import numpy as np


def softmax(x, axis=-1):
    """Softmax激活函数，用于将 logits 转换为概率分布"""
    # 减去最大值防止数值溢出
    e_x = np.exp(x - np.max(x, axis=axis, keepdims=True))
    return e_x / np.sum(e_x, axis=axis, keepdims=True)


def relu(x):
    """ReLU激活函数"""
    return np.maximum(0, x)


class Linear:
    """全连接层（线性层）"""
    def __init__(self, in_features, out_features):
        # 使用He初始化权重
        self.weight = np.random.randn(in_features, out_features) * np.sqrt(2.0 / in_features)
        # 偏置初始化为0
        self.bias = np.zeros(out_features)
        
    def forward(self, x):
        """前向传播：y = xW + b"""
        return x @ self.weight + self.bias


class LayerNorm:
    """层归一化（Layer Normalization）"""
    def __init__(self, d_model, eps=1e-6):
        # 缩放参数
        self.gamma = np.ones(d_model)
        # 偏移参数
        self.beta = np.zeros(d_model)
        # 防止除零的小常数
        self.eps = eps
        
    def forward(self, x):
        """前向传播：对最后一个维度进行归一化"""
        mean = np.mean(x, axis=-1, keepdims=True)
        std = np.std(x, axis=-1, keepdims=True)
        return self.gamma * (x - mean) / (std + self.eps) + self.beta


class MultiHeadAttention:
    """多头注意力机制（支持KV cache）"""
    def __init__(self, d_model, num_heads):
        self.d_model = d_model  # 模型维度
        self.num_heads = num_heads  # 注意力头数
        self.d_k = d_model // num_heads  # 每个头的维度
        
        # 定义Q、K、V的线性变换层
        self.W_q = Linear(d_model, d_model)  # Query变换
        self.W_k = Linear(d_model, d_model)  # Key变换
        self.W_v = Linear(d_model, d_model)  # Value变换
        self.W_o = Linear(d_model, d_model)  # 输出变换
        
        # KV cache：用于推理时缓存历史Key和Value
        self.k_cache = None  # 缓存的Key [batch_size, num_heads, seq_len, d_k]
        self.v_cache = None  # 缓存的Value [batch_size, num_heads, seq_len, d_k]
        
    def clear_cache(self):
        """清除KV cache"""
        self.k_cache = None
        self.v_cache = None
        
    def forward(self, Q, K, V, mask=None, use_cache=False):
        """
        前向传播
        Q: Query张量 [batch_size, seq_len, d_model]
        K: Key张量 [batch_size, seq_len, d_model]
        V: Value张量 [batch_size, seq_len, d_model]
        mask: 注意力掩码
        use_cache: 是否使用KV cache（推理时设为True）
        """
        batch_size = Q.shape[0]
        seq_len = Q.shape[1]
        
        # 线性变换并分割成多头
        Q = self.W_q.forward(Q).reshape(batch_size, seq_len, self.num_heads, self.d_k).transpose(0, 2, 1, 3)
        K = self.W_k.forward(K).reshape(batch_size, seq_len, self.num_heads, self.d_k).transpose(0, 2, 1, 3)
        V = self.W_v.forward(V).reshape(batch_size, seq_len, self.num_heads, self.d_k).transpose(0, 2, 1, 3)
        
        # KV cache逻辑
        if use_cache:
            if self.k_cache is None:
                # Prefill阶段：首次计算，缓存完整的K和V
                self.k_cache = K
                self.v_cache = V
            else:
                # Decode阶段：拼接历史缓存和新的K、V
                self.k_cache = np.concatenate([self.k_cache, K], axis=2)
                self.v_cache = np.concatenate([self.v_cache, V], axis=2)
                # 使用完整的K和V进行注意力计算
                K = self.k_cache
                V = self.v_cache
        
