drbaph / s2-pro-fp8

Fish Audio S2 Pro — FP8 Weight-Only Quantized

FP8 quantized version of fishaudio/s2-pro .

Original Model | Technical Report | GitHub | Playground | ComfyUI Node

---

What is this?

This is a weight-only FP8 quantization of Fish Audio S2 Pro — a state-of-the-art open-source TTS model with fine-grained inline prosody and emotion control across 80+ languages. The quantization cuts the on-disk size roughly in half and reduces VRAM usage from ~24 GB to ~12 GB, with no perceptible quality loss in practice.

---

Quantization Details

What is quantized: All nn.Linear weight matrices in both the Slow AR (4B) and Fast AR (400M) backbones — 201 layers in total. Non-linear weights (embeddings, layer norms, codec) remain in bfloat16.

Method: Per-row symmetric FP8

Each output row of every weight matrix has its own float32 scale factor:

scale = max(abs(row)) / FP8_MAX        # FP8_MAX = 448.0 for float8_e4m3fn
W_fp8 = round(W_bf16 / scale)          # quantize
W_bf16 = W_fp8.to(bfloat16) * scale   # dequantize at inference

Per-row scaling captures the per-channel magnitude variation in transformer weight matrices much better than a single per-tensor scale, significantly reducing quantization error at minimal overhead.

No external quantization library required. Dequantization is implemented in pure PyTorch inside a custom FP8Linear module — no torchao, bitsandbytes, or AutoGPTQ needed. The model loads and runs on any machine with PyTorch 2.1+.

File layout inside model.safetensors :

• <layer>.weight — float8_e4m3fn tensor (quantized weights)
• <layer>.weight.scale — float32 tensor, shape [out_features, 1] (per-row scales)
• All other tensors — bfloat16 (embeddings, norms, codec, etc.)

---

Hardware Requirements

• GPU: NVIDIA GPU with CUDA support
• VRAM: ~12 GB
• Native FP8 tensor cores: Ada Lovelace or Blackwell (RTX 4090, RTX 5090, H100, etc.) — recommended for full speed
• Older GPUs (Ampere and below): Will load and run correctly. Dequantization to bfloat16 happens on all hardware, so you still get the ~12 GB VRAM footprint benefit even without native FP8 cores.

---

Usage — ComfyUI (Recommended)

The easiest way to use this model is with ComfyUI-FishAudioS2 , which has native support for this FP8 model with zero extra setup.

Installation

1. Install the ComfyUI node via ComfyUI Manager (search FishAudioS2 ) or manually:

   cd ComfyUI/custom_nodes
   git clone https://github.com/Saganaki22/ComfyUI-FishAudioS2.git
   

2. The model auto-downloads on first use — select s2-pro-fp8 from the model dropdown in any Fish S2 node.

3. Or download manually:

   huggingface-cli download drbaph/s2-pro-fp8 --local-dir ComfyUI/models/fishaudioS2/s2-pro-fp8
   

Recommended settings

• precision : auto or bfloat16 — matches the activation dtype
• attention : auto or sage_attention for best performance
• keep_model_loaded : True if running multiple generations back-to-back

Works with all three nodes: Fish S2 TTS , Fish S2 Voice Clone TTS , Fish S2 Multi-Speaker TTS .

---

About Fish Audio S2 Pro

Fish Audio S2 Pro is a leading text-to-speech model with fine-grained inline control of prosody and emotion. Trained on over 10M+ hours of audio data across 80+ languages, it combines reinforcement learning alignment with a Dual-Autoregressive (Dual-AR) architecture.

Architecture

S2 Pro builds on a decoder-only transformer combined with an RVQ-based audio codec (10 codebooks, ~21 Hz frame rate):

• Slow AR (4B parameters): Operates along the time axis and predicts the primary semantic codebook.
• Fast AR (400M parameters): Generates the remaining 9 residual codebooks at each time step, reconstructing fine-grained acoustic detail.

This asymmetric design keeps inference efficient while preserving audio fidelity. The Dual-AR architecture is structurally isomorphic to standard autoregressive LLMs, inheriting LLM-native serving optimizations — continuous batching, paged KV cache, CUDA graph replay, RadixAttention-based prefix caching.

Fine-Grained Inline Control

Embed natural-language instructions directly in the text using [tag] syntax. S2 Pro accepts free-form descriptions — not a fixed tag vocabulary:

[pause] [emphasis] [laughing] [inhale] [chuckle] [tsk] [singing] [excited] [volume up] [echo] [angry] [sigh] [whisper] [screaming] [shouting] [surprised] [short pause] [exhale] [delight] [sad] [clearing throat] [shocked] [with strong accent] [professional broadcast tone] [pitch up] [pitch down]

Free-form examples: [whisper in small voice] · [super happy and excited] · [speaking slowly and clearly] · [sarcastic tone]

15,000+ unique tags supported.

Supported Languages

Tier 1 (Best Quality): Japanese (ja), English (en), Chinese (zh)

Tier 2: Korean (ko), Spanish (es), Portuguese (pt), Arabic (ar), Russian (ru), French (fr), German (de)

80+ total: sv, it, tr, no, nl, cy, eu, ca, da, gl, ta, hu, fi, pl, et, hi, la, ur, th, vi, jw, bn, yo, sl, cs, sw, nn, he, ms, uk, id, kk, bg, lv, my, tl, sk, ne, fa, af, el, bo, hr, ro, sn, mi, yi, am, be, km, is, az, sd, br, sq, ps, mn, ht, ml, sr, sa, te, ka, bs, pa, lt, kn, si, hy, mr, as, gu, fo

Production Streaming Performance (original model, H200)

• Real-Time Factor (RTF): 0.195
• Time-to-first-audio: ~100 ms
• Throughput: 3,000+ acoustic tokens/s while maintaining RTF below 0.5

---

Citation

@misc{liao2026fishaudios2technical,
      title={Fish Audio S2 Technical Report}, 
      author={Shijia Liao and Yuxuan Wang and Songting Liu and Yifan Cheng and Ruoyi Zhang and Tianyu Li and Shidong Li and Yisheng Zheng and Xingwei Liu and Qingzheng Wang and Zhizhuo Zhou and Jiahua Liu and Xin Chen and Dawei Han},
      year={2026},
      eprint={2603.08823},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2603.08823},
}

---

License

This model inherits the Fish Audio Research License from fishaudio/s2-pro . Research and non-commercial use is permitted free of charge. Commercial use requires a separate license from Fish Audio — contact [email protected] .

The FP8 quantization was produced by drbaph and is released under the same license.