Function quant_weight_fp16

utils/preprocess-huggingface-bitnet.py:5–9

(weight)

import torch

def quant_weight_fp16(weight):
    weight = weight.to(torch.float)
    s = 1.0 / weight.abs().mean().clamp_(min=1e-5)
    new_weight = (weight * s).round().clamp(-1, 1) / s
    return new_weight

def quant_model(input, output):
    tensors = {}