Swagcrew commited on
Commit
dbde30e
·
verified ·
1 Parent(s): 7bdebfa

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +38 -0
README.md ADDED
@@ -0,0 +1,38 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # 🐟 Fish Speech S2 Pro — Quantization Experiments
2
+
3
+ Multi-phase quantization of Fish Audio S2 Pro (4.56B params) on A100-SXM4-80GB.
4
+
5
+ ## Results Summary
6
+
7
+ | Phase | Method | Target | Disk Size (MB) | Compression |
8
+ |-------|--------|--------|----------------|-------------|
9
+ | 1a | FP8 (per-row) | Slow AR | 5240 | 4.91x |
10
+ | 1b | INT4 (group=128) | Slow AR | 5344 | 4.91x |
11
+ | 2a | INT4 (group=128) | All layers | 4962 | 8.87x |
12
+ | 2b | INT8 (per-row) | Slow AR | 5240 | 4.91x |
13
+ | 2c | INT3 (group=128) | Slow AR | 5344 | 4.91x |
14
+ | 3a | INT2 (group=64) | Slow AR | 5453 | 4.91x |
15
+ | 3b | INT2 (group=64) | All layers | 5082 | 8.87x |
16
+
17
+ **Note:** Disk sizes with safetensors format show similar sizes because int8/int4/int3/int2
18
+ values are stored in int8 dtype tensors (safetensors doesn't support sub-byte packing).
19
+ Actual memory usage during inference will differ significantly.
20
+
21
+ ## Baseline: BF16 (Original)
22
+ - **fishaudio/s2-pro**: ~10.8 GB total (model + codec)
23
+ - **drbaph/s2-pro-fp8**: ~6.2 GB (FP8 proven, zero quality loss)
24
+
25
+ ## Architecture
26
+ - **Slow AR**: 36 layers, dim 2560, 32 heads (GQA), ~4.0B params
27
+ - **Fast AR**: 4 layers, dim 2560, 32 heads (GQA), ~0.4B params
28
+ - **Codec**: DAC RVQ, 10 codebooks, 44.1kHz
29
+
30
+ ## Citation
31
+ ```bibtex
32
+ @misc{liao2026fishaudios2technical,
33
+ title={Fish Audio S2 Technical Report},
34
+ author={Shijia Liao and Yuxuan Wang and others},
35
+ year={2026}, eprint={2603.08823},
36
+ archivePrefix={arXiv}, primaryClass={cs.SD},
37
+ }
38
+ ```