=21.12s RTF=1.617 字数=83

[Zhyw0]

hello，后续我们会支持sglang的推理，现在可以看一下我们新更新的 #76 ，我们更新了用法，在warmup之后基本上首包延迟能达到180ms，RTF能达到0.5左右，如果没有warmup（每次使用新的音色进行生成），首包延迟和RTF就会略高。