一、前言

本文将介绍ChatGLM-6B的模型结构组件源码。

代练链接：https://huggingface.co/THUDM/chatglm-6b/blob/main/modeling_chatglm.py

二、激活函数

@torch.jit.script
def gelu_impl(x):
    """OpenAI's gelu implementation."""
    return 0.5 * x * (1.0 + torch.tanh(0.7978845608028654 * x *
                                       (1.0 + 0.044715 * x * x)))


def gelu(x):
    return gelu_impl(x)

三、位置编码

3.1、RoPE原理简介

ChatGLM-6B的位置编码采用的旋转位置编码（详细推导过程见原文：ROPE），简单来说其目的就是构建一个包含相对位置信息的Attention矩阵，其公式如下：
$${\left({\mathcal{R}}_m\mathbf{q}\right)}^{\top }\left({\mathcal{R}}_n\mathbf{k}\right)={\mathbf{q}}^{\top }{\mathcal{R}}_m^{\top }{\mathcal{R}}_n\mathbf{k}={\mathbf{q}}^{\top }{\mathcal{R}}_{n-m}\mathbf{k}$$
式中，$q$、$k$分别表示注意力机制中的query和key，$m$、$n$分别表示两个位置，$R_i$表示位置i处处理的矩阵，其中$R$的形式为：
$$\left(\begin{array}{ccccccc}\cos m{\theta }_0& -\sin m{\theta }_0& 0& 0& \cdots & 0& 0\\ \sin m{\theta }_0& \cos m{\theta }_0& 0& 0& \cdots & 0& 0\\ 0& 0& \cos m{\theta }_1& -\sin m{\theta }_1& \cdots & 0& 0\\ 0& 0& \sin m{\theta }_1& \cos m{\theta }_1& \cdots & 0& 0\\ ⋮& ⋮& ⋮& ⋮& \ddots & ⋮& ⋮\\ 0& 0& 0& 0& \cdots & \cos m{\theta }_{d/2-1}& -\sin m{\theta }_{d/2-1}\\ 0& 0& 0& 0& \cdots & \sin m{\theta }_{d/2-1}& \cos m{\theta }_{d/2-1}\end{array}\right)$$
原作者提到，由于$R$非常稀疏，直接用矩阵乘法来实现会很浪费算力，推荐通过下述方式来实现RoPE：
$$\left(\begin{array}{c}{q}_0\\ q_1\\ q_2\\ q_3\\ ⋮\\ q_{d-2}\\ q_{d-1}\end{array}\right)\otimes \left(\begin{array}{c}\cos m{\theta }_0\\ \cos m{\theta }_0\\ \cos m{\theta }_1\\ \cos m{\theta }_1\\ ⋮\\ \cos m{\theta }_{d/2-1}\\ \cos m{\theta }_{d/2-1}\end{array}\right)+\left(\begin{array}{c}-q_1\\ q_0\\ -q_3\\ q_2\\ ⋮\\ -q_{d-1}\\ q_{d-2}\end{array}\right)\otimes \left(\begin{array}{c}\sin m{\theta }_0\\ \sin m{\theta }_0\\ \sin m{\theta }_1\\ \sin m{\theta }_1\\ ⋮\\ \sin m{\theta }_{d/2-1}\\ \sin m{\theta }_{d/2-1}\end{array}\right)$$

3.2、ChatGLM-6B中RoPE代码实现

这里直接上代码阅读

class RotaryEmbedding(torch.nn.Module):
    def __init__(self, dim, base=10000, precision=torch.half, learnable=False):
        super().__init__()
        inv_freq = 1. / (base ** (torch.arange(0, dim, 2).float() / dim))
        inv_freq = inv_freq.half()
        self.learnable = learnable
        if learnable:
            self.inv_freq = torch.nn.Parameter(inv_freq)
            self.max_seq_len_cached = None
        else:
            self.register_buffer('inv_freq', inv_freq)
            self.max_seq_len_cached = None
            self.cos_cached = None
            self.sin_cached = None
        self.precision = precision

    def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys,
                              error_msgs):
        pass