        # 计算缩放点积注意力分数
        # Q: [batch, heads, q_len, d_k]
        # K: [batch, heads, kv_len, d_k]（decode时kv_len > q_len）
        attn_scores = Q @ K.transpose(0, 1, 3, 2) / np.sqrt(self.d_k)
        
        # 应用掩码（如果提供）
        if mask is not None:
            attn_scores = np.where(mask == 0, -1e9, attn_scores)
        
        # Softmax得到注意力权重
        attn_probs = softmax(attn_scores, axis=-1)
        
        # 加权求和
        attn_output = attn_probs @ V
        
        # 拼接多头输出
        attn_output = attn_output.transpose(0, 2, 1, 3).reshape(batch_size, seq_len, self.d_model)
        
        # 最终线性变换
        output = self.W_o.forward(attn_output)
        return output


class FeedForward:
    """前馈神经网络（两层全连接网络）"""
    def __init__(self, d_model, d_ff):
        # 第一层：扩展维度
        self.linear1 = Linear(d_model, d_ff)
        # 第二层：恢复维度
        self.linear2 = Linear(d_ff, d_model)
        
    def forward(self, x):
        """前向传播：线性 -> ReLU -> 线性"""
        return self.linear2.forward(relu(self.linear1.forward(x)))


class TransformerBlock:
    """Transformer编码器块"""
    def __init__(self, d_model, num_heads, d_ff):
        # 多头自注意力层
        self.attention = MultiHeadAttention(d_model, num_heads)
        # 前馈网络层
        self.feed_forward = FeedForward(d_model, d_ff)
        # 两个层归一化
        self.norm1 = LayerNorm(d_model)
        self.norm2 = LayerNorm(d_model)
        
    def forward(self, x, mask=None, use_cache=False):
        """
        前向传播（残差连接 + 层归一化）
        1. 自注意力 -> 残差连接 -> 层归一化
        2. 前馈网络 -> 残差连接 -> 层归一化
        x: 输入张量 [batch_size, seq_len, d_model]
        mask: 注意力掩码
        use_cache: 是否使用KV cache
        """
        # 自注意力子层（Q=K=V，自注意力）
        attn_output = self.attention.forward(x, x, x, mask, use_cache=use_cache)
        x = self.norm1.forward(x + attn_output)
        
        # 前馈网络子层
        ff_output = self.feed_forward.forward(x)
        x = self.norm2.forward(x + ff_output)
        return x
    
    def clear_cache(self):
        """清除该层的KV cache"""
        self.attention.clear_cache()


class SimpleTransformer:
    """简化的Transformer编码器模型（支持prefill/decode分离）"""
    def __init__(self, vocab_size, d_model, num_heads, num_layers, d_ff, max_seq_len):
        self.d_model = d_model
        self.vocab_size = vocab_size
        # 词嵌入层
        self.embedding = np.random.randn(vocab_size, d_model) * 0.01
        # 位置编码
        self.positional_encoding = self._create_positional_encoding(max_seq_len, d_model)
        # Transformer编码器层堆叠
        self.transformer_blocks = [
            TransformerBlock(d_model, num_heads, d_ff) for _ in range(num_layers)
        ]
        
    def _create_positional_encoding(self, max_seq_len, d_model):
        """
        创建正弦余弦位置编码
        使用不同频率的正弦和余弦函数生成位置信息
        """
        pe = np.zeros((max_seq_len, d_model))
        position = np.arange(0, max_seq_len).reshape(-1, 1).astype(float)
        div_term = np.exp(np.arange(0, d_model, 2).astype(float) * -(np.log(10000.0) / d_model))
        pe[:, 0::2] = np.sin(position * div_term)  # 偶数维度用正弦
        pe[:, 1::2] = np.cos(position * div_term)  # 奇数维度用余弦
        return pe
    
    def clear_cache(self):
        """清除所有层的KV cache"""
        for block in self.transformer_blocks:
            block.clear_cache()
        
    def prefill(self, x, mask=None):
        """
        Prefill阶段：处理完整的输入序列
        x: 输入序列 [batch_size, seq_len]
        mask: 注意力掩码
        返回: 输出张量 [batch_size, seq_len, d_model]
        """
        seq_len = x.shape[1]
        
