T5模型源码详解

参考资料：

(Alpha) T5 详解 | Blogs

整个源码主要围绕着Transformers库中的model_t5展开，也默认大家都已经熟悉了transformer的基本原理，知道了像GPT和Bert这些为代表的decoder和encoder架构，这里我们就介绍另一个流派的代表：基于encoder-decoder架构的T5模型

1.代码文件框架：

这里主要围绕modeling_t5展开（基本原理强力推荐上面的博客，下面主要是对官方代码的解读）：

class T5LayerNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        Construct a layernorm module in the T5 style. No bias and no subtraction of mean.
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)

        # convert into half-precision if necessary
        if self.weight.dtype in [torch.float16, torch.bfloat16]:
            hidden_states = hidden_states.to(self.weight.dtype)
        return self.weight * hidden_states

层归一化：hidden_states（B,L,C)，return （B,L,C)

class T5DenseActDense(nn.Module):
    def __init__(self, config: T5Config):
        super().__init__()
        self.wi = nn.Linear(config.d_model, config.d_ff, bias=False)
        self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)
        self.dropout = nn.Dropout(config.dropout_rate)
        self.act = ACT2FN[config.dense_act_fn]

    def forward(self, hidden_states):
        hidden_states = self.wi(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states = self.dropout(hidden_states)
        if (
            isinstance(self.wo.weight, torch.Tensor)
            and hidden_states.dtype != self.wo.weight.dtype
            and self.wo.weight.dtype != torch.int8
        ):
            hidden_states = hidden_states.to(self.wo.weight.dtype)
        hidden_states = self.wo(hidden_states)
        return hidden_states

主要作为FFN中的一部分，输入和输出维度没有变化，用于深层次加工和信息提炼

class T5DenseGatedActDense(nn.Module):
    def __init__(self, config: T5Config):
        super().__init__()
        self.wi_0 = nn.Linear(config.d_model, config.d_ff, bias=False)
        self.wi_1 = nn.Linear(config.d_model, config.d_ff, bias=False)
        self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)
        self.dropout = nn.Dropout(config.dropout_rate)
        self.act = ACT2FN[config.dense_act_fn]

    def forward(self, hidden_states):
        hidden_gelu = self.act(self.wi_0(hidden_states))
        hidden_linear = self.wi_1(hidden_states)
        hidden_states = hidden_gelu * hidden_linear
        hidden_states = self.dropout(hidden_states)

        # To make 8bit quantization work for google/flan-t5-xxl, self.wo is kept in float32.
        # See https://github.com/huggingface/transformers/issues/20287
        # we also make sure the weights are not in `int8` in case users will force `_keep_in_fp32_modules` to be `None``
        if (
            isinstance(self.wo.weight, torch.Tensor)
            and hidden_states.dtype != self.wo.weight.dtype
            and self.wo.weight.dtype != torch.int8
        ):
            hidden_states = hidden_states.to(self.wo.weight.dtype)

        hidden_states = self.wo(hidden_states)
        return hidden_states

这个和前面一个多了门控机制，也就是wi_0的输出当作模型重要性权重来控制wi_1，输入输出维度依旧不变

class T5LayerFF(nn.Module):
    def __init__(self, config: T5Config):
        super().__init__()
        if config.is_gated_act:
            self.DenseReluDense = T5DenseGatedActDense(config)
        else:
            self.DenseReluDense = T5DenseActDense(config)

        self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
        self.dropout = nn.Dropout(config.dropout_rate)

    def forward(self, hidden_states):
        forwarded_states = self.layer_norm(hidden_states)
        forwarded_states = self.DenseReluDense(forwarded_states)
        hidden_states = hidden_states + self.dropout(forwarded_states)
        return hidden_states

FFN，输入输出维度不变，用于特征融合。

下面介绍最重要的注意力机制：

1.初始化：

    def __init__(
        self,
        config: T5Config,
        has_relative_attention_bias=False,
        layer_idx: Optional[int] = None,
    ):
        super().__init__()
        self.is_decoder = config.is_decoder
        self.has_relative_attention_bias = has_relative_attention_bias
        self.relative_attention_num_buckets = config.relative_attention_num_buckets
        self.relative_attention_max_distance = config.relative_attention_max_distance
        self.d_model = config.d_model
        self.key_value_proj_dim = config.d_kv
        self.n_heads = config.num_heads
        self.dropout = config.dropout_rate
        self.inner_dim = self.n_heads * self.key_value_proj_dim
        self.layer_idx = layer_idx
        if layer_idx is None and self.is_decoder:
            logger.warning_once(
                f"Instantiating a decoder {self.__class__.__name__} without passing `layer_idx` is not recommended and "
                "will to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
                "when creating this class."
            )

        # Mesh TensorFlow initialization to avoid scaling before softmax
        self.q = nn.Linear(self.d_model, self.inner_dim, bias=False)
        self.k = nn.Linear(self.d_model, self.inner_dim, bias=False)
        self.v = nn.Linear(self.d_model, self.inner_dim, bias=False)
        self.o = nn.Linear(self.inner_dim, self.d_model, bias=False)

        if self.has_relative_attention_bias:
            self.relative_attention_bias = nn.Embedding(self.relative_attention_num_buckets, self.n_heads)
        self.pruned_heads = set()
        self.gradient_checkpointing = False

下面是各个参数的解释：

2.relative_position_bucket 函数

    def _relative_position_bucket(relative_position, bidirectional=True, num_buckets=32, max_distance=128):
        relative_buckets = 0
        if bidirectional:
            num_buckets //= 2
            relative_buckets += (relative_position > 0).to(torch.long) * num_buckets
            relative_position = torch.abs(relative_position)
        else:
            relative_position = -torch.min(relative_position, torch.zeros_like(relative_position))
        # now relative_position is in the range [0, inf)

        # half of the buckets are for exact increments in positions
        max_exact