    def forward(self, x, seq_dim=1, seq_len=None):
        if seq_len is None:
            seq_len = x.shape[seq_dim]
        if self.max_seq_len_cached is None or (seq_len > self.max_seq_len_cached):
            self.max_seq_len_cached = None if self.learnable else seq_len
            # 在计算旋转嵌入之前，根据当前的嵌入维度和基数计算频率因子 inv_freq。将其转换为半精度数据类型（如果指定的 precision 为 bfloat16，则转换为单精度）。
            t = torch.arange(seq_len, device=x.device, dtype=self.inv_freq.dtype)
            # 使用 爱因斯坦求和函数 einsum 将 t 和 inv_freq 相乘，得到频率矩阵 freqs。
            freqs = torch.einsum('i,j->ij', t, self.inv_freq)
            # 通过在频率矩阵 freqs 中进行重复和拼接操作，生成旋转嵌入矩阵 emb，其维度为 [seq_len, 2 * dim]。
            emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
            if self.precision == torch.bfloat16:
                emb = emb.float()

            # 将旋转嵌入矩阵 emb 分别进行余弦和正弦运算。
            cos_cached = emb.cos()[:, None, :]
            sin_cached = emb.sin()[:, None, :]
            if self.precision == torch.bfloat16:
                cos_cached = cos_cached.bfloat16()
                sin_cached = sin_cached.bfloat16()
            if self.learnable:
                return cos_cached, sin_cached
            self.cos_cached, self.sin_cached = cos_cached, sin_cached
        # 按照序列长度截取
        return self.cos_cached[:seq_len, ...], self.sin_cached[:seq_len, ...]

四、注意力层

4.1、2D位置编码

ChatGLM-6B代码中这一层采用的位置编码是GLM的中提出的2D位置编码，详细原理见原文：GLM: General Language Model Pretraining with Autoregressive Blank Infilling，其原理图如下图：

输入的序列是$x_1,x_2,x_3,x_4,x_5,x_6$，片段$x_3$和片段$x_5、x_6$被随机MASK，原始的输入序列则变为$x_1,x_2,[M],x_4,[M]$，如上图(a)和(b)所示。将三个片段拼接得到模型的输入$x_1,x_2,[M],x_4,[M],[S],x_5,x_6,[S],x_3$，模型的输出则是被遮蔽掉的片段，如上图©所示。这里使用了2种位置编码：第一种编码为整个输入嵌入位置信息，能够表示MASK片段在原始输入中的位置;第二种位置编码则是为MASK片段内的tokens输入位置信息。

4.2、注意力机制

ChatGLM-6B相比标准的自注意力机制在Q和K中注入了RoPE位置信息。

• 标准自注意力机制attention_fn：

def attention_fn(
        self,
        query_layer,
        key_layer,
        value_layer,
        attention_mask,
        hidden_size_per_partition,
        layer_id,
        layer_past=None,
        scaling_attention_score=True,
        use_cache=False,
):
    # 考虑过去的信息
    if layer_past is not None:
        past_key, past_value = layer_past
        key_layer = torch.cat((past_key, key_layer), dim=0)
        value_layer = torch.cat((past_value, value_layer), dim=0)

    # seqlen, batch, num_attention_heads, hidden_size_per_attention_head
    seq_len, b, nh, hidden_size = key_layer.shape

    if use_cache:
        present = (key_layer, value_layer)
    else:
        present = None

    # 对查询层进行缩放操作，即将其除以（隐藏层大小的平方根乘以查询层的缩放系数）。这是为了控制注意力得分的尺度。
    query_key_layer_scaling_coeff = float(layer_id + 1)
    if scaling_attention_score:
        query_layer = query_layer / (math.sqrt(hidden_size) * query_key_layer_scaling_coeff)

    # ===================================
    # Raw attention scores. [b, np, s, s]
    # ===================================

    # # 注意力分数的输出形状: [batch_size, num_heads, seq_length, seq_length]
    output_size = (query_layer.size(1), query_layer.size(2), query_layer.size(0), key_layer.size(0))

