s2-pro-BnB-4Bits

S2-Pro NF4

<!-- polished-overview:start -->

Overview

s2-pro-BnB-4Bits is a low-bit checkpoint prepared for memory-efficient inference, published by groxaxo. It is intended for open-source evaluation, reproducible experimentation, and compatible local or hosted inference workflows. The wording below is deliberately limited to what can be verified from this repository's metadata and artifacts.

At a glance

What is included

• *.pth (2 files)
• config.json
• tokenizer.json
• tokenizer_config.json
• chat_template.jinja
• LICENSE.md
• Additional configuration, tokenizer, processor, or shard files (9 visible artifacts total)

Quick start

Transformers + BitsAndBytes

Use a current Transformers release with BitsAndBytes support, keeping the tokenizer and model configuration files from this repository alongside the checkpoint.

Compatibility and responsible use

• Use a runtime that explicitly supports this format, architecture, and modality.
• Keep configuration, tokenizer, processor, projection, and weight files from the same revision together.
• Review the source model card and license before redistribution or deployment.
• Hardware needs depend on parameter count, context length, cache precision, quantization, and concurrency.
• Report reproducible issues with the runtime version, hardware, launch command, and a minimal example.

Quantization or conversion changes numerical behavior, memory use, and throughput relative to the source checkpoint; validate quality on your own workload.

Generated outputs may be inaccurate or unsuitable for a given use case. Users are responsible for testing behavior, applying appropriate safeguards, and complying with applicable licenses and laws.

<!-- polished-overview:end --> <img src="overview.png" alt="S2-Pro overview" width="100%">

GitHub Fork | Upstream Fish Speech | Technical Report | Fish Audio

This repository hosts the Groxaxo NF4 release of Fish Audio S2-Pro for lower-VRAM inference.

• Base model: Fish Audio S2-Pro
• Relation: Quantized release
• Format: bitsandbytes NF4 prequantized model.pth
• Target hardware: practical single-GPU inference on 12 GB+ VRAM setups
• Best paired with: groxaxo/fish-speech-int4-patch

This is a community-hosted release of the original Fish Audio model. Credit for the base model, research, and architecture belongs to the Fish Audio team.

Huge thanks to the original creators at Fish Audio and the upstream fishaudio/fish-speech project for building and open-sourcing S2-Pro.

If this NF4 release helps you, please star the companion GitHub project here:

https://github.com/groxaxo/fish-speech-int4-patch

The goal is simple: make the flagship S2-Pro experience easier to run, easier to share, and easier to deploy on real-world single-GPU machines.

What is in this repo

• model.pth: prequantized NF4 checkpoint
• codec.pth: codec weights
• tokenizer/config assets needed by the patched loader

The checkpoint is meant to be loaded through the fork's bnb4 path. It is not a legacy int4 or int8 export.

Recommended usage

Use the patched repo that defaults to the right settings for this checkpoint:

git clone https://github.com/groxaxo/fish-speech-int4-patch
cd fish-speech-int4-patch

./install_bnb4_3060.sh
./start_bnb4_3060.sh

That path starts the API/WebUI with the intended defaults:

• --bnb4
• --half
• lazy loading
• s2-pro as the canonical model name

Why people use this release

• lower-VRAM NF4 deployment path for S2-Pro
• companion GitHub fork with API, WebUI, Docker, and export tooling
• smoke-tested prequantized model.pth reload support
• clearer self-hosting path for 12 GB and 24 GB cards

Quick commands

WebUI

git clone https://github.com/groxaxo/fish-speech-int4-patch
cd fish-speech-int4-patch
./install_bnb4_3060.sh
./start_bnb4_3060.sh

API server

PYTHONPATH=. python tools/api_server.py \
  --checkpoint-path /path/to/s2-pro \
  --bnb4 \
  --half \
  --host 0.0.0.0 \
  --port 8880

OpenAI-style request

curl http://127.0.0.1:8880/v1/audio/speech \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "s2-pro",
    "input": "[warm, calm] Hello from the Groxaxo NF4 S2-Pro release.",
    "voice": "default"
  }' \
  --output speech.wav

Manual loading

If you want to point the repo at this checkpoint directly, keep --bnb4 enabled:

PYTHONPATH=. python tools/api_server.py \
  --checkpoint-path /path/to/s2-pro \
  --bnb4 \
  --half

Or in Python:

import torch
from fish_speech.models.text2semantic.inference import init_model

model, decode_one_token = init_model(
    checkpoint_path="/path/to/s2-pro",
    device="cuda:0",
    precision=torch.float16,
    compile=False,
    bnb4=True,
)

Why this release exists

Upstream S2-Pro is excellent, but many single-card workstations do not have enough VRAM for a comfortable default setup. This NF4 release makes S2-Pro much easier to run on common cards like the RTX 3060 while preserving the flagship model path.

Model notes

• S2-Pro uses a Dual-Autoregressive architecture with a 4B slow AR stack and a fast residual AR stack.
• It supports fine-grained inline control with natural-language tags such as [whisper], [laugh], and [sad].
• It supports multilingual generation, multi-speaker prompting, and strong voice cloning workflows.

Prompt examples

[whisper] We need to leave quietly before sunrise.
[excited] We actually got it working on a 12 GB card.
[sad] I waited for you at the station all night.

Links

• Groxaxo fork README
• Star the GitHub project
• GitHub issues and feature requests
• Fish Audio blog post
• Fish Audio S2 technical report

License

This model remains under the Fish Audio Research License. Research and non-commercial use is permitted under that license. Commercial use requires a separate agreement with Fish Audio.