【大规模语言模型：从理论到实践】Transformer中MultiHeadAttention详解

class MultiHeadAttention(nn.Module):
    def __init__(self, heads, d_model, dropout=0.1):
        super().__init__()
        self.d_model = d_model
        self.d_k = d_model // heads  # 每个头的维度，d_model 除以 heads
        self.h = heads  # 多头的数量
        
        # 定义线性变换层，用于将输入分别映射到 q, k, v 空间
        self.q_linear = nn.Linear(d_model, d_model)
        self.v_linear = nn.Linear(d_model, d_model)
        self.k_linear = nn.Linear(d_model, d_model)
        
        # 定义 dropout 层，用于