    # 形状重塑：[seq_length, batch_size, num_heads, head_dim] -> [seq_length, batch_size*num_heads, head_dim]
    # [sq, b, np, hn] -> [sq, b * np, hn]
    query_layer = query_layer.view(output_size[2], output_size[0] * output_size[1], -1)
    # [sk, b, np, hn] -> [sk, b * np, hn]
    key_layer = key_layer.view(output_size[3], output_size[0] * output_size[1], -1)

    matmul_result = torch.empty(
        output_size[0] * output_size[1],
        output_size[2],
        output_size[3],
        dtype=query_layer.dtype,
        device=query_layer.device,
    )
    # 计算原始的注意力得分，通过转置和重塑操作，将查询、键和值的张量形状调整为合适的形状。
    matmul_result = torch.baddbmm(
        matmul_result,
        query_layer.transpose(0, 1),  # [b * np, sq, hn]
        key_layer.transpose(0, 1).transpose(1, 2),  # [b * np, hn, sk]
        beta=0.0,
        alpha=1.0,
    )

    # 重塑形状为：[batch_size,num_head,seq_length,seq_length]
    attention_scores = matmul_result.view(*output_size)

    # 如果指定了缩放的掩码 softmax（scale_mask_softmax），则将注意力得分传递给缩放的掩码 softmax 函数进行处理，以获得归一化的注意力概率。
    # 否则，将应用 softmax 操作，并根据需要填充一个较大的负数值（-10000.0）来屏蔽无效位置。
    if self.scale_mask_softmax:
        self.scale_mask_softmax.scale = query_key_layer_scaling_coeff
        attention_probs = self.scale_mask_softmax(attention_scores, attention_mask.contiguous())
    else:
        # 对注意力分数进行mask
        if not (attention_mask == 0).all():
            # if auto-regressive, skip
            attention_scores.masked_fill_(attention_mask, -10000.0)
        dtype = attention_scores.type()
        attention_scores = attention_scores.float()
        attention_scores = attention_scores * query_key_layer_scaling_coeff

        attention_probs = F.softmax(attention_scores, dim=-1)

        attention_probs = attention_probs.type(dtype)

    # =========================
    # Context layer. [sq, b, hp]
    # =========================

    # value_layer -> context layer.
    # [sk, b, np, hn] --> [b, np, sq, hn]

    # context layer shape: [b, np, sq, hn]
    output_size = (value_layer.size(1), value_layer.size(2), query_layer.size(0), value_layer.size(3))

    # change view [sk, b * np, hn]
    value_layer = value_layer.view(value_layer.size(0), output_size[0] * output_size[1], -1)

    # 对注意力分数进行mask
    # change view [b * np, sq, sk]
    attention_probs = attention_probs.view(output_size[0] * output_size[1], output_size[2], -1)
    # matmul: [b * np, sq, hn]
    context_layer = torch.bmm(attention_probs, value_layer.transpose(0, 1))
    # change view [b, np, sq, hn]
    context_layer = context_layer.view(*output_size)
    # [b, np, sq, hn] --> [sq, b, np, hn]
    context_layer = context_layer.permute(2, 0, 1, 3).contiguous()
    # [sq, b, np, hn] --> [sq, b, hp]
    new_context_layer_shape = context_layer.size()[:-2] + (hidden_size_per_partition,)
    # 重塑上下文层
    context_layer = context_layer.view(*new_context_layer_shape)
    outputs = (context_layer, present, attention_probs)

    return outputs

• SelfAttention的目的是为了捕捉序列中的位置信息，应用RoPE将位置信息注入Q和K。

  class SelfAttention(torch.nn.Module):
      def __init__(self, hidden_size, num_attention_heads,
                   layer_id, hidden_size_per_attention_head=None, bias=True,
                   params_dtype=torch.float, position_encoding_2d=True):
          super(SelfAttention, self).__init__()
  
          self.layer_id = layer_id
          self.hidden_size = hidden_size
          self.hidden_size_per_partition = hidden_size
          self.num_attention_heads = num_attention_heads
          self.num_attention_heads_per_partition = num_attention_heads
          self.position_encoding_2d = position_encoding_2d
          self.rotary_emb = RotaryEmbedding(
              self.hidden_size // (self.num_attention_heads * 2)
              if position_encoding_2d
              else self.hidden_size // self.num_attention_heads,
              base=10000,
              precision=torch.half,
              learnable=False,
          )
  
          self.scale_mask_softmax = None
  
          if hidden_size_per_attention_head is None:
              self.hidden_size_per_attention_head = hidden_size // num_attention_heads
          else:
              self.hidden_size_per_attention_head = hidden_size_per_attention_head
  
          self.inner_hidden_size = num_attention_heads * self.hidden_size_per_attention_head
  
