s2-pro-fp8 / quantization_info.json
drbaph's picture
Upload 10 files
d3a2cd8 verified
{
"quantization_method": "torchao_Float8WeightOnly",
"weight_dtype": "float8_e4m3fn",
"scale_dtype": "float32",
"scale_granularity": "per_row",
"activation_dtype": "bfloat16",
"torchao_version": "0.16.0",
"torch_version": "2.9.0+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": 6.16,
"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)"
},
"inference_requirements": {
"torchao": ">= 0.8.0",
"compute_capability": ">= 8.9 (RTX 4090 / 5090) for native FP8 matmuls"
},
"notes": "All nn.Linear weights are float8_e4m3fn. Activations are bfloat16 (weight-only quantization). codec.pth is unchanged bfloat16."
}