Update model card for 5090 optimized Fish S2

NOTICE.md

@@ -0,0 +1,14 @@
+This repository documents an RTX 5090 local serving profile for Fish Audio S2-Pro.
+
+It does not redistribute Fish Audio model weights. Users must obtain the base model
+from the upstream Hugging Face repository:
+
+https://huggingface.co/fishaudio/s2-pro
+
+Fish Audio attribution:
+
+This model is licensed under the Fish Audio Research License, Copyright (c) 39 AI,
+INC. All Rights Reserved.
+
+Commercial use of Fish Audio S2-Pro or derivative works requires a separate license
+from Fish Audio.

README.md

@@ -1,3 +1,114 @@
 ---
-license: mit
+license: other
+license_name: fish-audio-research-license
+license_link: https://huggingface.co/fishaudio/s2-pro/blob/main/LICENSE.md
+base_model: fishaudio/s2-pro
+pipeline_tag: text-to-speech
+library_name: sglang
+tags:
+- text-to-speech
+- fish-audio
+- s2-pro
+- sglang
+- realtime
+- rtx-5090
+- no-weights
+inference: false
 ---
+
+# Fish Audio S2-Pro Realtime Optimized for RTX 5090
+
+This repository is a model-card-only release for an RTX 5090 local serving profile for
+[Fish Audio S2-Pro](https://huggingface.co/fishaudio/s2-pro).
+
+It does **not** redistribute Fish Audio S2-Pro weights. The underlying model files are
+the official upstream `fishaudio/s2-pro` checkpoint files, so users should accept the
+upstream license and download the weights directly from Fish Audio.
+
+## What This Is
+
+This release documents a local realtime optimization profile:
+
+- Base model: `fishaudio/s2-pro`
+- Target GPU: NVIDIA GeForce RTX 5090, 32 GB VRAM
+- Serving path: Docker-backed SGLang Omni
+- Intended use: single-user realtime local TTS
+- Example workstation split: RTX 5090 handles TTS; a separate RTX 3090 can handle realtime STT/ASR
+
+No fine-tuning, quantization, or checkpoint conversion is claimed here. The optimization
+work is in the serving/runtime profile around the official S2-Pro weights.
+
+## Why There Are No Weights Here
+
+The local checkpoint files match the official upstream `fishaudio/s2-pro` release. To
+avoid duplicating gated model files or confusing the license boundary, this repository
+does not upload:
+
+- `model-00001-of-00002.safetensors`
+- `model-00002-of-00002.safetensors`
+- `codec.pth`
+- tokenizer/config files from the upstream model
+
+Download those files from:
+
+```text
+https://huggingface.co/fishaudio/s2-pro
+```
+
+## Local Benchmark Summary
+
+Measured locally on an RTX 5090. Lower is better.
+
+![RTX 5090 benchmark bar chart](assets/rtx5090-benchmark-bars.svg)
+
+| Metric | Before: Python Fish server | After: SGLang Omni + cached reference | Change |
+|---|---:|---:|---:|
+| First audio | 25.1s | 0.36s | 98.6% lower latency |
+| Total request time | 25.1s | 2.10s | 91.6% lower latency |
+| Estimated RTF | 5.51 | 0.48 | 91.3% lower RTF |
+
+Benchmark notes:
+
+- Baseline was the local Python Fish server path.
+- Optimized path was the Docker-backed SGLang Omni path with a warm server and cached reference voice.
+- Warm cached SGLang generation was below realtime for the measured short live TTS sample.
+- The chart is a local serving benchmark, not an upstream Fish Audio benchmark claim.
+
+## Optimization Profile
+
+The measured realtime profile used:
+
+- SGLang Omni instead of the eager Python Fish server path.
+- The Docker image's pinned Torch/SGLang/FlashInfer stack.
+- SGLang CUDA graph replay enabled.
+- RTX 5090 / SM120-safe Fish audio-decoder path by disabling the incompatible `sgl-kernel` KV-cache flash-attention path.
+- Graph-safe fixed-cache SDPA fallback for the Fish audio decoder.
+- `flashinfer` text attention backend.
+- Single-user live memory profile:
+  - `mem_fraction_static=0.50`
+  - `chunked_prefill_size=2048`
+  - `max_running_requests=4`
+- Preloaded/cached reference VQ codes for repeated voice-reference requests.
+- Docker model/runtime volumes to avoid repeated slow checkpoint reads through Windows `/mnt/d` bind mounts.
+
+Measured local VRAM after the tuned live restart was about 24.6 GB on the RTX 5090,
+down from an earlier near-full 32.2 GB SGLang container reservation.
+
+## License
+
+The base model is governed by the
+[Fish Audio Research License](https://huggingface.co/fishaudio/s2-pro/blob/main/LICENSE.md).
+Research and non-commercial use are permitted by Fish Audio under that license.
+Commercial use requires a separate license from Fish Audio.
+
+This repository does not grant additional rights to the Fish Audio model weights.
+
+## Attribution
+
+Built with Fish Audio S2-Pro. Fish Audio S2-Pro is developed by Fish Audio / 39 AI, INC.
+
+Upstream model:
+
+```text
+fishaudio/s2-pro
+```