          # Strided linear layer.
          self.query_key_value = skip_init(
              torch.nn.Linear,
              hidden_size,
              3 * self.inner_hidden_size,
              bias=bias,
              dtype=params_dtype,
          )
  
          self.dense = skip_init(
              torch.nn.Linear,
              self.inner_hidden_size,
              hidden_size,
              bias=bias,
              dtype=params_dtype,
          )
  
      @staticmethod
      def attention_mask_func(attention_scores, attention_mask):
          attention_scores.masked_fill_(attention_mask, -10000.0)
          return attention_scores
  
      def split_tensor_along_last_dim(self, tensor, num_partitions,
                                      contiguous_split_chunks=False):
          """Split a tensor along its last dimension.
          Arguments:
              tensor: input tensor.
              num_partitions: number of partitions to split the tensor
              contiguous_split_chunks: If True, make each chunk contiguous
                                      in memory.
          """
          # Get the size and dimension.
          last_dim = tensor.dim() - 1
          last_dim_size = tensor.size()[last_dim] // num_partitions
          # Split.
          tensor_list = torch.split(tensor, last_dim_size, dim=last_dim)
          # Note: torch.split does not create contiguous tensors by default.
          if contiguous_split_chunks:
              return tuple(chunk.contiguous() for chunk in tensor_list)
  
          return tensor_list
  
      def forward(
              self,
              hidden_states: torch.Tensor,
              position_ids,
              attention_mask: torch.Tensor,
              layer_id,
              layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
              use_cache: bool = False,
              output_attentions: bool = False,
      ):
          """
          hidden_states: [seq_len, batch, hidden_size]
          attention_mask: [(1, 1), seq_len, seq_len]
          """
  
          # [seq_len, batch, 3 * hidden_size]
          mixed_raw_layer = self.query_key_value(hidden_states)
  
          # [seq_len, batch, 3 * hidden_size] --> [seq_len, batch, num_attention_heads, 3 * hidden_size_per_attention_head]
          new_tensor_shape = mixed_raw_layer.size()[:-1] + (
              self.num_attention_heads_per_partition,
              3 * self.hidden_size_per_attention_head,
          )
          mixed_raw_layer = mixed_raw_layer.view(*new_tensor_shape)
  
          # [seq_len, batch, num_attention_heads, hidden_size_per_attention_head]
          (query_layer, key_layer, value_layer) = self.split_tensor_along_last_dim(mixed_raw_layer, 3)
  
          # 根据是否使用二维位置编码，对查询和键应用旋转嵌入，并根据位置信息进行索引操作。
          if self.position_encoding_2d:
              q1, q2 = query_layer.chunk(2, dim=(query_layer.ndim - 1))
              k1, k2 = key_layer.chunk(2, dim=(key_layer.ndim - 1))
              cos, sin = self.rotary_emb(q1, seq_len=position_ids.max() + 1)
              position_ids, block_position_ids = position_ids[:, 0, :].transpose(0, 1).contiguous(), \
                                                 position_ids[:, 1, :].transpose(0, 1).contiguous()
              q1, k1 = apply_rotary_pos_emb_index(q1, k1, cos, sin, position_ids)
              q2, k2 = apply_rotary_pos_emb_index(q2, k2, cos, sin, block_position_ids)
              # 拼接嵌入不同位置信息的query和key，这样query和key中包含了两种位置信息
              query_layer = torch.concat([q1, q2], dim=(q1.ndim - 1))
              key_layer = torch.concat([k1, k2], dim=(k1.ndim - 1))
          else:
              # RoPE
              position_ids = position_ids.transpose(0, 1)
              cos, sin = self.rotary_emb(value_layer, seq_len=position_ids.max() + 1)
              # [seq_len, batch, num_attention_heads, hidden_size_per_attention_head]
              query_layer, key_layer = apply_rotary_pos_emb_index(query_layer, key_layer, cos, sin, position_ids)
  
