逐行拆解多头自注意力和训练代码

只是写法示意，不是完整程序。

# ANS 温度缩放与 Top-K 概率分布采样核心实现 (ans-mini-gpt.py)
import math
def ans_top_k_softmax(logits: list, temp: float = 0.8, k: int = 5) -> list:
    scaled = [x / max(0.05, temp) for x in logits]
    cutoff = sorted(scaled, reverse=True)[min(k - 1, len(scaled) - 1)]
    exps = [math.exp(x - cutoff) if x >= cutoff else 0.0 for x in scaled]
    s = sum(exps) or 1.0
    return [round(e / s, 4) for e in exps]