# s2pro-native

> Native C11 + CUDA streaming text-to-speech for Fish Audio S2-Pro (OpenAudio
> S2), designed and qualified for the NVIDIA DGX Spark (GB10, sm_121). The
> whole inference path is native — prompt building, the 36-layer slow-AR
> backbone, the 4-layer fast-AR residual decoder, the DAC vocoder, the
> scheduler and the HTTP server. Python, PyTorch, SGLang and vLLM are not in
> the runtime. Wall RTF 0.51 (compute / audio), 0.14 s to first audio.

Made by Luka Löhr — https://lukaloehr.com

Two commands on a Spark with Docker and the NVIDIA runtime:

    docker pull ghcr.io/luka-loehr/s2pro-native:1.0.1
    docker run --gpus all -p 8010:8010 -v s2pro-data:/data ghcr.io/luka-loehr/s2pro-native:1.0.1

The S2-Pro checkpoint is downloaded on first start into the volume. It is
licensed by Fish Audio under the Fish Audio Research License (non-commercial)
and is never part of the image. The runtime source is Apache-2.0.

## Setup

- [Agent setup runbook](https://s2pro.lukaloehr.com/setup.txt): end-to-end setup written for a coding agent. Asks first whether they actually have a DGX Spark, then pulls, runs and verifies with a real wav. Start here.
- [HTTP API](https://s2pro.lukaloehr.com/api.txt): /healthz, /v1/voices, /v1/tts, request fields, streaming semantics, voices and cloning.

## Reference

- [Performance](https://s2pro.lukaloehr.com/performance.txt): measured RTF and TTFA per serving path, the decode ladder, and the optimization record from 2.05 to 0.51.
- [Architecture](https://s2pro.lukaloehr.com/architecture.txt): the Dual-AR path, module layout, and the quantization ladder.
- [Everything in one file](https://s2pro.lukaloehr.com/llms-full.txt): all of the above concatenated.

## Source

- [GitHub repository](https://github.com/luka-loehr/s2pro-native): full source, docs, benchmarks, optimization reports. Apache-2.0.
- [Upstream checkpoint](https://huggingface.co/fishaudio/s2-pro): fishaudio/s2-pro, Fish Audio Research License.