          # 调用 attention_fn 方法计算注意力得分和上下文层，其中使用了注意力函数的代码块
          # [seq_len, batch, hidden_size]
          context_layer, present, attention_probs = attention_fn(
              self=self,
              query_layer=query_layer,
              key_layer=key_layer,
              value_layer=value_layer,
              attention_mask=attention_mask,
              hidden_size_per_partition=self.hidden_size_per_partition,
              layer_id=layer_id,
              layer_past=layer_past,
              use_cache=use_cache
          )
  
          output = self.dense(context_layer)
  
          outputs = (output, present)
  
          if output_attentions:
              outputs += (attention_probs,)
  
          return outputs  # output, present, attention_probs
  
  

  五、GLU层

  根据代码，GLU形式化表示为：
  $$GLU(X)=GELU(X{W}_1)W_2$$

  class GEGLU(torch.nn.Module):
      def __init__(self):
          super().__init__()
          self.activation_fn = F.gelu
  
      def forward(self, x):
          # dim=-1 breaks in jit for pt<1.10
          x1, x2 = x.chunk(2, dim=(x.ndim - 1))
          return x1 * self.activation_fn(x2)
  
  
  class GLU(torch.nn.Module):
      def __init__(self, hidden_size, inner_hidden_size=None,
                   layer_id=None, bias=True, activation_func=gelu, params_dtype=torch.float):
          super(GLU, self).__init__()
          self.layer_id = layer_id
          self.activation_func = activation_func
  
          # Project to 4h.
          self.hidden_size = hidden_size
          if inner_hidden_size is None:
              inner_hidden_size = 4 * hidden_size
          self.inner_hidden_size = inner_hidden_size
          self.dense_h_to_4h = skip_init(
              torch.nn.Linear,
              self.hidden_size,
              self.inner_hidden_size,
              bias=bias,
              dtype=params_dtype,
          )
          # Project back to h.
          self.dense_4h_to_h = skip_init(
              torch.nn.Linear,
              self.inner_hidden_size,
              self.hidden_size,
              bias=bias,
              dtype=params_dtype,
          )
  
      def forward(self, hidden_states):
          """
          hidden_states: [seq_len, batch, hidden_size]
          """
  
          # [seq_len, batch, inner_hidden_size]
          intermediate_parallel = self.dense_h_to_4h(hidden_states)
  
          intermediate_parallel = self.activation_func(intermediate_parallel)
  
          output = self.dense_4h_to_h(intermediate_parallel)
  
          return output
  

  六、GLMBlock

根据代码，GLMBlock由Layer Norm、Self Attention、Layer Norm和GLU模块构成。

class GLMBlock(torch.nn.Module):
    def __init__(
            self,
            hidden_size,
            num_attention_heads,
            layernorm_epsilon,
            layer_id,
            inner_hidden_size=None,
            hidden_size_per_attention_head=None,
            layernorm=LayerNorm,
            use_bias=True,
            params_dtype=torch.float,
            num_layers=28,
            position_encoding_2d=True
    ):
        super(GLMBlock, self).__init__()
        # Set output layer initialization if not provided.

        self.layer_id = layer_id

        # LayerNorm层
        self.input_layernorm = layernorm(hidden_size, eps=layernorm_epsilon)
        # 是否使用2维位置编码
        self.position_encoding_2d = position_encoding_2d

        # 自注意力层
        self.attention = SelfAttention(
            hidden_size,
            num_attention_heads,
            layer_id,
            hidden_size_per_attention_head=hidden_size_per_attention_head,
            bias=use_bias,
            params_dtype=params_dtype,
            position_encoding_2d=self.position_encoding_2d
        )

        # LayerNorm层
        self.post_attention_layernorm = layernorm(hidden_size, eps=layernorm_epsilon)

        self.num_layers = num_layers

        # GLU层
        self.mlp = GLU(
            hidden_size,
            inner_hidden_size=inner_hidden_size,
            bias=use_bias,
            layer_id=layer_id,
            params_dtype=params_dtype,
        )

    def forward(
            self,
            hidden_states: torch.Tensor,
            position_ids,
            attention_mask: torch.Tensor,
            layer_id,
            layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
            use_cache: bool = False,
            output_attentions: bool = False,
    ):
        """
        hidden_states: [seq_len, batch, hidden_size]
        attention_mask: [(1, 1), seq_len, seq_len]
        """