        # 词嵌入 + 缩放
        x = self.embedding[x] * np.sqrt(self.d_model)
        
        # 加上位置编码
        x = x + self.positional_encoding[:seq_len, :]
        
        # 通过所有Transformer块，使用KV cache
        for block in self.transformer_blocks:
            x = block.forward(x, mask, use_cache=True)
            
        return x
    
    def decode(self, x, position):
        """
        Decode阶段：处理单个新token
        x: 新token [batch_size, 1]
        position: 当前token在序列中的位置
        返回: 输出张量 [batch_size, 1, d_model]
        """
        # 词嵌入 + 缩放
        x = self.embedding[x] * np.sqrt(self.d_model)
        
        # 加上位置编码（使用当前位置）
        x = x + self.positional_encoding[position:position+1, :]
        
        # 通过所有Transformer块，使用KV cache
        for block in self.transformer_blocks:
            x = block.forward(x, use_cache=True)
            
        return x
    
    def forward(self, x, mask=None):
        """
        前向传播（不使用KV cache，用于训练）
        x: 输入序列 [batch_size, seq_len]
        mask: 注意力掩码
        """
        seq_len = x.shape[1]
        
        # 词嵌入 + 缩放
        x = self.embedding[x] * np.sqrt(self.d_model)
        
        # 加上位置编码
        x = x + self.positional_encoding[:seq_len, :]
        
        # 通过所有Transformer块
        for block in self.transformer_blocks:
            x = block.forward(x, mask)
            
        return x
    
    def count_parameters(self):
        """统计模型参数数量"""
        count = 0
        count += self.embedding.size  # 嵌入层参数
        for block in self.transformer_blocks:
            # 注意力层参数
            count += block.attention.W_q.weight.size + block.attention.W_q.bias.size
            count += block.attention.W_k.weight.size + block.attention.W_k.bias.size
            count += block.attention.W_v.weight.size + block.attention.W_v.bias.size
            count += block.attention.W_o.weight.size + block.attention.W_o.bias.size
            # 前馈网络参数
            count += block.feed_forward.linear1.weight.size + block.feed_forward.linear1.bias.size
            count += block.feed_forward.linear2.weight.size + block.feed_forward.linear2.bias.size
            # 层归一化参数
            count += block.norm1.gamma.size + block.norm1.beta.size
            count += block.norm2.gamma.size + block.norm2.beta.size
        return count


def create_padding_mask(seq, pad_idx=0):
    """
    创建填充掩码
    用于忽略序列中的填充位置（padding tokens）
    """
    return (seq != pad_idx).astype(float)[:, np.newaxis, np.newaxis, :]


def create_look_ahead_mask(size):
    """
    创建前瞻掩码（下三角掩码）
    用于解码器中，防止位置i看到i之后的信息
    """
    mask = np.triu(np.ones((size, size)), k=1).astype(bool)
    return (~mask).astype(float)


class InferenceEngine:
    """推理引擎：管理prefill和decode阶段"""
    def __init__(self, model):
        """
        初始化推理引擎
        model: SimpleTransformer模型实例
        """
        self.model = model
        self.generated_tokens = []  # 已生成的token列表
        self.current_position = 0   # 当前位置
        
    def reset(self):
        """重置推理状态"""
        self.model.clear_cache()
        self.generated_tokens = []
        self.current_position = 0
        
    def prefill(self, input_ids):
        """
        Prefill阶段：处理完整的输入序列
        input_ids: 输入token序列 [batch_size, seq_len]
        返回: 下一个token的logits [batch_size, vocab_size]
        """
        self.reset()
        batch_size = input_ids.shape[0]
        seq_len = input_ids.shape[1]
        
