百川Dynamic NTK-ALiBi的代码实现：无需微调即可推理更长文本

---

前言

NTK-ALiBi原理：NTK-ALiBi：通过插值实现大模型ALiBi位置编码的长文本外推

代码实现

打开百川模型文件夹中的modeling_baichuan.py

1、增加build_dynamically_alibi_tensor函数：

def build_dynamically_alibi_tensor(num_heads, max_pos) -> torch.Tensor:
    """Psuedo code for Dynamic NTK-ALiBi."""

    # dynamic ntk factor according to actual sequence length
    a0 = 1.0
    train_seq_len = 4096
    a = a0 * torch.tensor(max_pos) / train_seq_len  # [batch, 1]
    a = a.masked_fill(a < 1.0, 1.0)  # dynamic step 1: dynamic ntk scaling factor

    scale = a ** (1.0 / (num_heads-1))  # dynamic step 2: coefficient b, for computation convenience
    closest_power_of_2 = 2 ** math.floor(math.log2(num_heads))
    base = torch.tensor(
        2 ** (-(2