        # 输入进行Layer Norm
        # [seq_len, batch, hidden_size]
        attention_input = self.input_layernorm(hidden_states)

        # 自注意力
        attention_outputs = self.attention(
            attention_input,
            position_ids,
            attention_mask=attention_mask,
            layer_id=layer_id,
            layer_past=layer_past,
            use_cache=use_cache,
            output_attentions=output_attentions
        )

        attention_output = attention_outputs[0]

        outputs = attention_outputs[1:]

        # Residual connection.
        alpha = (2 * self.num_layers) ** 0.5
        # 执行注意力残差连接
        hidden_states = attention_input * alpha + attention_output
        # 对注意力残差连接后的输出进行层归一化
        mlp_input = self.post_attention_layernorm(hidden_states)

        # 使用GLU层对归一化后的输出进行非线性变换
        mlp_output = self.mlp(mlp_input)

        # 执行GLU残差连接
        output = mlp_input * alpha + mlp_output

        if use_cache:
            outputs = (output,) + outputs
        else:
            outputs = (output,) + outputs[1:]

        return outputs  # hidden_states, present, attentions

七、ChatGLMPreTrainedModel

这一块主要看看其中的MASK和Position_ids

7.1、ChatGLM-6B的Mask

ChatGLM-6B采用prefix-LM的Mask，其对于输入的前缀使用双向注意力，对于后续的生成部分则是Causal Mask。

def get_masks(self, input_ids, device):
    batch_size, seq_length = input_ids.shape
    # context_lengths记录了batch中每个样本的真实长度
    context_lengths = [seq.tolist().index(self.config.bos_token_id) for seq in input_ids]
    # 生成causal mask，即下三角以及对角线为1，上三角为0
    attention_mask = torch.ones((batch_size, seq_length, seq_length), device=device)
    attention_mask.tril_()
    # 将前缀部分的注意力改为双向注意力
    for i, context_length in enumerate(context_lengths):
        attention_mask[i, :, :context_length] = 1
    attention_mask.unsqueeze_(1)
    attention_mask = (attention_mask < 0.5).bool()
        
    return attention_mask

7.2、ChatGLM-6B的Position_ids

def get_position_ids(self, input_ids, mask_positions, device, use_gmasks=None):
    """
    input_ids: [batch_size, seq_length]
    mask_positions: [batch_size]，由于GLM系列中会使用[Mask]或[gMask]标志，mask_positions就是指这些标注的具体位置
    """
    batch_size, seq_length = input_ids.shape
    if use_gmasks is None:
        use_gmasks = [False] * batch_size
    # context_lengths：未被padding前，batch中各个样本的长度
    context_lengths = [seq.tolist().index(self.config.bos_token_id) for seq in input_ids]
    # 2维位置编码
    if self.position_encoding_2d:
        # [0,1,2,...,seq_length-1]
        position_ids = torch.arange(seq_length, dtype=torch.long, device=device).unsqueeze(0).repeat(batch_size, 1)
        # 将原始输入后所有位置的postion id都设置为[Mask]或者[gMask]的位置id
        for i, context_length in enumerate(context_lengths):
            position_ids[i, context_length:] = mask_positions[i]
        # 原始输入的位置编码全部设置为0，待生成的位置添加顺序的位置id
        # 例如：[0,0,0,0,1,2,3,4,5]
        block_position_ids = [torch.cat((
            torch.zeros(context_length, dtype=torch.long, device=device),
            torch.arange(seq_length - context_length, dtype=torch.long, device=device) + 1
        )) for context_length in context_lengths]
        block_position_ids = torch.stack(block_position_ids, dim=0)
        # 将postion_ids和block_position_ids堆叠在一起，用于后续的参数传入；
        # 在注意力层中，还有将这个position_ids拆分为两部分
        position_ids = torch.stack((position_ids, block_position_ids), dim=1)
    else:
        position_ids = torch.arange(seq_length, dtype=torch.long, device=device).unsqueeze(0).repeat(batch_size, 1)
        for i, context_length in enumerate(context_lengths):
            if not use_gmasks[i]:
                position_ids[i, context_length:] = mask_positions[i]