        # 创建前瞻掩码（防止看到未来信息）
        mask = create_look_ahead_mask(seq_len)
        
        # 执行prefill前向传播（使用KV cache）
        output = self.model.prefill(input_ids, mask)
        
        # 获取最后一个位置的输出（用于预测下一个token）
        last_output = output[:, -1, :]  # [batch_size, d_model]
        
        # 简单的logits计算：使用嵌入矩阵的转置作为输出投影
        # logits = last_output @ embedding.T
        logits = last_output @ self.model.embedding.T  # [batch_size, vocab_size]
        
        # 更新状态
        self.current_position = seq_len
        
        return logits
    
    def decode_step(self, input_token):
        """
        Decode阶段：处理单个新token
        input_token: 新token [batch_size, 1]
        返回: 下一个token的logits [batch_size, vocab_size]
        """
        # 执行decode前向传播（使用KV cache）
        output = self.model.decode(input_token, self.current_position)
        
        # 获取输出（单token，所以直接取[:, 0, :]）
        last_output = output[:, 0, :]  # [batch_size, d_model]
        
        # 计算logits
        logits = last_output @ self.model.embedding.T  # [batch_size, vocab_size]
        
        # 更新位置
        self.current_position += 1
        
        return logits
    
    def generate(self, input_ids, max_new_tokens=50, temperature=1.0):
        """
        自回归生成文本
        input_ids: 输入token序列 [batch_size, seq_len]
        max_new_tokens: 最大生成token数
        temperature: 温度参数（控制随机性）
        返回: 完整的生成序列 [batch_size, seq_len + max_new_tokens]
        """
        batch_size = input_ids.shape[0]
        
        # 保存原始输入
        generated = input_ids.copy()
        
        # Prefill阶段
        logits = self.prefill(input_ids)
        
        # 采样第一个生成的token
        next_token = self._sample(logits, temperature)
        generated = np.concatenate([generated, next_token], axis=1)
        
        # Decode阶段：逐个生成token
        for i in range(max_new_tokens - 1):
            logits = self.decode_step(next_token)
            next_token = self._sample(logits, temperature)
            generated = np.concatenate([generated, next_token], axis=1)
            
            # 检查是否生成了结束符（这里用0作为结束符）
            if np.all(next_token == 0):
                break
                
        return generated
    
    def _sample(self, logits, temperature=1.0):
        """
        从logits中采样
        logits: [batch_size, vocab_size]
        temperature: 温度参数
        返回: 采样的token [batch_size, 1]
        """
        # 应用温度缩放
        logits = logits / temperature
        
        # 计算概率分布
        probs = softmax(logits, axis=-1)
        
        # 按概率采样
        batch_size = logits.shape[0]
        next_tokens = np.zeros((batch_size, 1), dtype=int)
        
        for i in range(batch_size):
            # 使用numpy的random.choice采样
            next_tokens[i, 0] = np.random.choice(logits.shape[1], p=probs[i])
            
        return next_tokens


if __name__ == "__main__":
    # 模型超参数
    vocab_size = 1000   # 词汇表大小
    d_model = 512       # 模型维度
    num_heads = 8       # 注意力头数
    num_layers = 6      # Transformer层数
    d_ff = 2048         # 前馈网络隐藏层维度
    max_seq_len = 100   # 最大序列长度
    
    # 创建模型
    model = SimpleTransformer(vocab_size, d_model, num_heads, num_layers, d_ff, max_seq_len)
    
    # 创建随机输入
    batch_size = 2
    seq_len = 10
    x = np.random.randint(0, vocab_size, (batch_size, seq_len))
    
    # 前向传播
    output = model.forward(x)
    print(f"输入形状: {x.shape}")
    print(f"输出形状: {output.shape}")
    print(f"模型参数量: {model.count_parameters():,}")