SGLang-Omni serves Fish Audio's Drama 3 TTS, claims 72% throughput gain from SPG topology

SGLang says Drama 3, Fish Audio's text-to-speech model, is supported on day 0 in SGLang-Omni, serving multi-speaker TTS, voice cloning and streaming audio through /v1/audio/speech.

It reports an 89ms median time to first audio on a single H200 for a single stream.

SGLang also says its SPG topology, which moves the Drama 3 vocoder to a second H200, raises throughput 72% and cuts median time to first audio 37% at 256 concurrent streams. The claim comes from an SGLang post that links a blog page describing the approach. The reports give no independent measurements of either result.