    return position_ids

八、ChatGLMModel

这一块主要是模型的各部件的组合结构，直接看源码：

class ChatGLMModel(ChatGLMPreTrainedModel):
	def __init__(self, config: ChatGLMConfig, empty_init=True):
        super().__init__(config)
        if empty_init:
            init_method = skip_init
        else:
            init_method = default_init
        # recording parameters
        self.max_sequence_length = config.max_sequence_length
        self.hidden_size = config.hidden_size
        self.params_dtype = torch.half
        self.num_attention_heads = config.num_attention_heads
        self.vocab_size = config.vocab_size
        self.num_layers = config.num_layers
        self.layernorm_epsilon = config.layernorm_epsilon
        self.inner_hidden_size = config.inner_hidden_size
        self.hidden_size_per_attention_head = self.hidden_size // self.num_attention_heads
        self.position_encoding_2d = config.position_encoding_2d
        self.pre_seq_len = config.pre_seq_len
        self.prefix_projection = config.prefix_projection

        self.word_embeddings = init_method(
            torch.nn.Embedding,
            num_embeddings=self.vocab_size, embedding_dim=self.hidden_size,
            dtype=self.params_dtype
        )
        self.gradient_checkpointing = False

        def get_layer(layer_id):
            return GLMBlock(
                self.hidden_size,
                self.num_attention_heads,
                self.layernorm_epsilon,
                layer_id,
                inner_hidden_size=self.inner_hidden_size,
                hidden_size_per_attention_head=self.hidden_size_per_attention_head,
                layernorm=LayerNorm,
                use_bias=True,
                params_dtype=self.params_dtype,
                position_encoding_2d=self.position_encoding_2d,
                empty_init=empty_init
            )

        self.layers = torch.nn.ModuleList(
            [get_layer(layer_id) for layer_id in range(self.num_layers)]
        )

        # Final layer norm before output.
        self.final_layernorm = LayerNorm(self.hidden_size, eps=self.layernorm_epsilon)

        """
        pre_seq_len 为prompt部分长度，这部分仅编码，无反向传播

        """
        if self.pre_seq_len is not None:
            for param in self.parameters():
                param.requires_grad = False
            self.prefix_tokens = torch.arange(self.pre_seq_len).long()
            self.prefix_encoder = PrefixEncoder(config)
            self.dropout = torch.nn.Dropout(0.1)


    def get_prompt(self, batch_size, device, dtype=torch.half):
        """
        prompt 编码
        
        """
        prefix_tokens = self.prefix_tokens.unsqueeze(0).expand(batch_size, -1).to(device)
        past_key_values = self.prefix_encoder(prefix_tokens).type(dtype)
        past_key_values = past_key_values.view(
            batch_size,
            self.pre_seq_len,
            self.num_layers * 2,
            self.num_attention_heads,
            self.hidden_size // self.num_attention_heads
        )
        # seq_len, b, nh, hidden_size
        past_key_values = self.dropout(past_key_values)
        past_key_values = past_key_values.permute([2, 1, 0, 3, 4]).split(2)
        # past_key_values = [(v[0], v[1]) for v in past_key_values]
        return past_key_values


	def forward(
            self,
            input_ids: Optional[torch.LongTensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            attention_mask: Optional[torch.Tensor] = None,
            past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
            inputs_embeds: Optional[torch.LongTensor] = None,
            use_cache: Optional[bool] = None,
            output_attentions: Optional[bool] = None,
            output_hidden_states: Optional[bool] = None,
            return_dict: Optional[bool] = None,
    ) -> Union[Tuple[torch.Tensor, ...], BaseModelOutputWithPast]:
        .....
        """
	past_key_values机制是重要的机制，其可以防止模型在文本生成任务中重新计算上一次迭代
 	中已经计算好的上下文的值，大大提高了模型在文本生成任务中的计算效率。但要特别注意的是，
    在第一次迭代时由于不存在上一次迭代返回的past_key_values值，因此第一次迭代时
    past_key_values值为None。
    past_key_values 中每个元素的dim :
    num_layers * seq_len * batch_size * nh * hidden_size_per_head
        """
        if past_key_values is None:
            if self.pre_seq_len is not None:
                past_key_values = self.get_prompt(batch_size=input_ids.shape[0], device=input_ids.device,
                                                  dtype=inputs_embeds.dtype)
            else:
                past_key_values = tuple([None] * len(self.layers))

