fish-speech-s2-pro-fp8 / quantization_info.json
AEmotionStudio's picture
Upload quantization_info.json with huggingface_hub
264ed96 verified
raw
history blame contribute delete
887 Bytes
{
"quantization_method": "per_row_symmetric_fp8",
"weight_dtype": "float8_e4m3fn",
"scale_dtype": "float32",
"scale_granularity": "per_row",
"activation_dtype": "bfloat16",
"torch_version": "2.10.0+cu128",
"source_model": "fishaudio/s2-pro",
"total_params_B": 4.562,
"fp8_linear_params_B": 4.048,
"bf16_other_params_B": 0.514,
"output_size_GB": 4.73,
"linear_layers_quantized": 201,
"key_format": {
"<layer_name>": "float8_e4m3fn quantized weight",
"<layer_name>.scale": "float32 per-row dequantization scale",
"_buf.<name>": "bf16/fp32 buffer (freqs_cis, causal_mask, etc.)",
"other": "bfloat16 (embeddings, norms, non-linear layers)"
},
"notes": "All nn.Linear weights are float8_e4m3fn with per-row scales. Activations are bfloat16 (weight-only quantization). codec.pth is unchanged bfloat16. Created by AEmotionStudio/ComfyUI-FFMPEGA."
}