            if attention_mask is None:
                attention_mask = self.get_masks(
                    input_ids,
                    device=input_ids.device
                )


            if position_ids is None:
    """
    如果只有MASK无gMASK，则mask_positions 为第一个MASK的起始位置，
    如果有gMASK, 则mask_positions 为第一个gMASK的起始位置
    e.g.
    gMASK = 130001
    MASK = 130000
    seqs = [[11,22,MASK,33,MASK]]
    --> mask_positions:[2] use_gmask = [False]

    gMASK = 130001
    MASK = 130000
    seqs = seqs = [[11,22,MASK,33,MASK, gMASK, 55, 66, gMASK, 77]]
    --> mask_positions:[5] use_gmask = [True]
    
    把位置id结合mask位置信息由get_position_ids计算（为父类ChatGLMPreTrainedModel的方法）
    在使用2d position coding 时，position_ids dim = batch_size * 2 * seq_length 
    第二维包含 position_id 和 block_position_id 
    """
                MASK, gMASK = self.config.mask_token_id, self.config.gmask_token_id
                seqs = input_ids.tolist()

                mask_positions, use_gmasks = [], []
                for seq in seqs:
                    mask_token = gMASK if gMASK in seq else MASK
                    use_gmask = mask_token == gMASK
                    mask_positions.append(seq.index(mask_token))
                    use_gmasks.append(use_gmask)

                position_ids = self.get_position_ids(
                    input_ids,
                    mask_positions=mask_positions,
                    device=input_ids.device,
                    use_gmasks=use_gmasks
                )

        if self.pre_seq_len is not None and attention_mask is not None:
            prefix_attention_mask = torch.ones(batch_size, 1, input_ids.size(-1), self.pre_seq_len).to(
                attention_mask.device)
            prefix_attention_mask = (prefix_attention_mask < 0.5).bool()
            attention_mask = torch.cat((prefix_attention_mask, attention_mask), dim=3)

        """
        输入的embedding在这里进行了转置 
        batch_size * seq_len * hidden_size -> seq_len * batch_size * hidden_size
        """
        # [seq_len, batch, hidden_size]
        hidden_states = inputs_embeds.transpose(0, 1)

        presents = () if use_cache else None
        all_self_attentions = () if output_attentions else None
        all_hidden_states = () if output_hidden_states else None

        if attention_mask is None:
            attention_mask = torch.zeros(1, 1, device=input_ids.device).bool()
        else:
            attention_mask = attention_mask.to(hidden_states.device)

            
		for i, layer in enumerate(self.layers):

            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)
            layer_past = past_key_values[i]

            if self.gradient_checkpointing and self.training:
                layer_ret = torch.utils.checkpoint.checkpoint(
                    layer,
                    hidden_states,
                    position_ids,
                    attention_mask,
                    torch.tensor(i),
                    layer_past,
                    use_cache,
                    output_attentions
                )
            else:
                layer_ret = layer(
                    hidden_states,
                    position_ids=position_ids,
                    attention_mask=attention_mask,
                    layer_id=torch.tensor(i),
                    layer_past=layer_past,
                    use_cache=use_cache,
                    output_attentions=output_attentions
                )

            hidden_states = layer_ret[0]

            if use_cache:
                presents = presents + (layer_ret[1],)

            if output_attentions:
                all_self_attentions = all_self_attentions + (layer_ret[2 if use_cache else 1],)

        # Final layer norm.
        hidden_states = self.final_layernorm(hidden_states)

        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        if not return_dict:
            return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)

"""
	经过多个glm block堆叠，最后通过一个layernorm
"""
        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=presents,
            hidden_states=all_hidden_states,
            attentions=all_self_attentions,
        )

如又不详细之处，还望一起交流学习。

参考文献

1. GLM: General Language Model Pretraining with Autoregressive Blank Infilling)
2. modeling_chatglm.py · THUDM/chatglm-6b at main (huggingface.co)
3. Transformer升级之路：2、博采众长的旋转式位置编码 - 科学空间|Scientific Spaces (kexue.fm)